{ "schemaVersion": 1, "deviceProperties": [{"id": 0, "name": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition", "totalGlobalMem": 102033981440, "computeMajor": 12, "computeMinor": 0, "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 1536, "regsPerBlock": 65536, "warpSize": 32, "sharedMemPerBlock": 49152, "numSms": 188, "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 101376, "sharedMemPerMultiprocessor": 102400}], "trace_id": "78CD23097D134704B741DBBF197F3AB6", "cuda_driver_version": 13040, "cuda_runtime_version": 13000, "cupti_version": 130001, "host_name": "skyler-rtx-pro-6000", "displayTimeUnit": "ms", "baseTimeNanoseconds": 1782967788000000000, "traceEvents": [ { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 1728719, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 1728719, "tid": 0, "args": { "labels": "CPU" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 1728719, "tid": 0, "args": { "sort_index": 1728719 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 0, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 0, "tid": 0, "args": { "labels": "GPU 0" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 0, "tid": 0, "args": { "sort_index": 5000000 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 1, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 1, "tid": 0, "args": { "labels": "GPU 1" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 1, "tid": 0, "args": { "sort_index": 5000001 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 2, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 2, "tid": 0, "args": { "labels": "GPU 2" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 2, "tid": 0, "args": { "sort_index": 5000002 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 3, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 3, "tid": 0, "args": { "labels": "GPU 3" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 3, "tid": 0, "args": { "sort_index": 5000003 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 4, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 4, "tid": 0, "args": { "labels": "GPU 4" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 4, "tid": 0, "args": { "sort_index": 5000004 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 5, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 5, "tid": 0, "args": { "labels": "GPU 5" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 5, "tid": 0, "args": { "sort_index": 5000005 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 6, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 6, "tid": 0, "args": { "labels": "GPU 6" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 6, "tid": 0, "args": { "sort_index": 5000006 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 7, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 7, "tid": 0, "args": { "labels": "GPU 7" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 7, "tid": 0, "args": { "sort_index": 5000007 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 8, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 8, "tid": 0, "args": { "labels": "GPU 8" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 8, "tid": 0, "args": { "sort_index": 5000008 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 9, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 9, "tid": 0, "args": { "labels": "GPU 9" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 9, "tid": 0, "args": { "sort_index": 5000009 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 10, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 10, "tid": 0, "args": { "labels": "GPU 10" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 10, "tid": 0, "args": { "sort_index": 5000010 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 11, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 11, "tid": 0, "args": { "labels": "GPU 11" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 11, "tid": 0, "args": { "sort_index": 5000011 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 12, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 12, "tid": 0, "args": { "labels": "GPU 12" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 12, "tid": 0, "args": { "sort_index": 5000012 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 13, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 13, "tid": 0, "args": { "labels": "GPU 13" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 13, "tid": 0, "args": { "sort_index": 5000013 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 14, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 14, "tid": 0, "args": { "labels": "GPU 14" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 14, "tid": 0, "args": { "sort_index": 5000014 } }, { "ph": "M", "name": "process_name", "ts": 6950381108176.507, "pid": 15, "tid": 0, "args": { "name": "python" } }, { "ph": "M", "name": "process_labels", "ts": 6950381108176.507, "pid": 15, "tid": 0, "args": { "labels": "GPU 15" } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108176.507, "pid": 15, "tid": 0, "args": { "sort_index": 5000015 } }, { "ph": "M", "name": "thread_name", "ts": 6950381108176.507, "pid": 0, "tid": 7, "args": { "name": "stream 7 " } }, { "ph": "M", "name": "thread_sort_index", "ts": 6950381108176.507, "pid": 0, "tid": 7, "args": { "sort_index": 7 } }, { "ph": "M", "name": "thread_name", "ts": 6950381108176.507, "pid": 1728719, "tid": 1728719, "args": { "name": "thread 1728719 (python)" } }, { "ph": "M", "name": "thread_sort_index", "ts": 6950381108176.507, "pid": 1728719, "tid": 1728719, "args": { "sort_index": 1728719 } }, { "ph": "M", "name": "thread_name", "ts": 6950381108176.507, "pid": 1728719, "tid": 1728719, "args": { "name": "thread 1728719 (python)" } }, { "ph": "M", "name": "thread_sort_index", "ts": 6950381108176.507, "pid": 1728719, "tid": 1728719, "args": { "sort_index": 1728719 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381108236.290, "dur": 150.119 , "args": { "External id": 1,"Record function id": 0, "Ev Idx": 0 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381108239.957, "dur": 9.257 , "args": { "External id": 2,"Record function id": 0, "Ev Idx": 1 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381108243.724, "dur": 5.309 , "args": { "External id": 3,"Record function id": 0, "Ev Idx": 2 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108245.767, "dur": 1.493 , "args": { "External id": 4,"Record function id": 0, "Ev Idx": 3 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381108250.406, "dur": 135.793 , "args": { "External id": 5,"Record function id": 0, "Ev Idx": 4 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381108252.720, "dur": 5.240 , "args": { "External id": 6,"Record function id": 0, "Ev Idx": 5 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1728719, "tid": 1728719, "ts": 6950381108257.249, "dur": 0.491 , "args": { "External id": 7,"Record function id": 0, "Ev Idx": 6 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381108258.631, "dur": 124.993 , "args": { "External id": 8,"Record function id": 0, "Ev Idx": 7 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1728719, "tid": 1728719, "ts": 6950381108384.826, "dur": 0.812 , "args": { "External id": 9,"Record function id": 0, "Ev Idx": 8 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108394.374, "dur": 22.562 , "args": { "External id": 10,"Record function id": 0, "Ev Idx": 9 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381108395.105, "dur": 21.701 , "args": { "External id": 11,"Record function id": 0, "Ev Idx": 10 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108399.864, "dur": 4.028 , "args": { "External id": 12,"Record function id": 0, "Ev Idx": 11 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381108404.934, "dur": 11.671 , "args": { "External id": 13,"Record function id": 0, "Ev Idx": 12 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 1728719, "tid": 1728719, "ts": 6950381108421.835, "dur": 11.722 , "args": { "External id": 14,"Record function id": 0, "Ev Idx": 13 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1728719, "tid": 1728719, "ts": 6950381108423.007, "dur": 0.381 , "args": { "External id": 15,"Record function id": 0, "Ev Idx": 14 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108423.709, "dur": 0.270 , "args": { "External id": 16,"Record function id": 0, "Ev Idx": 15 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 1728719, "tid": 1728719, "ts": 6950381108437.184, "dur": 9.487 , "args": { "External id": 17,"Record function id": 0, "Ev Idx": 16 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1728719, "tid": 1728719, "ts": 6950381108452.873, "dur": 7.935 , "args": { "External id": 18,"Record function id": 0, "Ev Idx": 17 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 1728719, "tid": 1728719, "ts": 6950381108463.432, "dur": 6.102 , "args": { "External id": 19,"Record function id": 0, "Ev Idx": 18 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 1728719, "tid": 1728719, "ts": 6950381108472.028, "dur": 6.813 , "args": { "External id": 20,"Record function id": 0, "Ev Idx": 19 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108479.953, "dur": 6.853 , "args": { "External id": 21,"Record function id": 0, "Ev Idx": 20 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381108480.164, "dur": 6.552 , "args": { "External id": 22,"Record function id": 0, "Ev Idx": 21 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108480.695, "dur": 1.853 , "args": { "External id": 23,"Record function id": 0, "Ev Idx": 22 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381108482.909, "dur": 3.687 , "args": { "External id": 24,"Record function id": 0, "Ev Idx": 23 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1728719, "tid": 1728719, "ts": 6950381108488.259, "dur": 5.259 , "args": { "External id": 25,"Record function id": 0, "Ev Idx": 24 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 1728719, "tid": 1728719, "ts": 6950381108494.590, "dur": 4.349 , "args": { "External id": 26,"Record function id": 0, "Ev Idx": 25 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108503.046, "dur": 6.613 , "args": { "External id": 27,"Record function id": 0, "Ev Idx": 26 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381108503.287, "dur": 6.281 , "args": { "External id": 28,"Record function id": 0, "Ev Idx": 27 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108503.557, "dur": 1.503 , "args": { "External id": 29,"Record function id": 0, "Ev Idx": 28 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381108505.340, "dur": 4.108 , "args": { "External id": 30,"Record function id": 0, "Ev Idx": 29 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381108514.397, "dur": 22.262 , "args": { "External id": 31,"Record function id": 0, "Ev Idx": 30 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381108514.628, "dur": 1.493 , "args": { "External id": 32,"Record function id": 0, "Ev Idx": 31 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381108514.928, "dur": 1.102 , "args": { "External id": 33,"Record function id": 0, "Ev Idx": 32 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108515.529, "dur": 0.361 , "args": { "External id": 34,"Record function id": 0, "Ev Idx": 33 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381108516.371, "dur": 20.138 , "args": { "External id": 35,"Record function id": 0, "Ev Idx": 34 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381108517.213, "dur": 1.152 , "args": { "External id": 36,"Record function id": 0, "Ev Idx": 35 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381108517.734, "dur": 0.490 , "args": { "External id": 37,"Record function id": 0, "Ev Idx": 36 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381108518.675, "dur": 16.832 , "args": { "External id": 38,"Record function id": 0, "Ev Idx": 37 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1728719, "tid": 1728719, "ts": 6950381108536.058, "dur": 0.210 , "args": { "External id": 39,"Record function id": 0, "Ev Idx": 38 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::gelu", "pid": 1728719, "tid": 1728719, "ts": 6950381108549.613, "dur": 7.644 , "args": { "External id": 40,"Record function id": 0, "Ev Idx": 39 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381108560.273, "dur": 12.032 , "args": { "External id": 41,"Record function id": 0, "Ev Idx": 40 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381108560.463, "dur": 0.782 , "args": { "External id": 42,"Record function id": 0, "Ev Idx": 41 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381108560.663, "dur": 0.481 , "args": { "External id": 43,"Record function id": 0, "Ev Idx": 42 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108560.924, "dur": 0.110 , "args": { "External id": 44,"Record function id": 0, "Ev Idx": 43 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381108561.495, "dur": 10.690 , "args": { "External id": 45,"Record function id": 0, "Ev Idx": 44 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381108561.946, "dur": 0.481 , "args": { "External id": 46,"Record function id": 0, "Ev Idx": 45 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381108562.146, "dur": 0.181 , "args": { "External id": 47,"Record function id": 0, "Ev Idx": 46 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381108562.677, "dur": 8.927 , "args": { "External id": 48,"Record function id": 0, "Ev Idx": 47 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1728719, "tid": 1728719, "ts": 6950381108571.934, "dur": 0.121 , "args": { "External id": 49,"Record function id": 0, "Ev Idx": 48 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::where", "pid": 1728719, "tid": 1728719, "ts": 6950381108575.231, "dur": 36.427 , "args": { "External id": 50,"Record function id": 0, "Ev Idx": 49 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::scalar_tensor", "pid": 1728719, "tid": 1728719, "ts": 6950381108576.293, "dur": 10.389 , "args": { "External id": 51,"Record function id": 0, "Ev Idx": 50 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108577.775, "dur": 1.854 , "args": { "External id": 52,"Record function id": 0, "Ev Idx": 51 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1728719, "tid": 1728719, "ts": 6950381108580.691, "dur": 5.791 , "args": { "External id": 53,"Record function id": 0, "Ev Idx": 52 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::scalar_tensor", "pid": 1728719, "tid": 1728719, "ts": 6950381108587.043, "dur": 5.049 , "args": { "External id": 54,"Record function id": 0, "Ev Idx": 53 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108587.293, "dur": 1.262 , "args": { "External id": 55,"Record function id": 0, "Ev Idx": 54 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1728719, "tid": 1728719, "ts": 6950381108588.836, "dur": 3.146 , "args": { "External id": 56,"Record function id": 0, "Ev Idx": 55 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::where", "pid": 1728719, "tid": 1728719, "ts": 6950381108593.014, "dur": 17.843 , "args": { "External id": 57,"Record function id": 0, "Ev Idx": 56 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108593.965, "dur": 0.972 , "args": { "External id": 58,"Record function id": 0, "Ev Idx": 57 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381108603.263, "dur": 1.723 , "args": { "External id": 59,"Record function id": 0, "Ev Idx": 58 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381108614.985, "dur": 1.673 , "args": { "External id": 60,"Record function id": 0, "Ev Idx": 59 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108615.976, "dur": 0.411 , "args": { "External id": 61,"Record function id": 0, "Ev Idx": 60 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381108618.150, "dur": 0.481 , "args": { "External id": 62,"Record function id": 0, "Ev Idx": 61 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108618.401, "dur": 0.100 , "args": { "External id": 63,"Record function id": 0, "Ev Idx": 62 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 1728719, "tid": 1728719, "ts": 6950381108620.575, "dur": 107.190 , "args": { "External id": 64,"Record function id": 0, "Ev Idx": 63 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 1728719, "tid": 1728719, "ts": 6950381108621.356, "dur": 0.191 , "args": { "External id": 65,"Record function id": 0, "Ev Idx": 64 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 1728719, "tid": 1728719, "ts": 6950381108624.061, "dur": 85.670 , "args": { "External id": 66,"Record function id": 0, "Ev Idx": 65 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cumsum", "pid": 1728719, "tid": 1728719, "ts": 6950381108625.995, "dur": 21.210 , "args": { "External id": 67,"Record function id": 0, "Ev Idx": 66 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108630.023, "dur": 0.060 , "args": { "External id": 68,"Record function id": 0, "Ev Idx": 67 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381108647.595, "dur": 0.752 , "args": { "External id": 69,"Record function id": 0, "Ev Idx": 68 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108648.036, "dur": 0.180 , "args": { "External id": 70,"Record function id": 0, "Ev Idx": 69 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381108649.248, "dur": 24.837 , "args": { "External id": 71,"Record function id": 0, "Ev Idx": 70 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381108650.080, "dur": 23.844 , "args": { "External id": 72,"Record function id": 0, "Ev Idx": 71 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::ge", "pid": 1728719, "tid": 1728719, "ts": 6950381108674.636, "dur": 8.596 , "args": { "External id": 73,"Record function id": 0, "Ev Idx": 72 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::all", "pid": 1728719, "tid": 1728719, "ts": 6950381108683.973, "dur": 8.516 , "args": { "External id": 74,"Record function id": 0, "Ev Idx": 73 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108686.087, "dur": 0.290 , "args": { "External id": 75,"Record function id": 0, "Ev Idx": 74 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381108692.920, "dur": 8.445 , "args": { "External id": 76,"Record function id": 0, "Ev Idx": 75 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381108693.090, "dur": 8.165 , "args": { "External id": 77,"Record function id": 0, "Ev Idx": 76 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108702.417, "dur": 1.894 , "args": { "External id": 78,"Record function id": 0, "Ev Idx": 77 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_select", "pid": 1728719, "tid": 1728719, "ts": 6950381108710.703, "dur": 16.500 , "args": { "External id": 79,"Record function id": 0, "Ev Idx": 78 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108711.454, "dur": 0.942 , "args": { "External id": 80,"Record function id": 0, "Ev Idx": 79 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381108713.979, "dur": 0.982 , "args": { "External id": 81,"Record function id": 0, "Ev Idx": 80 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381108716.353, "dur": 0.311 , "args": { "External id": 82,"Record function id": 0, "Ev Idx": 81 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1728719, "tid": 1728719, "ts": 6950381108717.395, "dur": 1.092 , "args": { "External id": 83,"Record function id": 0, "Ev Idx": 82 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108718.267, "dur": 0.070 , "args": { "External id": 84,"Record function id": 0, "Ev Idx": 83 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::gather", "pid": 1728719, "tid": 1728719, "ts": 6950381108719.239, "dur": 7.474 , "args": { "External id": 85,"Record function id": 0, "Ev Idx": 84 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108721.072, "dur": 0.110 , "args": { "External id": 86,"Record function id": 0, "Ev Idx": 85 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108721.543, "dur": 0.120 , "args": { "External id": 87,"Record function id": 0, "Ev Idx": 86 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_zeros", "pid": 1728719, "tid": 1728719, "ts": 6950381108732.814, "dur": 10.830 , "args": { "External id": 88,"Record function id": 0, "Ev Idx": 87 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108733.555, "dur": 2.816 , "args": { "External id": 89,"Record function id": 0, "Ev Idx": 88 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108734.046, "dur": 2.194 , "args": { "External id": 90,"Record function id": 0, "Ev Idx": 89 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1728719, "tid": 1728719, "ts": 6950381108737.182, "dur": 6.322 , "args": { "External id": 91,"Record function id": 0, "Ev Idx": 90 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1728719, "tid": 1728719, "ts": 6950381108737.994, "dur": 5.139 , "args": { "External id": 92,"Record function id": 0, "Ev Idx": 91 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1728719, "tid": 1728719, "ts": 6950381108746.620, "dur": 8.265 , "args": { "External id": 93,"Record function id": 0, "Ev Idx": 92 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 1728719, "tid": 1728719, "ts": 6950381108757.350, "dur": 59.320 , "args": { "External id": 94,"Record function id": 0, "Ev Idx": 93 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 1728719, "tid": 1728719, "ts": 6950381108757.921, "dur": 43.911 , "args": { "External id": 95,"Record function id": 0, "Ev Idx": 94 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cumsum", "pid": 1728719, "tid": 1728719, "ts": 6950381108758.452, "dur": 11.421 , "args": { "External id": 96,"Record function id": 0, "Ev Idx": 95 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108759.774, "dur": 0.090 , "args": { "External id": 97,"Record function id": 0, "Ev Idx": 96 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381108770.234, "dur": 0.901 , "args": { "External id": 98,"Record function id": 0, "Ev Idx": 97 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108770.865, "dur": 0.160 , "args": { "External id": 99,"Record function id": 0, "Ev Idx": 98 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381108771.526, "dur": 7.103 , "args": { "External id": 100,"Record function id": 0, "Ev Idx": 99 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381108771.746, "dur": 6.783 , "args": { "External id": 101,"Record function id": 0, "Ev Idx": 100 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::ge", "pid": 1728719, "tid": 1728719, "ts": 6950381108778.960, "dur": 5.200 , "args": { "External id": 102,"Record function id": 0, "Ev Idx": 101 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::all", "pid": 1728719, "tid": 1728719, "ts": 6950381108784.470, "dur": 5.440 , "args": { "External id": 103,"Record function id": 0, "Ev Idx": 102 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108785.803, "dur": 0.120 , "args": { "External id": 104,"Record function id": 0, "Ev Idx": 103 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381108790.171, "dur": 6.542 , "args": { "External id": 105,"Record function id": 0, "Ev Idx": 104 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381108790.341, "dur": 6.262 , "args": { "External id": 106,"Record function id": 0, "Ev Idx": 105 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108797.404, "dur": 1.323 , "args": { "External id": 107,"Record function id": 0, "Ev Idx": 106 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_select", "pid": 1728719, "tid": 1728719, "ts": 6950381108802.193, "dur": 13.866 , "args": { "External id": 108,"Record function id": 0, "Ev Idx": 107 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381108802.624, "dur": 0.922 , "args": { "External id": 109,"Record function id": 0, "Ev Idx": 108 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381108804.738, "dur": 1.633 , "args": { "External id": 110,"Record function id": 0, "Ev Idx": 109 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381108807.122, "dur": 0.221 , "args": { "External id": 111,"Record function id": 0, "Ev Idx": 110 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1728719, "tid": 1728719, "ts": 6950381108807.693, "dur": 0.822 , "args": { "External id": 112,"Record function id": 0, "Ev Idx": 111 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108808.094, "dur": 0.331 , "args": { "External id": 113,"Record function id": 0, "Ev Idx": 112 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::gather", "pid": 1728719, "tid": 1728719, "ts": 6950381108808.886, "dur": 6.742 , "args": { "External id": 114,"Record function id": 0, "Ev Idx": 113 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108809.857, "dur": 0.121 , "args": { "External id": 115,"Record function id": 0, "Ev Idx": 114 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108810.338, "dur": 0.171 , "args": { "External id": 116,"Record function id": 0, "Ev Idx": 115 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381108821.970, "dur": 0.190 , "args": { "External id": 117,"Record function id": 0, "Ev Idx": 116 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_put_", "pid": 1728719, "tid": 1728719, "ts": 6950381108823.703, "dur": 48.400 , "args": { "External id": 118,"Record function id": 0, "Ev Idx": 117 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_index_put_impl_", "pid": 1728719, "tid": 1728719, "ts": 6950381108824.885, "dur": 46.757 , "args": { "External id": 119,"Record function id": 0, "Ev Idx": 118 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 1728719, "tid": 1728719, "ts": 6950381108826.629, "dur": 35.275 , "args": { "External id": 120,"Record function id": 0, "Ev Idx": 119 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381108844.692, "dur": 0.862 , "args": { "External id": 121,"Record function id": 0, "Ev Idx": 120 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381108855.963, "dur": 1.473 , "args": { "External id": 122,"Record function id": 0, "Ev Idx": 121 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381108856.544, "dur": 0.802 , "args": { "External id": 123,"Record function id": 0, "Ev Idx": 122 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108857.045, "dur": 0.140 , "args": { "External id": 124,"Record function id": 0, "Ev Idx": 123 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381108858.528, "dur": 2.525 , "args": { "External id": 125,"Record function id": 0, "Ev Idx": 124 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381108859.840, "dur": 0.922 , "args": { "External id": 126,"Record function id": 0, "Ev Idx": 125 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381108862.375, "dur": 0.651 , "args": { "External id": 127,"Record function id": 0, "Ev Idx": 126 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108862.726, "dur": 0.150 , "args": { "External id": 128,"Record function id": 0, "Ev Idx": 127 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108864.609, "dur": 0.161 , "args": { "External id": 129,"Record function id": 0, "Ev Idx": 128 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381108865.180, "dur": 0.501 , "args": { "External id": 130,"Record function id": 0, "Ev Idx": 129 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381108865.411, "dur": 0.170 , "args": { "External id": 131,"Record function id": 0, "Ev Idx": 130 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108880.709, "dur": 0.681 , "args": { "External id": 132,"Record function id": 0, "Ev Idx": 131 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108882.232, "dur": 0.361 , "args": { "External id": 133,"Record function id": 0, "Ev Idx": 132 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108883.024, "dur": 0.100 , "args": { "External id": 134,"Record function id": 0, "Ev Idx": 133 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 1728719, "tid": 1728719, "ts": 6950381108889.896, "dur": 0.111 , "args": { "External id": 135,"Record function id": 0, "Ev Idx": 134 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 1728719, "tid": 1728719, "ts": 6950381108890.738, "dur": 0.080 , "args": { "External id": 136,"Record function id": 0, "Ev Idx": 135 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381108891.529, "dur": 33.453 , "args": { "External id": 137,"Record function id": 0, "Ev Idx": 136 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381108891.950, "dur": 32.911 , "args": { "External id": 138,"Record function id": 0, "Ev Idx": 137 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381108892.862, "dur": 1.192 , "args": { "External id": 139,"Record function id": 0, "Ev Idx": 138 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381108894.465, "dur": 30.216 , "args": { "External id": 140,"Record function id": 0, "Ev Idx": 139 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109025.489, "dur": 0.731 , "args": { "External id": 141,"Record function id": 0, "Ev Idx": 140 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109142.156, "dur": 19.827 , "args": { "External id": 142,"Record function id": 0, "Ev Idx": 141 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109142.466, "dur": 19.387 , "args": { "External id": 143,"Record function id": 0, "Ev Idx": 142 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109142.947, "dur": 2.054 , "args": { "External id": 144,"Record function id": 0, "Ev Idx": 143 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109145.352, "dur": 16.360 , "args": { "External id": 145,"Record function id": 0, "Ev Idx": 144 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381109163.195, "dur": 0.110 , "args": { "External id": 146,"Record function id": 0, "Ev Idx": 145 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1728719, "tid": 1728719, "ts": 6950381109164.227, "dur": 0.110 , "args": { "External id": 147,"Record function id": 0, "Ev Idx": 146 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1728719, "tid": 1728719, "ts": 6950381109166.832, "dur": 9.928 , "args": { "External id": 148,"Record function id": 0, "Ev Idx": 147 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109167.393, "dur": 1.242 , "args": { "External id": 149,"Record function id": 0, "Ev Idx": 148 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109168.976, "dur": 7.644 , "args": { "External id": 150,"Record function id": 0, "Ev Idx": 149 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1728719, "tid": 1728719, "ts": 6950381109177.371, "dur": 5.541 , "args": { "External id": 151,"Record function id": 0, "Ev Idx": 150 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109177.602, "dur": 1.392 , "args": { "External id": 152,"Record function id": 0, "Ev Idx": 151 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109179.285, "dur": 3.527 , "args": { "External id": 153,"Record function id": 0, "Ev Idx": 152 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109184.455, "dur": 0.260 , "args": { "External id": 154,"Record function id": 0, "Ev Idx": 153 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109298.707, "dur": 16.220 , "args": { "External id": 155,"Record function id": 0, "Ev Idx": 154 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109298.938, "dur": 15.899 , "args": { "External id": 156,"Record function id": 0, "Ev Idx": 155 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109299.278, "dur": 1.744 , "args": { "External id": 157,"Record function id": 0, "Ev Idx": 156 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109301.282, "dur": 13.425 , "args": { "External id": 158,"Record function id": 0, "Ev Idx": 157 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381109315.398, "dur": 0.030 , "args": { "External id": 159,"Record function id": 0, "Ev Idx": 158 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1728719, "tid": 1728719, "ts": 6950381109315.709, "dur": 0.030 , "args": { "External id": 160,"Record function id": 0, "Ev Idx": 159 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_put_", "pid": 1728719, "tid": 1728719, "ts": 6950381109317.953, "dur": 40.986 , "args": { "External id": 161,"Record function id": 0, "Ev Idx": 160 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_index_put_impl_", "pid": 1728719, "tid": 1728719, "ts": 6950381109318.364, "dur": 40.465 , "args": { "External id": 162,"Record function id": 0, "Ev Idx": 161 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 1728719, "tid": 1728719, "ts": 6950381109319.105, "dur": 31.629 , "args": { "External id": 163,"Record function id": 0, "Ev Idx": 162 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381109338.041, "dur": 1.122 , "args": { "External id": 164,"Record function id": 0, "Ev Idx": 163 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381109347.648, "dur": 1.102 , "args": { "External id": 165,"Record function id": 0, "Ev Idx": 164 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381109347.889, "dur": 0.771 , "args": { "External id": 166,"Record function id": 0, "Ev Idx": 165 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109348.280, "dur": 0.250 , "args": { "External id": 167,"Record function id": 0, "Ev Idx": 166 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381109349.121, "dur": 0.852 , "args": { "External id": 168,"Record function id": 0, "Ev Idx": 167 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381109349.432, "dur": 0.340 , "args": { "External id": 169,"Record function id": 0, "Ev Idx": 168 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381109351.125, "dur": 0.821 , "args": { "External id": 170,"Record function id": 0, "Ev Idx": 169 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109351.596, "dur": 0.220 , "args": { "External id": 171,"Record function id": 0, "Ev Idx": 170 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109352.788, "dur": 0.120 , "args": { "External id": 172,"Record function id": 0, "Ev Idx": 171 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109353.229, "dur": 0.591 , "args": { "External id": 173,"Record function id": 0, "Ev Idx": 172 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109353.459, "dur": 0.261 , "args": { "External id": 174,"Record function id": 0, "Ev Idx": 173 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109361.915, "dur": 0.431 , "args": { "External id": 175,"Record function id": 0, "Ev Idx": 174 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109478.001, "dur": 16.781 , "args": { "External id": 176,"Record function id": 0, "Ev Idx": 175 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109478.392, "dur": 16.300 , "args": { "External id": 177,"Record function id": 0, "Ev Idx": 176 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109478.712, "dur": 1.563 , "args": { "External id": 178,"Record function id": 0, "Ev Idx": 177 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109480.576, "dur": 13.976 , "args": { "External id": 179,"Record function id": 0, "Ev Idx": 178 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381109495.233, "dur": 0.100 , "args": { "External id": 180,"Record function id": 0, "Ev Idx": 179 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1728719, "tid": 1728719, "ts": 6950381109495.704, "dur": 0.040 , "args": { "External id": 181,"Record function id": 0, "Ev Idx": 180 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_put_", "pid": 1728719, "tid": 1728719, "ts": 6950381109496.836, "dur": 37.881 , "args": { "External id": 182,"Record function id": 0, "Ev Idx": 181 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_index_put_impl_", "pid": 1728719, "tid": 1728719, "ts": 6950381109497.067, "dur": 37.539 , "args": { "External id": 183,"Record function id": 0, "Ev Idx": 182 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 1728719, "tid": 1728719, "ts": 6950381109497.547, "dur": 25.558 , "args": { "External id": 184,"Record function id": 0, "Ev Idx": 183 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381109512.646, "dur": 0.781 , "args": { "External id": 185,"Record function id": 0, "Ev Idx": 184 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381109520.761, "dur": 0.901 , "args": { "External id": 186,"Record function id": 0, "Ev Idx": 185 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381109520.981, "dur": 0.581 , "args": { "External id": 187,"Record function id": 0, "Ev Idx": 186 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109521.312, "dur": 0.130 , "args": { "External id": 188,"Record function id": 0, "Ev Idx": 187 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381109521.923, "dur": 0.561 , "args": { "External id": 189,"Record function id": 0, "Ev Idx": 188 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381109522.153, "dur": 0.130 , "args": { "External id": 190,"Record function id": 0, "Ev Idx": 189 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381109523.396, "dur": 0.500 , "args": { "External id": 191,"Record function id": 0, "Ev Idx": 190 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109523.656, "dur": 0.110 , "args": { "External id": 192,"Record function id": 0, "Ev Idx": 191 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109524.437, "dur": 0.081 , "args": { "External id": 193,"Record function id": 0, "Ev Idx": 192 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109524.788, "dur": 0.401 , "args": { "External id": 194,"Record function id": 0, "Ev Idx": 193 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109524.978, "dur": 0.111 , "args": { "External id": 195,"Record function id": 0, "Ev Idx": 194 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 1728719, "tid": 1728719, "ts": 6950381109537.341, "dur": 10.119 , "args": { "External id": 196,"Record function id": 0, "Ev Idx": 195 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109538.664, "dur": 0.190 , "args": { "External id": 197,"Record function id": 0, "Ev Idx": 196 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109539.135, "dur": 0.431 , "args": { "External id": 198,"Record function id": 0, "Ev Idx": 197 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109539.305, "dur": 0.170 , "args": { "External id": 199,"Record function id": 0, "Ev Idx": 198 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 1728719, "tid": 1728719, "ts": 6950381109548.933, "dur": 6.142 , "args": { "External id": 200,"Record function id": 0, "Ev Idx": 199 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109549.354, "dur": 0.100 , "args": { "External id": 201,"Record function id": 0, "Ev Idx": 200 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109549.705, "dur": 0.350 , "args": { "External id": 202,"Record function id": 0, "Ev Idx": 201 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109549.865, "dur": 0.100 , "args": { "External id": 203,"Record function id": 0, "Ev Idx": 202 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 1728719, "tid": 1728719, "ts": 6950381109556.157, "dur": 5.700 , "args": { "External id": 204,"Record function id": 0, "Ev Idx": 203 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109556.467, "dur": 0.110 , "args": { "External id": 205,"Record function id": 0, "Ev Idx": 204 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109556.828, "dur": 0.441 , "args": { "External id": 206,"Record function id": 0, "Ev Idx": 205 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109557.088, "dur": 0.090 , "args": { "External id": 207,"Record function id": 0, "Ev Idx": 206 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1728719, "tid": 1728719, "ts": 6950381109563.741, "dur": 7.443 , "args": { "External id": 208,"Record function id": 0, "Ev Idx": 207 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::nonzero_numpy", "pid": 1728719, "tid": 1728719, "ts": 6950381109587.044, "dur": 30.627 , "args": { "External id": 209,"Record function id": 0, "Ev Idx": 208 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::nonzero", "pid": 1728719, "tid": 1728719, "ts": 6950381109587.445, "dur": 28.112 , "args": { "External id": 210,"Record function id": 0, "Ev Idx": 209 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381109604.897, "dur": 0.942 , "args": { "External id": 211,"Record function id": 0, "Ev Idx": 210 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381109613.123, "dur": 0.941 , "args": { "External id": 212,"Record function id": 0, "Ev Idx": 211 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381109613.453, "dur": 0.511 , "args": { "External id": 213,"Record function id": 0, "Ev Idx": 212 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109613.754, "dur": 0.100 , "args": { "External id": 214,"Record function id": 0, "Ev Idx": 213 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381109614.345, "dur": 0.471 , "args": { "External id": 215,"Record function id": 0, "Ev Idx": 214 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1728719, "tid": 1728719, "ts": 6950381109614.585, "dur": 0.121 , "args": { "External id": 216,"Record function id": 0, "Ev Idx": 215 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::unbind", "pid": 1728719, "tid": 1728719, "ts": 6950381109616.308, "dur": 1.193 , "args": { "External id": 217,"Record function id": 0, "Ev Idx": 216 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381109616.799, "dur": 0.561 , "args": { "External id": 218,"Record function id": 0, "Ev Idx": 217 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109617.120, "dur": 0.090 , "args": { "External id": 219,"Record function id": 0, "Ev Idx": 218 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::full_like", "pid": 1728719, "tid": 1728719, "ts": 6950381109625.345, "dur": 8.025 , "args": { "External id": 220,"Record function id": 0, "Ev Idx": 219 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381109626.588, "dur": 1.903 , "args": { "External id": 221,"Record function id": 0, "Ev Idx": 220 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109626.998, "dur": 1.333 , "args": { "External id": 222,"Record function id": 0, "Ev Idx": 221 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1728719, "tid": 1728719, "ts": 6950381109629.042, "dur": 4.148 , "args": { "External id": 223,"Record function id": 0, "Ev Idx": 222 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 1728719, "tid": 1728719, "ts": 6950381109636.095, "dur": 9.728 , "args": { "External id": 224,"Record function id": 0, "Ev Idx": 223 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109636.887, "dur": 1.202 , "args": { "External id": 225,"Record function id": 0, "Ev Idx": 224 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 1728719, "tid": 1728719, "ts": 6950381109638.941, "dur": 6.682 , "args": { "External id": 226,"Record function id": 0, "Ev Idx": 225 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381109640.213, "dur": 0.922 , "args": { "External id": 227,"Record function id": 0, "Ev Idx": 226 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109648.308, "dur": 0.321 , "args": { "External id": 228,"Record function id": 0, "Ev Idx": 227 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109649.270, "dur": 10.610 , "args": { "External id": 229,"Record function id": 0, "Ev Idx": 228 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109649.590, "dur": 10.199 , "args": { "External id": 230,"Record function id": 0, "Ev Idx": 229 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109650.141, "dur": 1.674 , "args": { "External id": 231,"Record function id": 0, "Ev Idx": 230 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109652.135, "dur": 7.524 , "args": { "External id": 232,"Record function id": 0, "Ev Idx": 231 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381109660.381, "dur": 0.050 , "args": { "External id": 233,"Record function id": 0, "Ev Idx": 232 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1728719, "tid": 1728719, "ts": 6950381109660.731, "dur": 0.050 , "args": { "External id": 234,"Record function id": 0, "Ev Idx": 233 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 1728719, "tid": 1728719, "ts": 6950381109661.803, "dur": 67.155 , "args": { "External id": 235,"Record function id": 0, "Ev Idx": 234 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 1728719, "tid": 1728719, "ts": 6950381109662.174, "dur": 0.140 , "args": { "External id": 236,"Record function id": 0, "Ev Idx": 235 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109662.795, "dur": 0.501 , "args": { "External id": 237,"Record function id": 0, "Ev Idx": 236 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109662.985, "dur": 0.201 , "args": { "External id": 238,"Record function id": 0, "Ev Idx": 237 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1728719, "tid": 1728719, "ts": 6950381109663.767, "dur": 0.721 , "args": { "External id": 239,"Record function id": 0, "Ev Idx": 238 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109664.248, "dur": 0.130 , "args": { "External id": 240,"Record function id": 0, "Ev Idx": 239 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 1728719, "tid": 1728719, "ts": 6950381109664.969, "dur": 50.444 , "args": { "External id": 241,"Record function id": 0, "Ev Idx": 240 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cumsum", "pid": 1728719, "tid": 1728719, "ts": 6950381109665.520, "dur": 12.674 , "args": { "External id": 242,"Record function id": 0, "Ev Idx": 241 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109667.143, "dur": 0.100 , "args": { "External id": 243,"Record function id": 0, "Ev Idx": 242 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1728719, "tid": 1728719, "ts": 6950381109678.524, "dur": 0.652 , "args": { "External id": 244,"Record function id": 0, "Ev Idx": 243 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109678.885, "dur": 0.160 , "args": { "External id": 245,"Record function id": 0, "Ev Idx": 244 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381109679.807, "dur": 7.584 , "args": { "External id": 246,"Record function id": 0, "Ev Idx": 245 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381109680.237, "dur": 7.024 , "args": { "External id": 247,"Record function id": 0, "Ev Idx": 246 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::ge", "pid": 1728719, "tid": 1728719, "ts": 6950381109687.882, "dur": 6.301 , "args": { "External id": 248,"Record function id": 0, "Ev Idx": 247 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::all", "pid": 1728719, "tid": 1728719, "ts": 6950381109694.544, "dur": 8.696 , "args": { "External id": 249,"Record function id": 0, "Ev Idx": 248 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1728719, "tid": 1728719, "ts": 6950381109696.678, "dur": 0.731 , "args": { "External id": 250,"Record function id": 0, "Ev Idx": 249 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109697.069, "dur": 0.210 , "args": { "External id": 251,"Record function id": 0, "Ev Idx": 250 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109697.700, "dur": 0.040 , "args": { "External id": 252,"Record function id": 0, "Ev Idx": 251 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109698.011, "dur": 4.788 , "args": { "External id": 253,"Record function id": 0, "Ev Idx": 252 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381109703.571, "dur": 6.111 , "args": { "External id": 254,"Record function id": 0, "Ev Idx": 253 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381109703.741, "dur": 5.831 , "args": { "External id": 255,"Record function id": 0, "Ev Idx": 254 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109710.494, "dur": 1.442 , "args": { "External id": 256,"Record function id": 0, "Ev Idx": 255 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_select", "pid": 1728719, "tid": 1728719, "ts": 6950381109715.924, "dur": 12.603 , "args": { "External id": 257,"Record function id": 0, "Ev Idx": 256 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109716.445, "dur": 0.902 , "args": { "External id": 258,"Record function id": 0, "Ev Idx": 257 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381109718.679, "dur": 0.932 , "args": { "External id": 259,"Record function id": 0, "Ev Idx": 258 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109720.372, "dur": 0.171 , "args": { "External id": 260,"Record function id": 0, "Ev Idx": 259 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1728719, "tid": 1728719, "ts": 6950381109720.853, "dur": 0.501 , "args": { "External id": 261,"Record function id": 0, "Ev Idx": 260 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109721.104, "dur": 0.140 , "args": { "External id": 262,"Record function id": 0, "Ev Idx": 261 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::gather", "pid": 1728719, "tid": 1728719, "ts": 6950381109721.735, "dur": 6.332 , "args": { "External id": 263,"Record function id": 0, "Ev Idx": 262 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109722.787, "dur": 0.090 , "args": { "External id": 264,"Record function id": 0, "Ev Idx": 263 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109723.288, "dur": 0.330 , "args": { "External id": 265,"Record function id": 0, "Ev Idx": 264 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_put_", "pid": 1728719, "tid": 1728719, "ts": 6950381109731.793, "dur": 5.761 , "args": { "External id": 266,"Record function id": 0, "Ev Idx": 265 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_index_put_impl_", "pid": 1728719, "tid": 1728719, "ts": 6950381109732.124, "dur": 5.320 , "args": { "External id": 267,"Record function id": 0, "Ev Idx": 266 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109733.126, "dur": 0.100 , "args": { "External id": 268,"Record function id": 0, "Ev Idx": 267 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109733.507, "dur": 0.420 , "args": { "External id": 269,"Record function id": 0, "Ev Idx": 268 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109733.727, "dur": 0.110 , "args": { "External id": 270,"Record function id": 0, "Ev Idx": 269 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1728719, "tid": 1728719, "ts": 6950381109739.989, "dur": 8.265 , "args": { "External id": 271,"Record function id": 0, "Ev Idx": 270 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381109740.720, "dur": 1.753 , "args": { "External id": 272,"Record function id": 0, "Ev Idx": 271 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109741.011, "dur": 1.362 , "args": { "External id": 273,"Record function id": 0, "Ev Idx": 272 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1728719, "tid": 1728719, "ts": 6950381109742.884, "dur": 5.240 , "args": { "External id": 274,"Record function id": 0, "Ev Idx": 273 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1728719, "tid": 1728719, "ts": 6950381109743.395, "dur": 4.388 , "args": { "External id": 275,"Record function id": 0, "Ev Idx": 274 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1728719, "tid": 1728719, "ts": 6950381109751.079, "dur": 1.173 , "args": { "External id": 276,"Record function id": 0, "Ev Idx": 275 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109751.801, "dur": 0.100 , "args": { "External id": 277,"Record function id": 0, "Ev Idx": 276 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381109754.696, "dur": 0.030 , "args": { "External id": 278,"Record function id": 0, "Ev Idx": 277 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::index_put_", "pid": 1728719, "tid": 1728719, "ts": 6950381109755.528, "dur": 16.661 , "args": { "External id": 279,"Record function id": 0, "Ev Idx": 278 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_index_put_impl_", "pid": 1728719, "tid": 1728719, "ts": 6950381109755.778, "dur": 16.310 , "args": { "External id": 280,"Record function id": 0, "Ev Idx": 279 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109756.519, "dur": 9.648 , "args": { "External id": 281,"Record function id": 0, "Ev Idx": 280 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109757.000, "dur": 9.047 , "args": { "External id": 282,"Record function id": 0, "Ev Idx": 281 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109757.341, "dur": 1.272 , "args": { "External id": 283,"Record function id": 0, "Ev Idx": 282 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109758.904, "dur": 7.033 , "args": { "External id": 284,"Record function id": 0, "Ev Idx": 283 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109766.708, "dur": 0.101 , "args": { "External id": 285,"Record function id": 0, "Ev Idx": 284 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381109767.069, "dur": 0.381 , "args": { "External id": 286,"Record function id": 0, "Ev Idx": 285 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381109767.249, "dur": 0.101 , "args": { "External id": 287,"Record function id": 0, "Ev Idx": 286 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109777.128, "dur": 0.130 , "args": { "External id": 288,"Record function id": 0, "Ev Idx": 287 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_zeros", "pid": 1728719, "tid": 1728719, "ts": 6950381109779.492, "dur": 7.725 , "args": { "External id": 289,"Record function id": 0, "Ev Idx": 288 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109780.163, "dur": 1.714 , "args": { "External id": 290,"Record function id": 0, "Ev Idx": 289 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381109780.484, "dur": 1.252 , "args": { "External id": 291,"Record function id": 0, "Ev Idx": 290 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1728719, "tid": 1728719, "ts": 6950381109782.177, "dur": 4.929 , "args": { "External id": 292,"Record function id": 0, "Ev Idx": 291 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1728719, "tid": 1728719, "ts": 6950381109782.558, "dur": 4.238 , "args": { "External id": 293,"Record function id": 0, "Ev Idx": 292 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1728719, "tid": 1728719, "ts": 6950381109788.870, "dur": 6.101 , "args": { "External id": 294,"Record function id": 0, "Ev Idx": 293 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109802.245, "dur": 6.442 , "args": { "External id": 295,"Record function id": 0, "Ev Idx": 294 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109802.455, "dur": 6.131 , "args": { "External id": 296,"Record function id": 0, "Ev Idx": 295 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109802.926, "dur": 1.302 , "args": { "External id": 297,"Record function id": 0, "Ev Idx": 296 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109804.589, "dur": 3.887 , "args": { "External id": 298,"Record function id": 0, "Ev Idx": 297 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 1728719, "tid": 1728719, "ts": 6950381109810.630, "dur": 6.883 , "args": { "External id": 299,"Record function id": 0, "Ev Idx": 298 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1728719, "tid": 1728719, "ts": 6950381109820.288, "dur": 0.952 , "args": { "External id": 300,"Record function id": 0, "Ev Idx": 299 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109820.859, "dur": 0.201 , "args": { "External id": 301,"Record function id": 0, "Ev Idx": 300 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1728719, "tid": 1728719, "ts": 6950381109823.414, "dur": 0.571 , "args": { "External id": 302,"Record function id": 0, "Ev Idx": 301 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109823.785, "dur": 0.100 , "args": { "External id": 303,"Record function id": 0, "Ev Idx": 302 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109825.157, "dur": 0.211 , "args": { "External id": 304,"Record function id": 0, "Ev Idx": 303 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 1728719, "tid": 1728719, "ts": 6950381109826.039, "dur": 5.400 , "args": { "External id": 305,"Record function id": 0, "Ev Idx": 304 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cos", "pid": 1728719, "tid": 1728719, "ts": 6950381109833.062, "dur": 6.222 , "args": { "External id": 306,"Record function id": 0, "Ev Idx": 305 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::sin", "pid": 1728719, "tid": 1728719, "ts": 6950381109840.826, "dur": 5.100 , "args": { "External id": 307,"Record function id": 0, "Ev Idx": 306 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1728719, "tid": 1728719, "ts": 6950381109846.988, "dur": 7.043 , "args": { "External id": 308,"Record function id": 0, "Ev Idx": 307 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109856.626, "dur": 0.140 , "args": { "External id": 309,"Record function id": 0, "Ev Idx": 308 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109859.571, "dur": 6.082 , "args": { "External id": 310,"Record function id": 0, "Ev Idx": 309 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109859.812, "dur": 5.750 , "args": { "External id": 311,"Record function id": 0, "Ev Idx": 310 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109860.112, "dur": 1.403 , "args": { "External id": 312,"Record function id": 0, "Ev Idx": 311 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109861.795, "dur": 3.647 , "args": { "External id": 313,"Record function id": 0, "Ev Idx": 312 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381109869.620, "dur": 18.124 , "args": { "External id": 314,"Record function id": 0, "Ev Idx": 313 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381109869.951, "dur": 1.142 , "args": { "External id": 315,"Record function id": 0, "Ev Idx": 314 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381109870.231, "dur": 0.762 , "args": { "External id": 316,"Record function id": 0, "Ev Idx": 315 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109870.602, "dur": 0.230 , "args": { "External id": 317,"Record function id": 0, "Ev Idx": 316 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381109871.423, "dur": 16.130 , "args": { "External id": 318,"Record function id": 0, "Ev Idx": 317 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381109871.734, "dur": 15.649 , "args": { "External id": 319,"Record function id": 0, "Ev Idx": 318 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 1728719, "tid": 1728719, "ts": 6950381109891.831, "dur": 6.012 , "args": { "External id": 320,"Record function id": 0, "Ev Idx": 319 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381109901.259, "dur": 15.769 , "args": { "External id": 321,"Record function id": 0, "Ev Idx": 320 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381109901.439, "dur": 0.912 , "args": { "External id": 322,"Record function id": 0, "Ev Idx": 321 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381109901.720, "dur": 0.531 , "args": { "External id": 323,"Record function id": 0, "Ev Idx": 322 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109902.020, "dur": 0.121 , "args": { "External id": 324,"Record function id": 0, "Ev Idx": 323 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381109902.611, "dur": 14.297 , "args": { "External id": 325,"Record function id": 0, "Ev Idx": 324 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381109902.842, "dur": 13.956 , "args": { "External id": 326,"Record function id": 0, "Ev Idx": 325 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 1728719, "tid": 1728719, "ts": 6950381109923.641, "dur": 4.067 , "args": { "External id": 327,"Record function id": 0, "Ev Idx": 326 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381109929.461, "dur": 10.039 , "args": { "External id": 328,"Record function id": 0, "Ev Idx": 327 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381109929.632, "dur": 0.701 , "args": { "External id": 329,"Record function id": 0, "Ev Idx": 328 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381109929.832, "dur": 0.411 , "args": { "External id": 330,"Record function id": 0, "Ev Idx": 329 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109930.032, "dur": 0.111 , "args": { "External id": 331,"Record function id": 0, "Ev Idx": 330 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381109930.573, "dur": 8.827 , "args": { "External id": 332,"Record function id": 0, "Ev Idx": 331 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381109930.754, "dur": 8.526 , "args": { "External id": 333,"Record function id": 0, "Ev Idx": 332 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_not", "pid": 1728719, "tid": 1728719, "ts": 6950381109942.295, "dur": 5.831 , "args": { "External id": 334,"Record function id": 0, "Ev Idx": 333 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1728719, "tid": 1728719, "ts": 6950381109949.689, "dur": 17.533 , "args": { "External id": 335,"Record function id": 0, "Ev Idx": 334 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1728719, "tid": 1728719, "ts": 6950381109950.460, "dur": 16.611 , "args": { "External id": 336,"Record function id": 0, "Ev Idx": 335 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109952.244, "dur": 0.140 , "args": { "External id": 337,"Record function id": 0, "Ev Idx": 336 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381109952.715, "dur": 7.634 , "args": { "External id": 338,"Record function id": 0, "Ev Idx": 337 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1728719, "tid": 1728719, "ts": 6950381109952.935, "dur": 7.324 , "args": { "External id": 339,"Record function id": 0, "Ev Idx": 338 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381109953.476, "dur": 1.553 , "args": { "External id": 340,"Record function id": 0, "Ev Idx": 339 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1728719, "tid": 1728719, "ts": 6950381109955.320, "dur": 4.819 , "args": { "External id": 341,"Record function id": 0, "Ev Idx": 340 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381109968.905, "dur": 76.472 , "args": { "External id": 342,"Record function id": 0, "Ev Idx": 341 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381109969.185, "dur": 76.052 , "args": { "External id": 343,"Record function id": 0, "Ev Idx": 342 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1728719, "tid": 1728719, "ts": 6950381110048.152, "dur": 0.912 , "args": { "External id": 344,"Record function id": 0, "Ev Idx": 343 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110048.643, "dur": 0.251 , "args": { "External id": 345,"Record function id": 0, "Ev Idx": 344 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1728719, "tid": 1728719, "ts": 6950381110050.557, "dur": 0.521 , "args": { "External id": 346,"Record function id": 0, "Ev Idx": 345 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110050.807, "dur": 0.151 , "args": { "External id": 347,"Record function id": 0, "Ev Idx": 346 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1728719, "tid": 1728719, "ts": 6950381110052.080, "dur": 0.471 , "args": { "External id": 348,"Record function id": 0, "Ev Idx": 347 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110052.350, "dur": 0.090 , "args": { "External id": 349,"Record function id": 0, "Ev Idx": 348 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381110083.688, "dur": 75.060 , "args": { "External id": 350,"Record function id": 0, "Ev Idx": 349 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/2", "pid": 1728719, "tid": 1728719, "ts": 6950381110159.540, "dur": 575.611 , "args": { "External id": 351,"Record function id": 0, "Ev Idx": 350 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381110178.004, "dur": 17.923 , "args": { "External id": 352,"Record function id": 0, "Ev Idx": 351 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110186.410, "dur": 0.871 , "args": { "External id": 353,"Record function id": 0, "Ev Idx": 352 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381110188.133, "dur": 0.110 , "args": { "External id": 354,"Record function id": 0, "Ev Idx": 353 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381110188.814, "dur": 0.080 , "args": { "External id": 355,"Record function id": 0, "Ev Idx": 354 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110190.527, "dur": 0.301 , "args": { "External id": 356,"Record function id": 0, "Ev Idx": 355 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381110191.169, "dur": 0.030 , "args": { "External id": 357,"Record function id": 0, "Ev Idx": 356 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381110191.599, "dur": 0.030 , "args": { "External id": 358,"Record function id": 0, "Ev Idx": 357 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381110202.820, "dur": 2.475 , "args": { "External id": 359,"Record function id": 0, "Ev Idx": 358 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fctolslp5dyuqahezjtwlujwmt4gv74ywim3eqdqgrrjgix4fv3a ##", "pid": 1728719, "tid": 1728719, "ts": 6950381110211.046, "dur": 516.531 , "args": { "External id": 360,"Record function id": 0, "Ev Idx": 359 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110245.610, "dur": 19.256 , "args": { "External id": 361,"Record function id": 0, "Ev Idx": 360 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110282.919, "dur": 16.140 , "args": { "External id": 362,"Record function id": 0, "Ev Idx": 361 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110308.146, "dur": 7.284 , "args": { "External id": 363,"Record function id": 0, "Ev Idx": 362 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110322.784, "dur": 6.111 , "args": { "External id": 364,"Record function id": 0, "Ev Idx": 363 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110333.694, "dur": 23.223 , "args": { "External id": 365,"Record function id": 0, "Ev Idx": 364 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110334.856, "dur": 1.373 , "args": { "External id": 366,"Record function id": 0, "Ev Idx": 365 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110335.177, "dur": 0.931 , "args": { "External id": 367,"Record function id": 0, "Ev Idx": 366 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110335.738, "dur": 0.230 , "args": { "External id": 368,"Record function id": 0, "Ev Idx": 367 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110336.709, "dur": 0.702 , "args": { "External id": 369,"Record function id": 0, "Ev Idx": 368 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110336.870, "dur": 0.461 , "args": { "External id": 370,"Record function id": 0, "Ev Idx": 369 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110337.130, "dur": 0.090 , "args": { "External id": 371,"Record function id": 0, "Ev Idx": 370 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110338.042, "dur": 0.751 , "args": { "External id": 372,"Record function id": 0, "Ev Idx": 371 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110338.202, "dur": 0.501 , "args": { "External id": 373,"Record function id": 0, "Ev Idx": 372 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110338.383, "dur": 0.220 , "args": { "External id": 374,"Record function id": 0, "Ev Idx": 373 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110342.300, "dur": 1.873 , "args": { "External id": 375,"Record function id": 0, "Ev Idx": 374 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110342.701, "dur": 1.332 , "args": { "External id": 376,"Record function id": 0, "Ev Idx": 375 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381110369.330, "dur": 8.666 , "args": { "External id": 377,"Record function id": 0, "Ev Idx": 376 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110385.059, "dur": 2.355 , "args": { "External id": 378,"Record function id": 0, "Ev Idx": 377 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110391.181, "dur": 0.841 , "args": { "External id": 379,"Record function id": 0, "Ev Idx": 378 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381110397.473, "dur": 0.891 , "args": { "External id": 380,"Record function id": 0, "Ev Idx": 379 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381110397.783, "dur": 0.461 , "args": { "External id": 381,"Record function id": 0, "Ev Idx": 380 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381110406.860, "dur": 12.293 , "args": { "External id": 382,"Record function id": 0, "Ev Idx": 381 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110422.149, "dur": 12.292 , "args": { "External id": 383,"Record function id": 0, "Ev Idx": 382 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110422.840, "dur": 0.831 , "args": { "External id": 384,"Record function id": 0, "Ev Idx": 383 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110423.050, "dur": 0.531 , "args": { "External id": 385,"Record function id": 0, "Ev Idx": 384 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110423.341, "dur": 0.110 , "args": { "External id": 386,"Record function id": 0, "Ev Idx": 385 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110423.972, "dur": 0.721 , "args": { "External id": 387,"Record function id": 0, "Ev Idx": 386 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110424.192, "dur": 0.401 , "args": { "External id": 388,"Record function id": 0, "Ev Idx": 387 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110424.403, "dur": 0.090 , "args": { "External id": 389,"Record function id": 0, "Ev Idx": 388 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110424.944, "dur": 0.681 , "args": { "External id": 390,"Record function id": 0, "Ev Idx": 389 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110425.094, "dur": 0.441 , "args": { "External id": 391,"Record function id": 0, "Ev Idx": 390 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110425.304, "dur": 0.121 , "args": { "External id": 392,"Record function id": 0, "Ev Idx": 391 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110426.947, "dur": 1.403 , "args": { "External id": 393,"Record function id": 0, "Ev Idx": 392 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110427.178, "dur": 1.062 , "args": { "External id": 394,"Record function id": 0, "Ev Idx": 393 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381110438.589, "dur": 4.809 , "args": { "External id": 395,"Record function id": 0, "Ev Idx": 394 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110444.731, "dur": 0.671 , "args": { "External id": 396,"Record function id": 0, "Ev Idx": 395 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110446.915, "dur": 0.601 , "args": { "External id": 397,"Record function id": 0, "Ev Idx": 396 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381110452.395, "dur": 0.421 , "args": { "External id": 398,"Record function id": 0, "Ev Idx": 397 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381110452.665, "dur": 0.050 , "args": { "External id": 399,"Record function id": 0, "Ev Idx": 398 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381110456.122, "dur": 7.544 , "args": { "External id": 400,"Record function id": 0, "Ev Idx": 399 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110466.341, "dur": 11.010 , "args": { "External id": 401,"Record function id": 0, "Ev Idx": 400 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110466.892, "dur": 0.811 , "args": { "External id": 402,"Record function id": 0, "Ev Idx": 401 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110467.092, "dur": 0.521 , "args": { "External id": 403,"Record function id": 0, "Ev Idx": 402 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110467.393, "dur": 0.110 , "args": { "External id": 404,"Record function id": 0, "Ev Idx": 403 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110467.994, "dur": 0.681 , "args": { "External id": 405,"Record function id": 0, "Ev Idx": 404 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110468.194, "dur": 0.391 , "args": { "External id": 406,"Record function id": 0, "Ev Idx": 405 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110468.405, "dur": 0.080 , "args": { "External id": 407,"Record function id": 0, "Ev Idx": 406 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110468.926, "dur": 0.771 , "args": { "External id": 408,"Record function id": 0, "Ev Idx": 407 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110469.096, "dur": 0.511 , "args": { "External id": 409,"Record function id": 0, "Ev Idx": 408 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110469.286, "dur": 0.221 , "args": { "External id": 410,"Record function id": 0, "Ev Idx": 409 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110470.789, "dur": 1.312 , "args": { "External id": 411,"Record function id": 0, "Ev Idx": 410 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110470.999, "dur": 0.992 , "args": { "External id": 412,"Record function id": 0, "Ev Idx": 411 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381110486.298, "dur": 7.273 , "args": { "External id": 413,"Record function id": 0, "Ev Idx": 412 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381110504.061, "dur": 7.534 , "args": { "External id": 414,"Record function id": 0, "Ev Idx": 413 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381110520.822, "dur": 7.364 , "args": { "External id": 415,"Record function id": 0, "Ev Idx": 414 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381110532.995, "dur": 29.795 , "args": { "External id": 416,"Record function id": 0, "Ev Idx": 415 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110533.957, "dur": 0.821 , "args": { "External id": 417,"Record function id": 0, "Ev Idx": 416 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110534.317, "dur": 0.311 , "args": { "External id": 418,"Record function id": 0, "Ev Idx": 417 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110535.049, "dur": 0.430 , "args": { "External id": 419,"Record function id": 0, "Ev Idx": 418 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110535.259, "dur": 0.110 , "args": { "External id": 420,"Record function id": 0, "Ev Idx": 419 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110535.730, "dur": 0.481 , "args": { "External id": 421,"Record function id": 0, "Ev Idx": 420 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110536.020, "dur": 0.091 , "args": { "External id": 422,"Record function id": 0, "Ev Idx": 421 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381110541.651, "dur": 18.695 , "args": { "External id": 423,"Record function id": 0, "Ev Idx": 422 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381110543.504, "dur": 2.565 , "args": { "External id": 424,"Record function id": 0, "Ev Idx": 423 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110543.855, "dur": 2.064 , "args": { "External id": 425,"Record function id": 0, "Ev Idx": 424 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110546.720, "dur": 1.583 , "args": { "External id": 426,"Record function id": 0, "Ev Idx": 425 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110549.275, "dur": 1.483 , "args": { "External id": 427,"Record function id": 0, "Ev Idx": 426 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110551.089, "dur": 1.582 , "args": { "External id": 428,"Record function id": 0, "Ev Idx": 427 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110558.823, "dur": 1.002 , "args": { "External id": 429,"Record function id": 0, "Ev Idx": 428 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110560.807, "dur": 0.541 , "args": { "External id": 430,"Record function id": 0, "Ev Idx": 429 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110561.037, "dur": 0.200 , "args": { "External id": 431,"Record function id": 0, "Ev Idx": 430 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110573.510, "dur": 9.027 , "args": { "External id": 432,"Record function id": 0, "Ev Idx": 431 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110585.392, "dur": 12.243 , "args": { "External id": 433,"Record function id": 0, "Ev Idx": 432 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110586.104, "dur": 0.791 , "args": { "External id": 434,"Record function id": 0, "Ev Idx": 433 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110586.314, "dur": 0.491 , "args": { "External id": 435,"Record function id": 0, "Ev Idx": 434 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110586.555, "dur": 0.120 , "args": { "External id": 436,"Record function id": 0, "Ev Idx": 435 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110587.246, "dur": 0.731 , "args": { "External id": 437,"Record function id": 0, "Ev Idx": 436 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110587.456, "dur": 0.431 , "args": { "External id": 438,"Record function id": 0, "Ev Idx": 437 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110587.667, "dur": 0.100 , "args": { "External id": 439,"Record function id": 0, "Ev Idx": 438 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110588.248, "dur": 0.641 , "args": { "External id": 440,"Record function id": 0, "Ev Idx": 439 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110588.408, "dur": 0.391 , "args": { "External id": 441,"Record function id": 0, "Ev Idx": 440 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110588.618, "dur": 0.081 , "args": { "External id": 442,"Record function id": 0, "Ev Idx": 441 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110590.121, "dur": 1.353 , "args": { "External id": 443,"Record function id": 0, "Ev Idx": 442 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110590.352, "dur": 1.001 , "args": { "External id": 444,"Record function id": 0, "Ev Idx": 443 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381110606.011, "dur": 8.315 , "args": { "External id": 445,"Record function id": 0, "Ev Idx": 444 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110618.624, "dur": 6.352 , "args": { "External id": 446,"Record function id": 0, "Ev Idx": 445 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110629.034, "dur": 5.640 , "args": { "External id": 447,"Record function id": 0, "Ev Idx": 446 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110639.213, "dur": 10.910 , "args": { "External id": 448,"Record function id": 0, "Ev Idx": 447 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110639.734, "dur": 0.741 , "args": { "External id": 449,"Record function id": 0, "Ev Idx": 448 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110639.924, "dur": 0.451 , "args": { "External id": 450,"Record function id": 0, "Ev Idx": 449 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110640.164, "dur": 0.091 , "args": { "External id": 451,"Record function id": 0, "Ev Idx": 450 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110640.765, "dur": 0.662 , "args": { "External id": 452,"Record function id": 0, "Ev Idx": 451 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110640.936, "dur": 0.401 , "args": { "External id": 453,"Record function id": 0, "Ev Idx": 452 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110641.166, "dur": 0.070 , "args": { "External id": 454,"Record function id": 0, "Ev Idx": 453 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110641.677, "dur": 0.641 , "args": { "External id": 455,"Record function id": 0, "Ev Idx": 454 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110641.837, "dur": 0.391 , "args": { "External id": 456,"Record function id": 0, "Ev Idx": 455 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110642.028, "dur": 0.090 , "args": { "External id": 457,"Record function id": 0, "Ev Idx": 456 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110643.420, "dur": 1.283 , "args": { "External id": 458,"Record function id": 0, "Ev Idx": 457 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110643.611, "dur": 0.982 , "args": { "External id": 459,"Record function id": 0, "Ev Idx": 458 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110654.641, "dur": 10.029 , "args": { "External id": 460,"Record function id": 0, "Ev Idx": 459 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110655.132, "dur": 0.641 , "args": { "External id": 461,"Record function id": 0, "Ev Idx": 460 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110655.292, "dur": 0.391 , "args": { "External id": 462,"Record function id": 0, "Ev Idx": 461 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110655.503, "dur": 0.070 , "args": { "External id": 463,"Record function id": 0, "Ev Idx": 462 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110656.034, "dur": 0.641 , "args": { "External id": 464,"Record function id": 0, "Ev Idx": 463 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110656.194, "dur": 0.391 , "args": { "External id": 465,"Record function id": 0, "Ev Idx": 464 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110656.405, "dur": 0.070 , "args": { "External id": 466,"Record function id": 0, "Ev Idx": 465 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110656.926, "dur": 0.641 , "args": { "External id": 467,"Record function id": 0, "Ev Idx": 466 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110657.096, "dur": 0.381 , "args": { "External id": 468,"Record function id": 0, "Ev Idx": 467 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110657.296, "dur": 0.080 , "args": { "External id": 469,"Record function id": 0, "Ev Idx": 468 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110658.559, "dur": 1.242 , "args": { "External id": 470,"Record function id": 0, "Ev Idx": 469 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110658.749, "dur": 0.952 , "args": { "External id": 471,"Record function id": 0, "Ev Idx": 470 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381110673.015, "dur": 7.224 , "args": { "External id": 472,"Record function id": 0, "Ev Idx": 471 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381110687.162, "dur": 7.013 , "args": { "External id": 473,"Record function id": 0, "Ev Idx": 472 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110698.493, "dur": 10.379 , "args": { "External id": 474,"Record function id": 0, "Ev Idx": 473 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110699.004, "dur": 0.761 , "args": { "External id": 475,"Record function id": 0, "Ev Idx": 474 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110699.174, "dur": 0.501 , "args": { "External id": 476,"Record function id": 0, "Ev Idx": 475 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110699.495, "dur": 0.080 , "args": { "External id": 477,"Record function id": 0, "Ev Idx": 476 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110700.036, "dur": 0.631 , "args": { "External id": 478,"Record function id": 0, "Ev Idx": 477 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110700.196, "dur": 0.381 , "args": { "External id": 479,"Record function id": 0, "Ev Idx": 478 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110700.406, "dur": 0.071 , "args": { "External id": 480,"Record function id": 0, "Ev Idx": 479 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110700.927, "dur": 0.642 , "args": { "External id": 481,"Record function id": 0, "Ev Idx": 480 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110701.098, "dur": 0.380 , "args": { "External id": 482,"Record function id": 0, "Ev Idx": 481 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110701.298, "dur": 0.080 , "args": { "External id": 483,"Record function id": 0, "Ev Idx": 482 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110702.631, "dur": 1.262 , "args": { "External id": 484,"Record function id": 0, "Ev Idx": 483 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110702.821, "dur": 0.972 , "args": { "External id": 485,"Record function id": 0, "Ev Idx": 484 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381110717.478, "dur": 7.404 , "args": { "External id": 486,"Record function id": 0, "Ev Idx": 485 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381110759.436, "dur": 49.112 , "args": { "External id": 487,"Record function id": 0, "Ev Idx": 486 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381110809.039, "dur": 443.054 , "args": { "External id": 488,"Record function id": 0, "Ev Idx": 487 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381110817.224, "dur": 13.435 , "args": { "External id": 489,"Record function id": 0, "Ev Idx": 488 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110821.983, "dur": 0.551 , "args": { "External id": 490,"Record function id": 0, "Ev Idx": 489 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381110823.125, "dur": 0.040 , "args": { "External id": 491,"Record function id": 0, "Ev Idx": 490 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381110823.706, "dur": 0.070 , "args": { "External id": 492,"Record function id": 0, "Ev Idx": 491 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110825.339, "dur": 0.361 , "args": { "External id": 493,"Record function id": 0, "Ev Idx": 492 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381110826.060, "dur": 0.020 , "args": { "External id": 494,"Record function id": 0, "Ev Idx": 493 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381110826.451, "dur": 0.020 , "args": { "External id": 495,"Record function id": 0, "Ev Idx": 494 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381110835.288, "dur": 1.793 , "args": { "External id": 496,"Record function id": 0, "Ev Idx": 495 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381110840.708, "dur": 406.516 , "args": { "External id": 497,"Record function id": 0, "Ev Idx": 496 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110856.237, "dur": 10.178 , "args": { "External id": 498,"Record function id": 0, "Ev Idx": 497 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110876.935, "dur": 9.287 , "args": { "External id": 499,"Record function id": 0, "Ev Idx": 498 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110893.005, "dur": 6.031 , "args": { "External id": 500,"Record function id": 0, "Ev Idx": 499 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381110905.248, "dur": 5.480 , "args": { "External id": 501,"Record function id": 0, "Ev Idx": 500 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110913.663, "dur": 15.519 , "args": { "External id": 502,"Record function id": 0, "Ev Idx": 501 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110914.405, "dur": 0.972 , "args": { "External id": 503,"Record function id": 0, "Ev Idx": 502 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110914.625, "dur": 0.641 , "args": { "External id": 504,"Record function id": 0, "Ev Idx": 503 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110914.956, "dur": 0.190 , "args": { "External id": 505,"Record function id": 0, "Ev Idx": 504 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110915.697, "dur": 0.742 , "args": { "External id": 506,"Record function id": 0, "Ev Idx": 505 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110915.938, "dur": 0.410 , "args": { "External id": 507,"Record function id": 0, "Ev Idx": 506 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110916.158, "dur": 0.080 , "args": { "External id": 508,"Record function id": 0, "Ev Idx": 507 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110916.709, "dur": 0.681 , "args": { "External id": 509,"Record function id": 0, "Ev Idx": 508 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110916.859, "dur": 0.441 , "args": { "External id": 510,"Record function id": 0, "Ev Idx": 509 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110917.040, "dur": 0.160 , "args": { "External id": 511,"Record function id": 0, "Ev Idx": 510 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110918.853, "dur": 3.787 , "args": { "External id": 512,"Record function id": 0, "Ev Idx": 511 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110921.318, "dur": 1.202 , "args": { "External id": 513,"Record function id": 0, "Ev Idx": 512 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381110938.961, "dur": 6.592 , "args": { "External id": 514,"Record function id": 0, "Ev Idx": 513 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110948.448, "dur": 1.122 , "args": { "External id": 515,"Record function id": 0, "Ev Idx": 514 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110951.965, "dur": 0.591 , "args": { "External id": 516,"Record function id": 0, "Ev Idx": 515 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381110956.854, "dur": 0.611 , "args": { "External id": 517,"Record function id": 0, "Ev Idx": 516 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381110957.204, "dur": 0.151 , "args": { "External id": 518,"Record function id": 0, "Ev Idx": 517 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381110963.717, "dur": 9.387 , "args": { "External id": 519,"Record function id": 0, "Ev Idx": 518 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381110975.509, "dur": 11.271 , "args": { "External id": 520,"Record function id": 0, "Ev Idx": 519 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110976.090, "dur": 0.761 , "args": { "External id": 521,"Record function id": 0, "Ev Idx": 520 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110976.290, "dur": 0.461 , "args": { "External id": 522,"Record function id": 0, "Ev Idx": 521 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110976.530, "dur": 0.111 , "args": { "External id": 523,"Record function id": 0, "Ev Idx": 522 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110977.142, "dur": 0.641 , "args": { "External id": 524,"Record function id": 0, "Ev Idx": 523 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110977.302, "dur": 0.391 , "args": { "External id": 525,"Record function id": 0, "Ev Idx": 524 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110977.522, "dur": 0.070 , "args": { "External id": 526,"Record function id": 0, "Ev Idx": 525 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381110978.033, "dur": 1.062 , "args": { "External id": 527,"Record function id": 0, "Ev Idx": 526 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381110978.194, "dur": 0.811 , "args": { "External id": 528,"Record function id": 0, "Ev Idx": 527 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381110978.674, "dur": 0.231 , "args": { "External id": 529,"Record function id": 0, "Ev Idx": 528 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110980.247, "dur": 1.303 , "args": { "External id": 530,"Record function id": 0, "Ev Idx": 529 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381110980.438, "dur": 1.012 , "args": { "External id": 531,"Record function id": 0, "Ev Idx": 530 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381110990.516, "dur": 4.298 , "args": { "External id": 532,"Record function id": 0, "Ev Idx": 531 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110996.057, "dur": 0.611 , "args": { "External id": 533,"Record function id": 0, "Ev Idx": 532 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381110998.131, "dur": 0.581 , "args": { "External id": 534,"Record function id": 0, "Ev Idx": 533 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381111003.060, "dur": 0.350 , "args": { "External id": 535,"Record function id": 0, "Ev Idx": 534 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381111003.270, "dur": 0.040 , "args": { "External id": 536,"Record function id": 0, "Ev Idx": 535 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381111006.166, "dur": 6.832 , "args": { "External id": 537,"Record function id": 0, "Ev Idx": 536 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111015.263, "dur": 10.569 , "args": { "External id": 538,"Record function id": 0, "Ev Idx": 537 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111015.834, "dur": 0.841 , "args": { "External id": 539,"Record function id": 0, "Ev Idx": 538 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111016.044, "dur": 0.541 , "args": { "External id": 540,"Record function id": 0, "Ev Idx": 539 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111016.365, "dur": 0.110 , "args": { "External id": 541,"Record function id": 0, "Ev Idx": 540 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111016.956, "dur": 0.661 , "args": { "External id": 542,"Record function id": 0, "Ev Idx": 541 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111017.136, "dur": 0.391 , "args": { "External id": 543,"Record function id": 0, "Ev Idx": 542 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111017.336, "dur": 0.081 , "args": { "External id": 544,"Record function id": 0, "Ev Idx": 543 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111017.857, "dur": 0.642 , "args": { "External id": 545,"Record function id": 0, "Ev Idx": 544 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111018.018, "dur": 0.390 , "args": { "External id": 546,"Record function id": 0, "Ev Idx": 545 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111018.208, "dur": 0.090 , "args": { "External id": 547,"Record function id": 0, "Ev Idx": 546 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111019.571, "dur": 1.272 , "args": { "External id": 548,"Record function id": 0, "Ev Idx": 547 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111019.771, "dur": 0.972 , "args": { "External id": 549,"Record function id": 0, "Ev Idx": 548 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381111033.577, "dur": 5.820 , "args": { "External id": 550,"Record function id": 0, "Ev Idx": 549 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381111048.985, "dur": 6.563 , "args": { "External id": 551,"Record function id": 0, "Ev Idx": 550 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381111064.364, "dur": 6.823 , "args": { "External id": 552,"Record function id": 0, "Ev Idx": 551 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381111074.583, "dur": 20.428 , "args": { "External id": 553,"Record function id": 0, "Ev Idx": 552 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111075.144, "dur": 0.631 , "args": { "External id": 554,"Record function id": 0, "Ev Idx": 553 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111075.415, "dur": 0.240 , "args": { "External id": 555,"Record function id": 0, "Ev Idx": 554 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111076.026, "dur": 0.621 , "args": { "External id": 556,"Record function id": 0, "Ev Idx": 555 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111076.426, "dur": 0.111 , "args": { "External id": 557,"Record function id": 0, "Ev Idx": 556 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111076.887, "dur": 0.702 , "args": { "External id": 558,"Record function id": 0, "Ev Idx": 557 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111077.388, "dur": 0.080 , "args": { "External id": 559,"Record function id": 0, "Ev Idx": 558 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381111078.039, "dur": 15.129 , "args": { "External id": 560,"Record function id": 0, "Ev Idx": 559 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381111079.312, "dur": 2.174 , "args": { "External id": 561,"Record function id": 0, "Ev Idx": 560 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111079.632, "dur": 1.714 , "args": { "External id": 562,"Record function id": 0, "Ev Idx": 561 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111082.017, "dur": 1.362 , "args": { "External id": 563,"Record function id": 0, "Ev Idx": 562 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111083.990, "dur": 1.493 , "args": { "External id": 564,"Record function id": 0, "Ev Idx": 563 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111085.754, "dur": 1.483 , "args": { "External id": 565,"Record function id": 0, "Ev Idx": 564 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111091.755, "dur": 1.032 , "args": { "External id": 566,"Record function id": 0, "Ev Idx": 565 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111093.458, "dur": 0.651 , "args": { "External id": 567,"Record function id": 0, "Ev Idx": 566 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111093.669, "dur": 0.330 , "args": { "External id": 568,"Record function id": 0, "Ev Idx": 567 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111103.607, "dur": 7.484 , "args": { "External id": 569,"Record function id": 0, "Ev Idx": 568 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111113.616, "dur": 10.920 , "args": { "External id": 570,"Record function id": 0, "Ev Idx": 569 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111114.267, "dur": 0.701 , "args": { "External id": 571,"Record function id": 0, "Ev Idx": 570 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111114.457, "dur": 0.421 , "args": { "External id": 572,"Record function id": 0, "Ev Idx": 571 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111114.678, "dur": 0.090 , "args": { "External id": 573,"Record function id": 0, "Ev Idx": 572 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111115.269, "dur": 0.651 , "args": { "External id": 574,"Record function id": 0, "Ev Idx": 573 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111115.419, "dur": 0.411 , "args": { "External id": 575,"Record function id": 0, "Ev Idx": 574 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111115.639, "dur": 0.091 , "args": { "External id": 576,"Record function id": 0, "Ev Idx": 575 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111116.200, "dur": 0.642 , "args": { "External id": 577,"Record function id": 0, "Ev Idx": 576 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111116.371, "dur": 0.380 , "args": { "External id": 578,"Record function id": 0, "Ev Idx": 577 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111116.581, "dur": 0.070 , "args": { "External id": 579,"Record function id": 0, "Ev Idx": 578 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111117.974, "dur": 1.282 , "args": { "External id": 580,"Record function id": 0, "Ev Idx": 579 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111118.184, "dur": 0.972 , "args": { "External id": 581,"Record function id": 0, "Ev Idx": 580 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381111132.100, "dur": 8.396 , "args": { "External id": 582,"Record function id": 0, "Ev Idx": 581 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111143.902, "dur": 5.691 , "args": { "External id": 583,"Record function id": 0, "Ev Idx": 582 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111153.310, "dur": 5.860 , "args": { "External id": 584,"Record function id": 0, "Ev Idx": 583 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111163.338, "dur": 10.640 , "args": { "External id": 585,"Record function id": 0, "Ev Idx": 584 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111163.889, "dur": 0.712 , "args": { "External id": 586,"Record function id": 0, "Ev Idx": 585 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111164.060, "dur": 0.440 , "args": { "External id": 587,"Record function id": 0, "Ev Idx": 586 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111164.290, "dur": 0.100 , "args": { "External id": 588,"Record function id": 0, "Ev Idx": 587 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111164.891, "dur": 0.651 , "args": { "External id": 589,"Record function id": 0, "Ev Idx": 588 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111165.041, "dur": 0.411 , "args": { "External id": 590,"Record function id": 0, "Ev Idx": 589 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111165.282, "dur": 0.070 , "args": { "External id": 591,"Record function id": 0, "Ev Idx": 590 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111165.803, "dur": 0.641 , "args": { "External id": 592,"Record function id": 0, "Ev Idx": 591 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111165.963, "dur": 0.391 , "args": { "External id": 593,"Record function id": 0, "Ev Idx": 592 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111166.173, "dur": 0.081 , "args": { "External id": 594,"Record function id": 0, "Ev Idx": 593 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111167.566, "dur": 1.312 , "args": { "External id": 595,"Record function id": 0, "Ev Idx": 594 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111167.796, "dur": 0.972 , "args": { "External id": 596,"Record function id": 0, "Ev Idx": 595 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111178.005, "dur": 9.999 , "args": { "External id": 597,"Record function id": 0, "Ev Idx": 596 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111178.506, "dur": 0.642 , "args": { "External id": 598,"Record function id": 0, "Ev Idx": 597 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111178.677, "dur": 0.380 , "args": { "External id": 599,"Record function id": 0, "Ev Idx": 598 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111178.877, "dur": 0.070 , "args": { "External id": 600,"Record function id": 0, "Ev Idx": 599 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111179.408, "dur": 0.631 , "args": { "External id": 601,"Record function id": 0, "Ev Idx": 600 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111179.568, "dur": 0.381 , "args": { "External id": 602,"Record function id": 0, "Ev Idx": 601 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111179.779, "dur": 0.070 , "args": { "External id": 603,"Record function id": 0, "Ev Idx": 602 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111180.300, "dur": 0.621 , "args": { "External id": 604,"Record function id": 0, "Ev Idx": 603 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111180.460, "dur": 0.371 , "args": { "External id": 605,"Record function id": 0, "Ev Idx": 604 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111180.650, "dur": 0.081 , "args": { "External id": 606,"Record function id": 0, "Ev Idx": 605 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111181.923, "dur": 1.222 , "args": { "External id": 607,"Record function id": 0, "Ev Idx": 606 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111182.093, "dur": 0.942 , "args": { "External id": 608,"Record function id": 0, "Ev Idx": 607 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381111195.258, "dur": 5.680 , "args": { "External id": 609,"Record function id": 0, "Ev Idx": 608 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381111207.240, "dur": 6.632 , "args": { "External id": 610,"Record function id": 0, "Ev Idx": 609 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111217.349, "dur": 10.159 , "args": { "External id": 611,"Record function id": 0, "Ev Idx": 610 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111217.860, "dur": 0.681 , "args": { "External id": 612,"Record function id": 0, "Ev Idx": 611 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111218.040, "dur": 0.411 , "args": { "External id": 613,"Record function id": 0, "Ev Idx": 612 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111218.250, "dur": 0.091 , "args": { "External id": 614,"Record function id": 0, "Ev Idx": 613 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111218.811, "dur": 0.632 , "args": { "External id": 615,"Record function id": 0, "Ev Idx": 614 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111218.982, "dur": 0.370 , "args": { "External id": 616,"Record function id": 0, "Ev Idx": 615 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111219.172, "dur": 0.070 , "args": { "External id": 617,"Record function id": 0, "Ev Idx": 616 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111219.703, "dur": 0.651 , "args": { "External id": 618,"Record function id": 0, "Ev Idx": 617 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111219.853, "dur": 0.411 , "args": { "External id": 619,"Record function id": 0, "Ev Idx": 618 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111220.084, "dur": 0.070 , "args": { "External id": 620,"Record function id": 0, "Ev Idx": 619 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111221.416, "dur": 1.253 , "args": { "External id": 621,"Record function id": 0, "Ev Idx": 620 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111221.617, "dur": 0.951 , "args": { "External id": 622,"Record function id": 0, "Ev Idx": 621 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381111235.693, "dur": 9.347 , "args": { "External id": 623,"Record function id": 0, "Ev Idx": 622 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381111272.271, "dur": 29.044 , "args": { "External id": 624,"Record function id": 0, "Ev Idx": 623 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381111301.706, "dur": 372.774 , "args": { "External id": 625,"Record function id": 0, "Ev Idx": 624 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381111307.727, "dur": 12.363 , "args": { "External id": 626,"Record function id": 0, "Ev Idx": 625 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111312.045, "dur": 0.491 , "args": { "External id": 627,"Record function id": 0, "Ev Idx": 626 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381111313.117, "dur": 0.050 , "args": { "External id": 628,"Record function id": 0, "Ev Idx": 627 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381111313.588, "dur": 0.060 , "args": { "External id": 629,"Record function id": 0, "Ev Idx": 628 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111315.171, "dur": 0.290 , "args": { "External id": 630,"Record function id": 0, "Ev Idx": 629 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381111315.822, "dur": 0.040 , "args": { "External id": 631,"Record function id": 0, "Ev Idx": 630 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381111316.223, "dur": 0.020 , "args": { "External id": 632,"Record function id": 0, "Ev Idx": 631 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381111323.466, "dur": 1.263 , "args": { "External id": 633,"Record function id": 0, "Ev Idx": 632 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381111327.774, "dur": 343.019 , "args": { "External id": 634,"Record function id": 0, "Ev Idx": 633 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111338.374, "dur": 7.394 , "args": { "External id": 635,"Record function id": 0, "Ev Idx": 634 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111353.071, "dur": 8.075 , "args": { "External id": 636,"Record function id": 0, "Ev Idx": 635 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111367.618, "dur": 5.691 , "args": { "External id": 637,"Record function id": 0, "Ev Idx": 636 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111378.950, "dur": 5.410 , "args": { "External id": 638,"Record function id": 0, "Ev Idx": 637 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111387.065, "dur": 12.623 , "args": { "External id": 639,"Record function id": 0, "Ev Idx": 638 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111387.726, "dur": 1.072 , "args": { "External id": 640,"Record function id": 0, "Ev Idx": 639 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111387.936, "dur": 0.752 , "args": { "External id": 641,"Record function id": 0, "Ev Idx": 640 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111388.237, "dur": 0.310 , "args": { "External id": 642,"Record function id": 0, "Ev Idx": 641 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111389.118, "dur": 0.642 , "args": { "External id": 643,"Record function id": 0, "Ev Idx": 642 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111389.269, "dur": 0.390 , "args": { "External id": 644,"Record function id": 0, "Ev Idx": 643 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111389.449, "dur": 0.100 , "args": { "External id": 645,"Record function id": 0, "Ev Idx": 644 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111390.030, "dur": 0.732 , "args": { "External id": 646,"Record function id": 0, "Ev Idx": 645 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111390.190, "dur": 0.481 , "args": { "External id": 647,"Record function id": 0, "Ev Idx": 646 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111390.381, "dur": 0.180 , "args": { "External id": 648,"Record function id": 0, "Ev Idx": 647 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111392.134, "dur": 1.413 , "args": { "External id": 649,"Record function id": 0, "Ev Idx": 648 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111392.344, "dur": 1.103 , "args": { "External id": 650,"Record function id": 0, "Ev Idx": 649 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381111406.030, "dur": 4.889 , "args": { "External id": 651,"Record function id": 0, "Ev Idx": 650 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111413.494, "dur": 0.982 , "args": { "External id": 652,"Record function id": 0, "Ev Idx": 651 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111416.580, "dur": 0.581 , "args": { "External id": 653,"Record function id": 0, "Ev Idx": 652 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381111421.228, "dur": 0.501 , "args": { "External id": 654,"Record function id": 0, "Ev Idx": 653 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381111421.509, "dur": 0.110 , "args": { "External id": 655,"Record function id": 0, "Ev Idx": 654 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381111424.875, "dur": 7.344 , "args": { "External id": 656,"Record function id": 0, "Ev Idx": 655 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111434.323, "dur": 10.790 , "args": { "External id": 657,"Record function id": 0, "Ev Idx": 656 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111434.884, "dur": 0.791 , "args": { "External id": 658,"Record function id": 0, "Ev Idx": 657 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111435.084, "dur": 0.501 , "args": { "External id": 659,"Record function id": 0, "Ev Idx": 658 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111435.365, "dur": 0.110 , "args": { "External id": 660,"Record function id": 0, "Ev Idx": 659 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111435.966, "dur": 0.641 , "args": { "External id": 661,"Record function id": 0, "Ev Idx": 660 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111436.136, "dur": 0.381 , "args": { "External id": 662,"Record function id": 0, "Ev Idx": 661 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111436.336, "dur": 0.081 , "args": { "External id": 663,"Record function id": 0, "Ev Idx": 662 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111436.867, "dur": 0.702 , "args": { "External id": 664,"Record function id": 0, "Ev Idx": 663 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111437.018, "dur": 0.460 , "args": { "External id": 665,"Record function id": 0, "Ev Idx": 664 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111437.208, "dur": 0.170 , "args": { "External id": 666,"Record function id": 0, "Ev Idx": 665 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111438.691, "dur": 1.292 , "args": { "External id": 667,"Record function id": 0, "Ev Idx": 666 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111438.901, "dur": 0.982 , "args": { "External id": 668,"Record function id": 0, "Ev Idx": 667 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381111448.509, "dur": 4.068 , "args": { "External id": 669,"Record function id": 0, "Ev Idx": 668 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111453.649, "dur": 0.621 , "args": { "External id": 670,"Record function id": 0, "Ev Idx": 669 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111455.672, "dur": 0.541 , "args": { "External id": 671,"Record function id": 0, "Ev Idx": 670 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381111460.071, "dur": 0.330 , "args": { "External id": 672,"Record function id": 0, "Ev Idx": 671 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381111460.261, "dur": 0.040 , "args": { "External id": 673,"Record function id": 0, "Ev Idx": 672 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381111463.236, "dur": 6.573 , "args": { "External id": 674,"Record function id": 0, "Ev Idx": 673 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111472.093, "dur": 10.499 , "args": { "External id": 675,"Record function id": 0, "Ev Idx": 674 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111472.634, "dur": 0.731 , "args": { "External id": 676,"Record function id": 0, "Ev Idx": 675 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111472.824, "dur": 0.451 , "args": { "External id": 677,"Record function id": 0, "Ev Idx": 676 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111473.065, "dur": 0.100 , "args": { "External id": 678,"Record function id": 0, "Ev Idx": 677 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111473.636, "dur": 0.611 , "args": { "External id": 679,"Record function id": 0, "Ev Idx": 678 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111473.796, "dur": 0.361 , "args": { "External id": 680,"Record function id": 0, "Ev Idx": 679 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111473.976, "dur": 0.081 , "args": { "External id": 681,"Record function id": 0, "Ev Idx": 680 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111474.507, "dur": 0.832 , "args": { "External id": 682,"Record function id": 0, "Ev Idx": 681 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111474.658, "dur": 0.591 , "args": { "External id": 683,"Record function id": 0, "Ev Idx": 682 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111474.938, "dur": 0.201 , "args": { "External id": 684,"Record function id": 0, "Ev Idx": 683 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111476.381, "dur": 1.272 , "args": { "External id": 685,"Record function id": 0, "Ev Idx": 684 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111476.581, "dur": 0.972 , "args": { "External id": 686,"Record function id": 0, "Ev Idx": 685 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381111487.462, "dur": 4.548 , "args": { "External id": 687,"Record function id": 0, "Ev Idx": 686 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381111497.630, "dur": 5.491 , "args": { "External id": 688,"Record function id": 0, "Ev Idx": 687 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381111508.421, "dur": 5.380 , "args": { "External id": 689,"Record function id": 0, "Ev Idx": 688 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381111516.586, "dur": 19.686 , "args": { "External id": 690,"Record function id": 0, "Ev Idx": 689 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111517.067, "dur": 0.611 , "args": { "External id": 691,"Record function id": 0, "Ev Idx": 690 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111517.347, "dur": 0.221 , "args": { "External id": 692,"Record function id": 0, "Ev Idx": 691 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111517.918, "dur": 0.591 , "args": { "External id": 693,"Record function id": 0, "Ev Idx": 692 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111518.299, "dur": 0.110 , "args": { "External id": 694,"Record function id": 0, "Ev Idx": 693 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111518.760, "dur": 0.380 , "args": { "External id": 695,"Record function id": 0, "Ev Idx": 694 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111518.950, "dur": 0.090 , "args": { "External id": 696,"Record function id": 0, "Ev Idx": 695 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381111519.511, "dur": 15.198 , "args": { "External id": 697,"Record function id": 0, "Ev Idx": 696 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381111520.493, "dur": 2.214 , "args": { "External id": 698,"Record function id": 0, "Ev Idx": 697 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111520.944, "dur": 1.633 , "args": { "External id": 699,"Record function id": 0, "Ev Idx": 698 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111523.208, "dur": 1.323 , "args": { "External id": 700,"Record function id": 0, "Ev Idx": 699 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111525.001, "dur": 1.303 , "args": { "External id": 701,"Record function id": 0, "Ev Idx": 700 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111526.594, "dur": 2.295 , "args": { "External id": 702,"Record function id": 0, "Ev Idx": 701 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111533.277, "dur": 1.052 , "args": { "External id": 703,"Record function id": 0, "Ev Idx": 702 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111535.010, "dur": 0.461 , "args": { "External id": 704,"Record function id": 0, "Ev Idx": 703 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111535.220, "dur": 0.151 , "args": { "External id": 705,"Record function id": 0, "Ev Idx": 704 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111547.634, "dur": 7.604 , "args": { "External id": 706,"Record function id": 0, "Ev Idx": 705 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111557.762, "dur": 11.161 , "args": { "External id": 707,"Record function id": 0, "Ev Idx": 706 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111558.414, "dur": 0.731 , "args": { "External id": 708,"Record function id": 0, "Ev Idx": 707 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111558.584, "dur": 0.471 , "args": { "External id": 709,"Record function id": 0, "Ev Idx": 708 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111558.814, "dur": 0.121 , "args": { "External id": 710,"Record function id": 0, "Ev Idx": 709 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111559.466, "dur": 0.661 , "args": { "External id": 711,"Record function id": 0, "Ev Idx": 710 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111559.636, "dur": 0.401 , "args": { "External id": 712,"Record function id": 0, "Ev Idx": 711 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111559.846, "dur": 0.080 , "args": { "External id": 713,"Record function id": 0, "Ev Idx": 712 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111560.417, "dur": 0.652 , "args": { "External id": 714,"Record function id": 0, "Ev Idx": 713 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111560.578, "dur": 0.390 , "args": { "External id": 715,"Record function id": 0, "Ev Idx": 714 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111560.798, "dur": 0.070 , "args": { "External id": 716,"Record function id": 0, "Ev Idx": 715 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111562.281, "dur": 1.262 , "args": { "External id": 717,"Record function id": 0, "Ev Idx": 716 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111562.481, "dur": 0.962 , "args": { "External id": 718,"Record function id": 0, "Ev Idx": 717 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381111573.662, "dur": 5.821 , "args": { "External id": 719,"Record function id": 0, "Ev Idx": 718 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111582.649, "dur": 5.510 , "args": { "External id": 720,"Record function id": 0, "Ev Idx": 719 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111592.227, "dur": 5.029 , "args": { "External id": 721,"Record function id": 0, "Ev Idx": 720 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111601.193, "dur": 10.299 , "args": { "External id": 722,"Record function id": 0, "Ev Idx": 721 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111601.704, "dur": 0.661 , "args": { "External id": 723,"Record function id": 0, "Ev Idx": 722 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111601.864, "dur": 0.411 , "args": { "External id": 724,"Record function id": 0, "Ev Idx": 723 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111602.085, "dur": 0.090 , "args": { "External id": 725,"Record function id": 0, "Ev Idx": 724 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111602.646, "dur": 0.661 , "args": { "External id": 726,"Record function id": 0, "Ev Idx": 725 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111602.796, "dur": 0.421 , "args": { "External id": 727,"Record function id": 0, "Ev Idx": 726 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111603.037, "dur": 0.080 , "args": { "External id": 728,"Record function id": 0, "Ev Idx": 727 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111603.558, "dur": 0.641 , "args": { "External id": 729,"Record function id": 0, "Ev Idx": 728 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111603.718, "dur": 0.391 , "args": { "External id": 730,"Record function id": 0, "Ev Idx": 729 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111603.928, "dur": 0.080 , "args": { "External id": 731,"Record function id": 0, "Ev Idx": 730 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111605.231, "dur": 1.242 , "args": { "External id": 732,"Record function id": 0, "Ev Idx": 731 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111605.411, "dur": 0.952 , "args": { "External id": 733,"Record function id": 0, "Ev Idx": 732 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111615.600, "dur": 10.169 , "args": { "External id": 734,"Record function id": 0, "Ev Idx": 733 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111616.111, "dur": 0.631 , "args": { "External id": 735,"Record function id": 0, "Ev Idx": 734 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111616.271, "dur": 0.381 , "args": { "External id": 736,"Record function id": 0, "Ev Idx": 735 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111616.472, "dur": 0.070 , "args": { "External id": 737,"Record function id": 0, "Ev Idx": 736 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111617.013, "dur": 0.641 , "args": { "External id": 738,"Record function id": 0, "Ev Idx": 737 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111617.173, "dur": 0.391 , "args": { "External id": 739,"Record function id": 0, "Ev Idx": 738 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111617.383, "dur": 0.081 , "args": { "External id": 740,"Record function id": 0, "Ev Idx": 739 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111617.914, "dur": 0.702 , "args": { "External id": 741,"Record function id": 0, "Ev Idx": 740 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111618.075, "dur": 0.450 , "args": { "External id": 742,"Record function id": 0, "Ev Idx": 741 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111618.275, "dur": 0.140 , "args": { "External id": 743,"Record function id": 0, "Ev Idx": 742 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111619.608, "dur": 1.232 , "args": { "External id": 744,"Record function id": 0, "Ev Idx": 743 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111619.788, "dur": 0.952 , "args": { "External id": 745,"Record function id": 0, "Ev Idx": 744 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381111630.828, "dur": 4.459 , "args": { "External id": 746,"Record function id": 0, "Ev Idx": 745 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381111638.793, "dur": 5.721 , "args": { "External id": 747,"Record function id": 0, "Ev Idx": 746 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111647.900, "dur": 10.209 , "args": { "External id": 748,"Record function id": 0, "Ev Idx": 747 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111648.421, "dur": 0.661 , "args": { "External id": 749,"Record function id": 0, "Ev Idx": 748 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111648.581, "dur": 0.411 , "args": { "External id": 750,"Record function id": 0, "Ev Idx": 749 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111648.802, "dur": 0.080 , "args": { "External id": 751,"Record function id": 0, "Ev Idx": 750 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111649.373, "dur": 0.661 , "args": { "External id": 752,"Record function id": 0, "Ev Idx": 751 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111649.533, "dur": 0.411 , "args": { "External id": 753,"Record function id": 0, "Ev Idx": 752 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111649.754, "dur": 0.090 , "args": { "External id": 754,"Record function id": 0, "Ev Idx": 753 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111650.285, "dur": 0.621 , "args": { "External id": 755,"Record function id": 0, "Ev Idx": 754 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111650.455, "dur": 0.361 , "args": { "External id": 756,"Record function id": 0, "Ev Idx": 755 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111650.645, "dur": 0.080 , "args": { "External id": 757,"Record function id": 0, "Ev Idx": 756 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111651.928, "dur": 1.232 , "args": { "External id": 758,"Record function id": 0, "Ev Idx": 757 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111652.108, "dur": 0.952 , "args": { "External id": 759,"Record function id": 0, "Ev Idx": 758 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381111663.489, "dur": 5.260 , "args": { "External id": 760,"Record function id": 0, "Ev Idx": 759 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381111691.772, "dur": 26.599 , "args": { "External id": 761,"Record function id": 0, "Ev Idx": 760 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381111718.752, "dur": 353.047 , "args": { "External id": 762,"Record function id": 0, "Ev Idx": 761 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381111724.322, "dur": 11.612 , "args": { "External id": 763,"Record function id": 0, "Ev Idx": 762 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111728.170, "dur": 0.440 , "args": { "External id": 764,"Record function id": 0, "Ev Idx": 763 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381111729.141, "dur": 0.050 , "args": { "External id": 765,"Record function id": 0, "Ev Idx": 764 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381111729.612, "dur": 0.040 , "args": { "External id": 766,"Record function id": 0, "Ev Idx": 765 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111731.325, "dur": 0.241 , "args": { "External id": 767,"Record function id": 0, "Ev Idx": 766 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381111731.907, "dur": 0.040 , "args": { "External id": 768,"Record function id": 0, "Ev Idx": 767 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381111732.277, "dur": 0.020 , "args": { "External id": 769,"Record function id": 0, "Ev Idx": 768 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381111739.020, "dur": 1.262 , "args": { "External id": 770,"Record function id": 0, "Ev Idx": 769 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381111742.657, "dur": 325.686 , "args": { "External id": 771,"Record function id": 0, "Ev Idx": 770 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111752.515, "dur": 6.622 , "args": { "External id": 772,"Record function id": 0, "Ev Idx": 771 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111766.110, "dur": 7.594 , "args": { "External id": 773,"Record function id": 0, "Ev Idx": 772 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111779.946, "dur": 5.520 , "args": { "External id": 774,"Record function id": 0, "Ev Idx": 773 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111791.507, "dur": 5.110 , "args": { "External id": 775,"Record function id": 0, "Ev Idx": 774 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111799.202, "dur": 12.383 , "args": { "External id": 776,"Record function id": 0, "Ev Idx": 775 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111799.883, "dur": 0.972 , "args": { "External id": 777,"Record function id": 0, "Ev Idx": 776 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111800.103, "dur": 0.642 , "args": { "External id": 778,"Record function id": 0, "Ev Idx": 777 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111800.404, "dur": 0.210 , "args": { "External id": 779,"Record function id": 0, "Ev Idx": 778 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111801.155, "dur": 0.712 , "args": { "External id": 780,"Record function id": 0, "Ev Idx": 779 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111801.326, "dur": 0.451 , "args": { "External id": 781,"Record function id": 0, "Ev Idx": 780 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111801.586, "dur": 0.090 , "args": { "External id": 782,"Record function id": 0, "Ev Idx": 781 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111802.127, "dur": 0.631 , "args": { "External id": 783,"Record function id": 0, "Ev Idx": 782 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111802.288, "dur": 0.380 , "args": { "External id": 784,"Record function id": 0, "Ev Idx": 783 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111802.478, "dur": 0.090 , "args": { "External id": 785,"Record function id": 0, "Ev Idx": 784 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111804.091, "dur": 1.433 , "args": { "External id": 786,"Record function id": 0, "Ev Idx": 785 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111804.311, "dur": 1.102 , "args": { "External id": 787,"Record function id": 0, "Ev Idx": 786 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381111817.376, "dur": 4.839 , "args": { "External id": 788,"Record function id": 0, "Ev Idx": 787 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111824.619, "dur": 1.012 , "args": { "External id": 789,"Record function id": 0, "Ev Idx": 788 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111827.454, "dur": 0.581 , "args": { "External id": 790,"Record function id": 0, "Ev Idx": 789 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381111831.993, "dur": 0.431 , "args": { "External id": 791,"Record function id": 0, "Ev Idx": 790 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381111832.253, "dur": 0.060 , "args": { "External id": 792,"Record function id": 0, "Ev Idx": 791 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381111835.499, "dur": 6.563 , "args": { "External id": 793,"Record function id": 0, "Ev Idx": 792 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111844.186, "dur": 10.709 , "args": { "External id": 794,"Record function id": 0, "Ev Idx": 793 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111844.717, "dur": 0.731 , "args": { "External id": 795,"Record function id": 0, "Ev Idx": 794 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111844.907, "dur": 0.451 , "args": { "External id": 796,"Record function id": 0, "Ev Idx": 795 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111845.157, "dur": 0.101 , "args": { "External id": 797,"Record function id": 0, "Ev Idx": 796 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111845.738, "dur": 0.632 , "args": { "External id": 798,"Record function id": 0, "Ev Idx": 797 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111845.909, "dur": 0.370 , "args": { "External id": 799,"Record function id": 0, "Ev Idx": 798 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111846.099, "dur": 0.080 , "args": { "External id": 800,"Record function id": 0, "Ev Idx": 799 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111846.620, "dur": 0.711 , "args": { "External id": 801,"Record function id": 0, "Ev Idx": 800 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111846.780, "dur": 0.461 , "args": { "External id": 802,"Record function id": 0, "Ev Idx": 801 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111846.981, "dur": 0.160 , "args": { "External id": 803,"Record function id": 0, "Ev Idx": 802 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111848.403, "dur": 1.333 , "args": { "External id": 804,"Record function id": 0, "Ev Idx": 803 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111848.614, "dur": 1.012 , "args": { "External id": 805,"Record function id": 0, "Ev Idx": 804 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381111858.272, "dur": 3.817 , "args": { "External id": 806,"Record function id": 0, "Ev Idx": 805 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111863.141, "dur": 0.601 , "args": { "External id": 807,"Record function id": 0, "Ev Idx": 806 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381111865.195, "dur": 0.571 , "args": { "External id": 808,"Record function id": 0, "Ev Idx": 807 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381111869.823, "dur": 0.341 , "args": { "External id": 809,"Record function id": 0, "Ev Idx": 808 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381111870.014, "dur": 0.040 , "args": { "External id": 810,"Record function id": 0, "Ev Idx": 809 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381111873.169, "dur": 6.583 , "args": { "External id": 811,"Record function id": 0, "Ev Idx": 810 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111882.196, "dur": 10.440 , "args": { "External id": 812,"Record function id": 0, "Ev Idx": 811 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111882.757, "dur": 0.692 , "args": { "External id": 813,"Record function id": 0, "Ev Idx": 812 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111882.948, "dur": 0.410 , "args": { "External id": 814,"Record function id": 0, "Ev Idx": 813 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111883.168, "dur": 0.080 , "args": { "External id": 815,"Record function id": 0, "Ev Idx": 814 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111883.719, "dur": 0.601 , "args": { "External id": 816,"Record function id": 0, "Ev Idx": 815 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111883.879, "dur": 0.351 , "args": { "External id": 817,"Record function id": 0, "Ev Idx": 816 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111884.060, "dur": 0.060 , "args": { "External id": 818,"Record function id": 0, "Ev Idx": 817 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111884.581, "dur": 0.701 , "args": { "External id": 819,"Record function id": 0, "Ev Idx": 818 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111884.751, "dur": 0.441 , "args": { "External id": 820,"Record function id": 0, "Ev Idx": 819 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111884.931, "dur": 0.161 , "args": { "External id": 821,"Record function id": 0, "Ev Idx": 820 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111886.344, "dur": 1.272 , "args": { "External id": 822,"Record function id": 0, "Ev Idx": 821 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111886.534, "dur": 0.982 , "args": { "External id": 823,"Record function id": 0, "Ev Idx": 822 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381111897.545, "dur": 4.248 , "args": { "External id": 824,"Record function id": 0, "Ev Idx": 823 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381111907.453, "dur": 4.869 , "args": { "External id": 825,"Record function id": 0, "Ev Idx": 824 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381111917.572, "dur": 4.639 , "args": { "External id": 826,"Record function id": 0, "Ev Idx": 825 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381111924.756, "dur": 18.133 , "args": { "External id": 827,"Record function id": 0, "Ev Idx": 826 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111925.226, "dur": 0.541 , "args": { "External id": 828,"Record function id": 0, "Ev Idx": 827 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111925.487, "dur": 0.170 , "args": { "External id": 829,"Record function id": 0, "Ev Idx": 828 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111926.008, "dur": 0.441 , "args": { "External id": 830,"Record function id": 0, "Ev Idx": 829 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111926.248, "dur": 0.090 , "args": { "External id": 831,"Record function id": 0, "Ev Idx": 830 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111926.699, "dur": 0.521 , "args": { "External id": 832,"Record function id": 0, "Ev Idx": 831 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111927.030, "dur": 0.080 , "args": { "External id": 833,"Record function id": 0, "Ev Idx": 832 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381111927.591, "dur": 13.765 , "args": { "External id": 834,"Record function id": 0, "Ev Idx": 833 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381111928.553, "dur": 2.023 , "args": { "External id": 835,"Record function id": 0, "Ev Idx": 834 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111928.833, "dur": 1.623 , "args": { "External id": 836,"Record function id": 0, "Ev Idx": 835 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111931.057, "dur": 1.263 , "args": { "External id": 837,"Record function id": 0, "Ev Idx": 836 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111932.770, "dur": 1.303 , "args": { "External id": 838,"Record function id": 0, "Ev Idx": 837 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111934.343, "dur": 1.493 , "args": { "External id": 839,"Record function id": 0, "Ev Idx": 838 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111940.044, "dur": 0.942 , "args": { "External id": 840,"Record function id": 0, "Ev Idx": 839 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111941.657, "dur": 0.481 , "args": { "External id": 841,"Record function id": 0, "Ev Idx": 840 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111941.877, "dur": 0.161 , "args": { "External id": 842,"Record function id": 0, "Ev Idx": 841 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111950.493, "dur": 6.703 , "args": { "External id": 843,"Record function id": 0, "Ev Idx": 842 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381111959.681, "dur": 10.740 , "args": { "External id": 844,"Record function id": 0, "Ev Idx": 843 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111960.302, "dur": 0.691 , "args": { "External id": 845,"Record function id": 0, "Ev Idx": 844 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111960.482, "dur": 0.421 , "args": { "External id": 846,"Record function id": 0, "Ev Idx": 845 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111960.702, "dur": 0.091 , "args": { "External id": 847,"Record function id": 0, "Ev Idx": 846 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111961.273, "dur": 0.652 , "args": { "External id": 848,"Record function id": 0, "Ev Idx": 847 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111961.454, "dur": 0.381 , "args": { "External id": 849,"Record function id": 0, "Ev Idx": 848 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111961.664, "dur": 0.070 , "args": { "External id": 850,"Record function id": 0, "Ev Idx": 849 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381111962.185, "dur": 0.631 , "args": { "External id": 851,"Record function id": 0, "Ev Idx": 850 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381111962.345, "dur": 0.381 , "args": { "External id": 852,"Record function id": 0, "Ev Idx": 851 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381111962.546, "dur": 0.080 , "args": { "External id": 853,"Record function id": 0, "Ev Idx": 852 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111963.948, "dur": 1.263 , "args": { "External id": 854,"Record function id": 0, "Ev Idx": 853 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381111964.149, "dur": 0.952 , "args": { "External id": 855,"Record function id": 0, "Ev Idx": 854 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381111974.708, "dur": 5.511 , "args": { "External id": 856,"Record function id": 0, "Ev Idx": 855 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111983.104, "dur": 5.300 , "args": { "External id": 857,"Record function id": 0, "Ev Idx": 856 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381111991.890, "dur": 4.880 , "args": { "External id": 858,"Record function id": 0, "Ev Idx": 857 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112000.466, "dur": 10.169 , "args": { "External id": 859,"Record function id": 0, "Ev Idx": 858 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112000.987, "dur": 0.672 , "args": { "External id": 860,"Record function id": 0, "Ev Idx": 859 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112001.158, "dur": 0.410 , "args": { "External id": 861,"Record function id": 0, "Ev Idx": 860 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112001.378, "dur": 0.080 , "args": { "External id": 862,"Record function id": 0, "Ev Idx": 861 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112001.939, "dur": 0.651 , "args": { "External id": 863,"Record function id": 0, "Ev Idx": 862 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112002.099, "dur": 0.401 , "args": { "External id": 864,"Record function id": 0, "Ev Idx": 863 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112002.320, "dur": 0.080 , "args": { "External id": 865,"Record function id": 0, "Ev Idx": 864 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112002.841, "dur": 0.621 , "args": { "External id": 866,"Record function id": 0, "Ev Idx": 865 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112002.991, "dur": 0.381 , "args": { "External id": 867,"Record function id": 0, "Ev Idx": 866 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112003.192, "dur": 0.080 , "args": { "External id": 868,"Record function id": 0, "Ev Idx": 867 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112004.494, "dur": 1.262 , "args": { "External id": 869,"Record function id": 0, "Ev Idx": 868 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112004.674, "dur": 0.972 , "args": { "External id": 870,"Record function id": 0, "Ev Idx": 869 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112014.503, "dur": 9.908 , "args": { "External id": 871,"Record function id": 0, "Ev Idx": 870 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112015.004, "dur": 0.641 , "args": { "External id": 872,"Record function id": 0, "Ev Idx": 871 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112015.174, "dur": 0.381 , "args": { "External id": 873,"Record function id": 0, "Ev Idx": 872 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112015.384, "dur": 0.070 , "args": { "External id": 874,"Record function id": 0, "Ev Idx": 873 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112015.905, "dur": 0.641 , "args": { "External id": 875,"Record function id": 0, "Ev Idx": 874 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112016.065, "dur": 0.391 , "args": { "External id": 876,"Record function id": 0, "Ev Idx": 875 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112016.276, "dur": 0.070 , "args": { "External id": 877,"Record function id": 0, "Ev Idx": 876 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112016.787, "dur": 0.631 , "args": { "External id": 878,"Record function id": 0, "Ev Idx": 877 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112016.947, "dur": 0.381 , "args": { "External id": 879,"Record function id": 0, "Ev Idx": 878 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112017.158, "dur": 0.070 , "args": { "External id": 880,"Record function id": 0, "Ev Idx": 879 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112018.430, "dur": 1.212 , "args": { "External id": 881,"Record function id": 0, "Ev Idx": 880 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112018.600, "dur": 0.942 , "args": { "External id": 882,"Record function id": 0, "Ev Idx": 881 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381112029.230, "dur": 4.358 , "args": { "External id": 883,"Record function id": 0, "Ev Idx": 882 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112037.175, "dur": 5.550 , "args": { "External id": 884,"Record function id": 0, "Ev Idx": 883 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112046.182, "dur": 10.279 , "args": { "External id": 885,"Record function id": 0, "Ev Idx": 884 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112046.682, "dur": 0.662 , "args": { "External id": 886,"Record function id": 0, "Ev Idx": 885 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112046.843, "dur": 0.411 , "args": { "External id": 887,"Record function id": 0, "Ev Idx": 886 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112047.063, "dur": 0.080 , "args": { "External id": 888,"Record function id": 0, "Ev Idx": 887 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112047.604, "dur": 0.661 , "args": { "External id": 889,"Record function id": 0, "Ev Idx": 888 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112047.764, "dur": 0.411 , "args": { "External id": 890,"Record function id": 0, "Ev Idx": 889 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112048.005, "dur": 0.070 , "args": { "External id": 891,"Record function id": 0, "Ev Idx": 890 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112048.526, "dur": 0.631 , "args": { "External id": 892,"Record function id": 0, "Ev Idx": 891 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112048.686, "dur": 0.381 , "args": { "External id": 893,"Record function id": 0, "Ev Idx": 892 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112048.877, "dur": 0.080 , "args": { "External id": 894,"Record function id": 0, "Ev Idx": 893 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112050.179, "dur": 1.242 , "args": { "External id": 895,"Record function id": 0, "Ev Idx": 894 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112050.349, "dur": 0.962 , "args": { "External id": 896,"Record function id": 0, "Ev Idx": 895 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381112061.370, "dur": 5.230 , "args": { "External id": 897,"Record function id": 0, "Ev Idx": 896 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381112088.200, "dur": 25.918 , "args": { "External id": 898,"Record function id": 0, "Ev Idx": 897 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381112114.489, "dur": 351.093 , "args": { "External id": 899,"Record function id": 0, "Ev Idx": 898 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381112120.049, "dur": 11.141 , "args": { "External id": 900,"Record function id": 0, "Ev Idx": 899 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112123.886, "dur": 0.381 , "args": { "External id": 901,"Record function id": 0, "Ev Idx": 900 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381112124.778, "dur": 0.040 , "args": { "External id": 902,"Record function id": 0, "Ev Idx": 901 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381112125.209, "dur": 0.040 , "args": { "External id": 903,"Record function id": 0, "Ev Idx": 902 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112126.511, "dur": 0.230 , "args": { "External id": 904,"Record function id": 0, "Ev Idx": 903 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381112127.082, "dur": 0.030 , "args": { "External id": 905,"Record function id": 0, "Ev Idx": 904 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381112127.453, "dur": 0.020 , "args": { "External id": 906,"Record function id": 0, "Ev Idx": 905 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381112134.326, "dur": 1.212 , "args": { "External id": 907,"Record function id": 0, "Ev Idx": 906 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381112137.862, "dur": 324.414 , "args": { "External id": 908,"Record function id": 0, "Ev Idx": 907 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112147.320, "dur": 7.073 , "args": { "External id": 909,"Record function id": 0, "Ev Idx": 908 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112161.206, "dur": 7.544 , "args": { "External id": 910,"Record function id": 0, "Ev Idx": 909 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112174.741, "dur": 5.320 , "args": { "External id": 911,"Record function id": 0, "Ev Idx": 910 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112185.611, "dur": 4.799 , "args": { "External id": 912,"Record function id": 0, "Ev Idx": 911 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112192.935, "dur": 12.523 , "args": { "External id": 913,"Record function id": 0, "Ev Idx": 912 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112193.546, "dur": 0.982 , "args": { "External id": 914,"Record function id": 0, "Ev Idx": 913 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112193.766, "dur": 0.661 , "args": { "External id": 915,"Record function id": 0, "Ev Idx": 914 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112194.087, "dur": 0.190 , "args": { "External id": 916,"Record function id": 0, "Ev Idx": 915 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112194.818, "dur": 0.641 , "args": { "External id": 917,"Record function id": 0, "Ev Idx": 916 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112194.979, "dur": 0.390 , "args": { "External id": 918,"Record function id": 0, "Ev Idx": 917 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112195.179, "dur": 0.080 , "args": { "External id": 919,"Record function id": 0, "Ev Idx": 918 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112195.710, "dur": 0.731 , "args": { "External id": 920,"Record function id": 0, "Ev Idx": 919 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112195.880, "dur": 0.471 , "args": { "External id": 921,"Record function id": 0, "Ev Idx": 920 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112196.081, "dur": 0.160 , "args": { "External id": 922,"Record function id": 0, "Ev Idx": 921 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112197.784, "dur": 1.462 , "args": { "External id": 923,"Record function id": 0, "Ev Idx": 922 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112198.014, "dur": 1.132 , "args": { "External id": 924,"Record function id": 0, "Ev Idx": 923 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112211.750, "dur": 4.729 , "args": { "External id": 925,"Record function id": 0, "Ev Idx": 924 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112218.673, "dur": 0.941 , "args": { "External id": 926,"Record function id": 0, "Ev Idx": 925 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112221.478, "dur": 0.551 , "args": { "External id": 927,"Record function id": 0, "Ev Idx": 926 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381112225.946, "dur": 0.461 , "args": { "External id": 928,"Record function id": 0, "Ev Idx": 927 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381112226.217, "dur": 0.090 , "args": { "External id": 929,"Record function id": 0, "Ev Idx": 928 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381112230.715, "dur": 6.502 , "args": { "External id": 930,"Record function id": 0, "Ev Idx": 929 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112239.411, "dur": 11.101 , "args": { "External id": 931,"Record function id": 0, "Ev Idx": 930 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112239.932, "dur": 0.912 , "args": { "External id": 932,"Record function id": 0, "Ev Idx": 931 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112240.123, "dur": 0.631 , "args": { "External id": 933,"Record function id": 0, "Ev Idx": 932 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112240.473, "dur": 0.171 , "args": { "External id": 934,"Record function id": 0, "Ev Idx": 933 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112241.134, "dur": 0.632 , "args": { "External id": 935,"Record function id": 0, "Ev Idx": 934 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112241.305, "dur": 0.370 , "args": { "External id": 936,"Record function id": 0, "Ev Idx": 935 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112241.505, "dur": 0.070 , "args": { "External id": 937,"Record function id": 0, "Ev Idx": 936 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112242.016, "dur": 0.862 , "args": { "External id": 938,"Record function id": 0, "Ev Idx": 937 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112242.186, "dur": 0.602 , "args": { "External id": 939,"Record function id": 0, "Ev Idx": 938 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112242.527, "dur": 0.160 , "args": { "External id": 940,"Record function id": 0, "Ev Idx": 939 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112243.980, "dur": 1.342 , "args": { "External id": 941,"Record function id": 0, "Ev Idx": 940 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112244.160, "dur": 1.062 , "args": { "External id": 942,"Record function id": 0, "Ev Idx": 941 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112253.978, "dur": 3.978 , "args": { "External id": 943,"Record function id": 0, "Ev Idx": 942 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112259.048, "dur": 0.621 , "args": { "External id": 944,"Record function id": 0, "Ev Idx": 943 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112261.082, "dur": 0.571 , "args": { "External id": 945,"Record function id": 0, "Ev Idx": 944 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381112265.670, "dur": 0.331 , "args": { "External id": 946,"Record function id": 0, "Ev Idx": 945 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381112265.860, "dur": 0.041 , "args": { "External id": 947,"Record function id": 0, "Ev Idx": 946 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381112268.616, "dur": 5.530 , "args": { "External id": 948,"Record function id": 0, "Ev Idx": 947 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112276.350, "dur": 10.319 , "args": { "External id": 949,"Record function id": 0, "Ev Idx": 948 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112276.891, "dur": 0.731 , "args": { "External id": 950,"Record function id": 0, "Ev Idx": 949 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112277.081, "dur": 0.451 , "args": { "External id": 951,"Record function id": 0, "Ev Idx": 950 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112277.322, "dur": 0.100 , "args": { "External id": 952,"Record function id": 0, "Ev Idx": 951 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112277.893, "dur": 0.641 , "args": { "External id": 953,"Record function id": 0, "Ev Idx": 952 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112278.063, "dur": 0.381 , "args": { "External id": 954,"Record function id": 0, "Ev Idx": 953 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112278.274, "dur": 0.070 , "args": { "External id": 955,"Record function id": 0, "Ev Idx": 954 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112278.795, "dur": 0.631 , "args": { "External id": 956,"Record function id": 0, "Ev Idx": 955 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112278.965, "dur": 0.371 , "args": { "External id": 957,"Record function id": 0, "Ev Idx": 956 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112279.145, "dur": 0.090 , "args": { "External id": 958,"Record function id": 0, "Ev Idx": 957 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112280.478, "dur": 1.292 , "args": { "External id": 959,"Record function id": 0, "Ev Idx": 958 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112280.658, "dur": 1.012 , "args": { "External id": 960,"Record function id": 0, "Ev Idx": 959 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381112291.368, "dur": 4.368 , "args": { "External id": 961,"Record function id": 0, "Ev Idx": 960 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381112301.427, "dur": 4.779 , "args": { "External id": 962,"Record function id": 0, "Ev Idx": 961 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381112311.325, "dur": 4.559 , "args": { "External id": 963,"Record function id": 0, "Ev Idx": 962 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381112318.298, "dur": 17.573 , "args": { "External id": 964,"Record function id": 0, "Ev Idx": 963 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112318.779, "dur": 0.531 , "args": { "External id": 965,"Record function id": 0, "Ev Idx": 964 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112319.039, "dur": 0.161 , "args": { "External id": 966,"Record function id": 0, "Ev Idx": 965 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112319.550, "dur": 0.401 , "args": { "External id": 967,"Record function id": 0, "Ev Idx": 966 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112319.751, "dur": 0.100 , "args": { "External id": 968,"Record function id": 0, "Ev Idx": 967 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112320.192, "dur": 0.380 , "args": { "External id": 969,"Record function id": 0, "Ev Idx": 968 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112320.392, "dur": 0.070 , "args": { "External id": 970,"Record function id": 0, "Ev Idx": 969 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381112320.943, "dur": 13.425 , "args": { "External id": 971,"Record function id": 0, "Ev Idx": 970 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381112321.935, "dur": 1.863 , "args": { "External id": 972,"Record function id": 0, "Ev Idx": 971 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112322.225, "dur": 1.453 , "args": { "External id": 973,"Record function id": 0, "Ev Idx": 972 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112324.279, "dur": 1.273 , "args": { "External id": 974,"Record function id": 0, "Ev Idx": 973 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112326.002, "dur": 1.323 , "args": { "External id": 975,"Record function id": 0, "Ev Idx": 974 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112327.605, "dur": 1.273 , "args": { "External id": 976,"Record function id": 0, "Ev Idx": 975 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112333.066, "dur": 0.951 , "args": { "External id": 977,"Record function id": 0, "Ev Idx": 976 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112334.659, "dur": 0.440 , "args": { "External id": 978,"Record function id": 0, "Ev Idx": 977 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112334.869, "dur": 0.130 , "args": { "External id": 979,"Record function id": 0, "Ev Idx": 978 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112343.064, "dur": 6.823 , "args": { "External id": 980,"Record function id": 0, "Ev Idx": 979 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112352.191, "dur": 10.690 , "args": { "External id": 981,"Record function id": 0, "Ev Idx": 980 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112352.782, "dur": 0.672 , "args": { "External id": 982,"Record function id": 0, "Ev Idx": 981 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112352.953, "dur": 0.410 , "args": { "External id": 983,"Record function id": 0, "Ev Idx": 982 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112353.163, "dur": 0.090 , "args": { "External id": 984,"Record function id": 0, "Ev Idx": 983 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112353.734, "dur": 0.651 , "args": { "External id": 985,"Record function id": 0, "Ev Idx": 984 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112353.904, "dur": 0.391 , "args": { "External id": 986,"Record function id": 0, "Ev Idx": 985 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112354.105, "dur": 0.090 , "args": { "External id": 987,"Record function id": 0, "Ev Idx": 986 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112354.646, "dur": 0.641 , "args": { "External id": 988,"Record function id": 0, "Ev Idx": 987 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112354.806, "dur": 0.391 , "args": { "External id": 989,"Record function id": 0, "Ev Idx": 988 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112355.016, "dur": 0.081 , "args": { "External id": 990,"Record function id": 0, "Ev Idx": 989 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112356.459, "dur": 1.272 , "args": { "External id": 991,"Record function id": 0, "Ev Idx": 990 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112356.669, "dur": 0.952 , "args": { "External id": 992,"Record function id": 0, "Ev Idx": 991 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381112366.969, "dur": 5.670 , "args": { "External id": 993,"Record function id": 0, "Ev Idx": 992 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112375.505, "dur": 5.059 , "args": { "External id": 994,"Record function id": 0, "Ev Idx": 993 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112383.870, "dur": 4.999 , "args": { "External id": 995,"Record function id": 0, "Ev Idx": 994 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112392.466, "dur": 10.199 , "args": { "External id": 996,"Record function id": 0, "Ev Idx": 995 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112392.977, "dur": 0.691 , "args": { "External id": 997,"Record function id": 0, "Ev Idx": 996 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112393.157, "dur": 0.421 , "args": { "External id": 998,"Record function id": 0, "Ev Idx": 997 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112393.388, "dur": 0.090 , "args": { "External id": 999,"Record function id": 0, "Ev Idx": 998 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112393.949, "dur": 0.661 , "args": { "External id": 1000,"Record function id": 0, "Ev Idx": 999 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112394.119, "dur": 0.391 , "args": { "External id": 1001,"Record function id": 0, "Ev Idx": 1000 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112394.330, "dur": 0.080 , "args": { "External id": 1002,"Record function id": 0, "Ev Idx": 1001 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112394.861, "dur": 0.631 , "args": { "External id": 1003,"Record function id": 0, "Ev Idx": 1002 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112395.031, "dur": 0.371 , "args": { "External id": 1004,"Record function id": 0, "Ev Idx": 1003 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112395.221, "dur": 0.080 , "args": { "External id": 1005,"Record function id": 0, "Ev Idx": 1004 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112396.534, "dur": 1.252 , "args": { "External id": 1006,"Record function id": 0, "Ev Idx": 1005 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112396.734, "dur": 0.952 , "args": { "External id": 1007,"Record function id": 0, "Ev Idx": 1006 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112406.482, "dur": 9.859 , "args": { "External id": 1008,"Record function id": 0, "Ev Idx": 1007 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112406.983, "dur": 0.631 , "args": { "External id": 1009,"Record function id": 0, "Ev Idx": 1008 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112407.143, "dur": 0.381 , "args": { "External id": 1010,"Record function id": 0, "Ev Idx": 1009 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112407.344, "dur": 0.080 , "args": { "External id": 1011,"Record function id": 0, "Ev Idx": 1010 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112407.875, "dur": 0.611 , "args": { "External id": 1012,"Record function id": 0, "Ev Idx": 1011 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112408.035, "dur": 0.361 , "args": { "External id": 1013,"Record function id": 0, "Ev Idx": 1012 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112408.225, "dur": 0.071 , "args": { "External id": 1014,"Record function id": 0, "Ev Idx": 1013 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112408.736, "dur": 0.622 , "args": { "External id": 1015,"Record function id": 0, "Ev Idx": 1014 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112408.897, "dur": 0.370 , "args": { "External id": 1016,"Record function id": 0, "Ev Idx": 1015 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112409.087, "dur": 0.080 , "args": { "External id": 1017,"Record function id": 0, "Ev Idx": 1016 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112410.400, "dur": 1.212 , "args": { "External id": 1018,"Record function id": 0, "Ev Idx": 1017 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112410.570, "dur": 0.942 , "args": { "External id": 1019,"Record function id": 0, "Ev Idx": 1018 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381112421.059, "dur": 4.218 , "args": { "External id": 1020,"Record function id": 0, "Ev Idx": 1019 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112428.563, "dur": 5.481 , "args": { "External id": 1021,"Record function id": 0, "Ev Idx": 1020 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112437.420, "dur": 13.184 , "args": { "External id": 1022,"Record function id": 0, "Ev Idx": 1021 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112437.921, "dur": 3.406 , "args": { "External id": 1023,"Record function id": 0, "Ev Idx": 1022 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112438.101, "dur": 3.136 , "args": { "External id": 1024,"Record function id": 0, "Ev Idx": 1023 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112441.017, "dur": 0.110 , "args": { "External id": 1025,"Record function id": 0, "Ev Idx": 1024 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112441.608, "dur": 0.691 , "args": { "External id": 1026,"Record function id": 0, "Ev Idx": 1025 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112441.768, "dur": 0.441 , "args": { "External id": 1027,"Record function id": 0, "Ev Idx": 1026 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112442.028, "dur": 0.071 , "args": { "External id": 1028,"Record function id": 0, "Ev Idx": 1027 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112442.549, "dur": 0.652 , "args": { "External id": 1029,"Record function id": 0, "Ev Idx": 1028 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112442.720, "dur": 0.390 , "args": { "External id": 1030,"Record function id": 0, "Ev Idx": 1029 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112442.930, "dur": 0.080 , "args": { "External id": 1031,"Record function id": 0, "Ev Idx": 1030 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112444.253, "dur": 1.252 , "args": { "External id": 1032,"Record function id": 0, "Ev Idx": 1031 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112444.433, "dur": 0.972 , "args": { "External id": 1033,"Record function id": 0, "Ev Idx": 1032 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381112455.594, "dur": 4.919 , "args": { "External id": 1034,"Record function id": 0, "Ev Idx": 1033 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381112481.001, "dur": 24.636 , "args": { "External id": 1035,"Record function id": 0, "Ev Idx": 1034 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381112506.038, "dur": 356.714 , "args": { "External id": 1036,"Record function id": 0, "Ev Idx": 1035 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381112510.947, "dur": 11.241 , "args": { "External id": 1037,"Record function id": 0, "Ev Idx": 1036 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112514.824, "dur": 0.381 , "args": { "External id": 1038,"Record function id": 0, "Ev Idx": 1037 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381112515.665, "dur": 0.031 , "args": { "External id": 1039,"Record function id": 0, "Ev Idx": 1038 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381112516.086, "dur": 0.040 , "args": { "External id": 1040,"Record function id": 0, "Ev Idx": 1039 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112517.519, "dur": 0.250 , "args": { "External id": 1041,"Record function id": 0, "Ev Idx": 1040 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381112518.120, "dur": 0.030 , "args": { "External id": 1042,"Record function id": 0, "Ev Idx": 1041 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381112518.571, "dur": 0.020 , "args": { "External id": 1043,"Record function id": 0, "Ev Idx": 1042 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381112525.083, "dur": 1.182 , "args": { "External id": 1044,"Record function id": 0, "Ev Idx": 1043 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381112535.001, "dur": 324.645 , "args": { "External id": 1045,"Record function id": 0, "Ev Idx": 1044 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112544.649, "dur": 7.124 , "args": { "External id": 1046,"Record function id": 0, "Ev Idx": 1045 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112558.545, "dur": 7.825 , "args": { "External id": 1047,"Record function id": 0, "Ev Idx": 1046 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112572.401, "dur": 5.390 , "args": { "External id": 1048,"Record function id": 0, "Ev Idx": 1047 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112583.452, "dur": 5.099 , "args": { "External id": 1049,"Record function id": 0, "Ev Idx": 1048 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112591.136, "dur": 13.235 , "args": { "External id": 1050,"Record function id": 0, "Ev Idx": 1049 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112591.817, "dur": 1.062 , "args": { "External id": 1051,"Record function id": 0, "Ev Idx": 1050 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112592.028, "dur": 0.751 , "args": { "External id": 1052,"Record function id": 0, "Ev Idx": 1051 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112592.448, "dur": 0.191 , "args": { "External id": 1053,"Record function id": 0, "Ev Idx": 1052 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112593.210, "dur": 0.621 , "args": { "External id": 1054,"Record function id": 0, "Ev Idx": 1053 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112593.370, "dur": 0.371 , "args": { "External id": 1055,"Record function id": 0, "Ev Idx": 1054 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112593.560, "dur": 0.081 , "args": { "External id": 1056,"Record function id": 0, "Ev Idx": 1055 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112594.101, "dur": 0.642 , "args": { "External id": 1057,"Record function id": 0, "Ev Idx": 1056 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112594.262, "dur": 0.390 , "args": { "External id": 1058,"Record function id": 0, "Ev Idx": 1057 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112594.472, "dur": 0.080 , "args": { "External id": 1059,"Record function id": 0, "Ev Idx": 1058 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112596.065, "dur": 1.753 , "args": { "External id": 1060,"Record function id": 0, "Ev Idx": 1059 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112596.296, "dur": 1.412 , "args": { "External id": 1061,"Record function id": 0, "Ev Idx": 1060 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112610.682, "dur": 4.679 , "args": { "External id": 1062,"Record function id": 0, "Ev Idx": 1061 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112617.776, "dur": 0.971 , "args": { "External id": 1063,"Record function id": 0, "Ev Idx": 1062 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112620.601, "dur": 0.531 , "args": { "External id": 1064,"Record function id": 0, "Ev Idx": 1063 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381112625.049, "dur": 0.481 , "args": { "External id": 1065,"Record function id": 0, "Ev Idx": 1064 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381112625.330, "dur": 0.090 , "args": { "External id": 1066,"Record function id": 0, "Ev Idx": 1065 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381112628.516, "dur": 7.053 , "args": { "External id": 1067,"Record function id": 0, "Ev Idx": 1066 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112637.632, "dur": 10.580 , "args": { "External id": 1068,"Record function id": 0, "Ev Idx": 1067 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112638.184, "dur": 0.741 , "args": { "External id": 1069,"Record function id": 0, "Ev Idx": 1068 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112638.354, "dur": 0.481 , "args": { "External id": 1070,"Record function id": 0, "Ev Idx": 1069 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112638.614, "dur": 0.111 , "args": { "External id": 1071,"Record function id": 0, "Ev Idx": 1070 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112639.195, "dur": 0.642 , "args": { "External id": 1072,"Record function id": 0, "Ev Idx": 1071 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112639.356, "dur": 0.390 , "args": { "External id": 1073,"Record function id": 0, "Ev Idx": 1072 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112639.566, "dur": 0.070 , "args": { "External id": 1074,"Record function id": 0, "Ev Idx": 1073 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112640.097, "dur": 0.641 , "args": { "External id": 1075,"Record function id": 0, "Ev Idx": 1074 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112640.267, "dur": 0.381 , "args": { "External id": 1076,"Record function id": 0, "Ev Idx": 1075 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112640.458, "dur": 0.080 , "args": { "External id": 1077,"Record function id": 0, "Ev Idx": 1076 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112641.810, "dur": 1.323 , "args": { "External id": 1078,"Record function id": 0, "Ev Idx": 1077 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112642.001, "dur": 1.032 , "args": { "External id": 1079,"Record function id": 0, "Ev Idx": 1078 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112651.578, "dur": 3.707 , "args": { "External id": 1080,"Record function id": 0, "Ev Idx": 1079 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112656.387, "dur": 0.581 , "args": { "External id": 1081,"Record function id": 0, "Ev Idx": 1080 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381112658.281, "dur": 0.551 , "args": { "External id": 1082,"Record function id": 0, "Ev Idx": 1081 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381112662.809, "dur": 0.321 , "args": { "External id": 1083,"Record function id": 0, "Ev Idx": 1082 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381112662.990, "dur": 0.040 , "args": { "External id": 1084,"Record function id": 0, "Ev Idx": 1083 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381112665.675, "dur": 6.301 , "args": { "External id": 1085,"Record function id": 0, "Ev Idx": 1084 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112674.341, "dur": 10.359 , "args": { "External id": 1086,"Record function id": 0, "Ev Idx": 1085 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112674.902, "dur": 0.751 , "args": { "External id": 1087,"Record function id": 0, "Ev Idx": 1086 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112675.102, "dur": 0.461 , "args": { "External id": 1088,"Record function id": 0, "Ev Idx": 1087 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112675.353, "dur": 0.090 , "args": { "External id": 1089,"Record function id": 0, "Ev Idx": 1088 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112675.924, "dur": 0.621 , "args": { "External id": 1090,"Record function id": 0, "Ev Idx": 1089 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112676.084, "dur": 0.371 , "args": { "External id": 1091,"Record function id": 0, "Ev Idx": 1090 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112676.284, "dur": 0.071 , "args": { "External id": 1092,"Record function id": 0, "Ev Idx": 1091 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112676.795, "dur": 0.642 , "args": { "External id": 1093,"Record function id": 0, "Ev Idx": 1092 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112676.966, "dur": 0.380 , "args": { "External id": 1094,"Record function id": 0, "Ev Idx": 1093 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112677.156, "dur": 0.090 , "args": { "External id": 1095,"Record function id": 0, "Ev Idx": 1094 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112678.519, "dur": 1.252 , "args": { "External id": 1096,"Record function id": 0, "Ev Idx": 1095 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112678.699, "dur": 0.972 , "args": { "External id": 1097,"Record function id": 0, "Ev Idx": 1096 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381112690.070, "dur": 4.188 , "args": { "External id": 1098,"Record function id": 0, "Ev Idx": 1097 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381112699.688, "dur": 5.140 , "args": { "External id": 1099,"Record function id": 0, "Ev Idx": 1098 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381112709.977, "dur": 4.629 , "args": { "External id": 1100,"Record function id": 0, "Ev Idx": 1099 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381112717.171, "dur": 18.574 , "args": { "External id": 1101,"Record function id": 0, "Ev Idx": 1100 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112717.702, "dur": 0.541 , "args": { "External id": 1102,"Record function id": 0, "Ev Idx": 1101 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112717.922, "dur": 0.200 , "args": { "External id": 1103,"Record function id": 0, "Ev Idx": 1102 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112718.483, "dur": 0.411 , "args": { "External id": 1104,"Record function id": 0, "Ev Idx": 1103 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112718.703, "dur": 0.091 , "args": { "External id": 1105,"Record function id": 0, "Ev Idx": 1104 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112719.134, "dur": 0.401 , "args": { "External id": 1106,"Record function id": 0, "Ev Idx": 1105 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112719.345, "dur": 0.090 , "args": { "External id": 1107,"Record function id": 0, "Ev Idx": 1106 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381112719.866, "dur": 14.286 , "args": { "External id": 1108,"Record function id": 0, "Ev Idx": 1107 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381112720.897, "dur": 2.225 , "args": { "External id": 1109,"Record function id": 0, "Ev Idx": 1108 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112721.438, "dur": 1.573 , "args": { "External id": 1110,"Record function id": 0, "Ev Idx": 1109 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112723.552, "dur": 1.253 , "args": { "External id": 1111,"Record function id": 0, "Ev Idx": 1110 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112725.266, "dur": 1.322 , "args": { "External id": 1112,"Record function id": 0, "Ev Idx": 1111 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112726.859, "dur": 1.623 , "args": { "External id": 1113,"Record function id": 0, "Ev Idx": 1112 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112732.770, "dur": 1.011 , "args": { "External id": 1114,"Record function id": 0, "Ev Idx": 1113 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112734.473, "dur": 0.451 , "args": { "External id": 1115,"Record function id": 0, "Ev Idx": 1114 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112734.693, "dur": 0.130 , "args": { "External id": 1116,"Record function id": 0, "Ev Idx": 1115 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112743.179, "dur": 6.572 , "args": { "External id": 1117,"Record function id": 0, "Ev Idx": 1116 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112752.226, "dur": 10.950 , "args": { "External id": 1118,"Record function id": 0, "Ev Idx": 1117 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112752.897, "dur": 0.691 , "args": { "External id": 1119,"Record function id": 0, "Ev Idx": 1118 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112753.077, "dur": 0.411 , "args": { "External id": 1120,"Record function id": 0, "Ev Idx": 1119 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112753.288, "dur": 0.090 , "args": { "External id": 1121,"Record function id": 0, "Ev Idx": 1120 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112753.859, "dur": 0.651 , "args": { "External id": 1122,"Record function id": 0, "Ev Idx": 1121 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112754.019, "dur": 0.401 , "args": { "External id": 1123,"Record function id": 0, "Ev Idx": 1122 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112754.230, "dur": 0.080 , "args": { "External id": 1124,"Record function id": 0, "Ev Idx": 1123 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112754.761, "dur": 0.641 , "args": { "External id": 1125,"Record function id": 0, "Ev Idx": 1124 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112754.921, "dur": 0.391 , "args": { "External id": 1126,"Record function id": 0, "Ev Idx": 1125 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112755.131, "dur": 0.070 , "args": { "External id": 1127,"Record function id": 0, "Ev Idx": 1126 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112756.524, "dur": 1.292 , "args": { "External id": 1128,"Record function id": 0, "Ev Idx": 1127 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112756.724, "dur": 0.982 , "args": { "External id": 1129,"Record function id": 0, "Ev Idx": 1128 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381112767.264, "dur": 5.370 , "args": { "External id": 1130,"Record function id": 0, "Ev Idx": 1129 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112775.549, "dur": 5.210 , "args": { "External id": 1131,"Record function id": 0, "Ev Idx": 1130 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112784.115, "dur": 4.839 , "args": { "External id": 1132,"Record function id": 0, "Ev Idx": 1131 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112792.491, "dur": 10.299 , "args": { "External id": 1133,"Record function id": 0, "Ev Idx": 1132 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112793.002, "dur": 0.681 , "args": { "External id": 1134,"Record function id": 0, "Ev Idx": 1133 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112793.172, "dur": 0.411 , "args": { "External id": 1135,"Record function id": 0, "Ev Idx": 1134 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112793.402, "dur": 0.071 , "args": { "External id": 1136,"Record function id": 0, "Ev Idx": 1135 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112793.953, "dur": 0.652 , "args": { "External id": 1137,"Record function id": 0, "Ev Idx": 1136 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112794.114, "dur": 0.401 , "args": { "External id": 1138,"Record function id": 0, "Ev Idx": 1137 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112794.334, "dur": 0.080 , "args": { "External id": 1139,"Record function id": 0, "Ev Idx": 1138 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112794.865, "dur": 0.631 , "args": { "External id": 1140,"Record function id": 0, "Ev Idx": 1139 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112795.035, "dur": 0.371 , "args": { "External id": 1141,"Record function id": 0, "Ev Idx": 1140 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112795.236, "dur": 0.070 , "args": { "External id": 1142,"Record function id": 0, "Ev Idx": 1141 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112796.548, "dur": 1.273 , "args": { "External id": 1143,"Record function id": 0, "Ev Idx": 1142 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112796.739, "dur": 0.981 , "args": { "External id": 1144,"Record function id": 0, "Ev Idx": 1143 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112806.747, "dur": 9.959 , "args": { "External id": 1145,"Record function id": 0, "Ev Idx": 1144 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112807.238, "dur": 0.661 , "args": { "External id": 1146,"Record function id": 0, "Ev Idx": 1145 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112807.419, "dur": 0.390 , "args": { "External id": 1147,"Record function id": 0, "Ev Idx": 1146 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112807.639, "dur": 0.060 , "args": { "External id": 1148,"Record function id": 0, "Ev Idx": 1147 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112808.160, "dur": 0.661 , "args": { "External id": 1149,"Record function id": 0, "Ev Idx": 1148 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112808.340, "dur": 0.391 , "args": { "External id": 1150,"Record function id": 0, "Ev Idx": 1149 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112808.551, "dur": 0.070 , "args": { "External id": 1151,"Record function id": 0, "Ev Idx": 1150 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112809.072, "dur": 0.621 , "args": { "External id": 1152,"Record function id": 0, "Ev Idx": 1151 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112809.232, "dur": 0.371 , "args": { "External id": 1153,"Record function id": 0, "Ev Idx": 1152 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112809.422, "dur": 0.070 , "args": { "External id": 1154,"Record function id": 0, "Ev Idx": 1153 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112810.695, "dur": 1.232 , "args": { "External id": 1155,"Record function id": 0, "Ev Idx": 1154 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112810.885, "dur": 0.942 , "args": { "External id": 1156,"Record function id": 0, "Ev Idx": 1155 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381112821.695, "dur": 4.408 , "args": { "External id": 1157,"Record function id": 0, "Ev Idx": 1156 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112829.450, "dur": 5.600 , "args": { "External id": 1158,"Record function id": 0, "Ev Idx": 1157 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112838.366, "dur": 10.129 , "args": { "External id": 1159,"Record function id": 0, "Ev Idx": 1158 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112838.867, "dur": 0.651 , "args": { "External id": 1160,"Record function id": 0, "Ev Idx": 1159 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112839.027, "dur": 0.401 , "args": { "External id": 1161,"Record function id": 0, "Ev Idx": 1160 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112839.238, "dur": 0.080 , "args": { "External id": 1162,"Record function id": 0, "Ev Idx": 1161 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112839.789, "dur": 0.631 , "args": { "External id": 1163,"Record function id": 0, "Ev Idx": 1162 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112839.939, "dur": 0.391 , "args": { "External id": 1164,"Record function id": 0, "Ev Idx": 1163 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112840.149, "dur": 0.081 , "args": { "External id": 1165,"Record function id": 0, "Ev Idx": 1164 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112840.670, "dur": 0.662 , "args": { "External id": 1166,"Record function id": 0, "Ev Idx": 1165 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112840.851, "dur": 0.390 , "args": { "External id": 1167,"Record function id": 0, "Ev Idx": 1166 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112841.051, "dur": 0.090 , "args": { "External id": 1168,"Record function id": 0, "Ev Idx": 1167 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112842.344, "dur": 1.262 , "args": { "External id": 1169,"Record function id": 0, "Ev Idx": 1168 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112842.534, "dur": 0.962 , "args": { "External id": 1170,"Record function id": 0, "Ev Idx": 1169 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381112853.184, "dur": 4.819 , "args": { "External id": 1171,"Record function id": 0, "Ev Idx": 1170 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381112878.621, "dur": 24.245 , "args": { "External id": 1172,"Record function id": 0, "Ev Idx": 1171 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381112903.217, "dur": 340.804 , "args": { "External id": 1173,"Record function id": 0, "Ev Idx": 1172 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381112908.286, "dur": 10.871 , "args": { "External id": 1174,"Record function id": 0, "Ev Idx": 1173 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112911.903, "dur": 0.391 , "args": { "External id": 1175,"Record function id": 0, "Ev Idx": 1174 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381112912.845, "dur": 0.050 , "args": { "External id": 1176,"Record function id": 0, "Ev Idx": 1175 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381112913.296, "dur": 0.040 , "args": { "External id": 1177,"Record function id": 0, "Ev Idx": 1176 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112914.618, "dur": 0.251 , "args": { "External id": 1178,"Record function id": 0, "Ev Idx": 1177 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381112915.229, "dur": 0.030 , "args": { "External id": 1179,"Record function id": 0, "Ev Idx": 1178 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381112915.590, "dur": 0.030 , "args": { "External id": 1180,"Record function id": 0, "Ev Idx": 1179 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381112922.022, "dur": 1.062 , "args": { "External id": 1181,"Record function id": 0, "Ev Idx": 1180 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381112925.468, "dur": 315.397 , "args": { "External id": 1182,"Record function id": 0, "Ev Idx": 1181 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112934.275, "dur": 6.321 , "args": { "External id": 1183,"Record function id": 0, "Ev Idx": 1182 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112946.988, "dur": 7.324 , "args": { "External id": 1184,"Record function id": 0, "Ev Idx": 1183 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112960.103, "dur": 5.510 , "args": { "External id": 1185,"Record function id": 0, "Ev Idx": 1184 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381112971.133, "dur": 5.030 , "args": { "External id": 1186,"Record function id": 0, "Ev Idx": 1185 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381112978.617, "dur": 12.413 , "args": { "External id": 1187,"Record function id": 0, "Ev Idx": 1186 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112979.238, "dur": 0.922 , "args": { "External id": 1188,"Record function id": 0, "Ev Idx": 1187 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112979.449, "dur": 0.611 , "args": { "External id": 1189,"Record function id": 0, "Ev Idx": 1188 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112979.759, "dur": 0.151 , "args": { "External id": 1190,"Record function id": 0, "Ev Idx": 1189 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112980.451, "dur": 0.711 , "args": { "External id": 1191,"Record function id": 0, "Ev Idx": 1190 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112980.611, "dur": 0.461 , "args": { "External id": 1192,"Record function id": 0, "Ev Idx": 1191 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112980.901, "dur": 0.071 , "args": { "External id": 1193,"Record function id": 0, "Ev Idx": 1192 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381112981.412, "dur": 0.702 , "args": { "External id": 1194,"Record function id": 0, "Ev Idx": 1193 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381112981.573, "dur": 0.451 , "args": { "External id": 1195,"Record function id": 0, "Ev Idx": 1194 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381112981.763, "dur": 0.160 , "args": { "External id": 1196,"Record function id": 0, "Ev Idx": 1195 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112983.396, "dur": 1.403 , "args": { "External id": 1197,"Record function id": 0, "Ev Idx": 1196 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381112983.596, "dur": 1.082 , "args": { "External id": 1198,"Record function id": 0, "Ev Idx": 1197 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381112996.440, "dur": 4.699 , "args": { "External id": 1199,"Record function id": 0, "Ev Idx": 1198 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113003.433, "dur": 0.962 , "args": { "External id": 1200,"Record function id": 0, "Ev Idx": 1199 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113006.108, "dur": 0.581 , "args": { "External id": 1201,"Record function id": 0, "Ev Idx": 1200 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381113010.507, "dur": 0.511 , "args": { "External id": 1202,"Record function id": 0, "Ev Idx": 1201 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381113010.817, "dur": 0.080 , "args": { "External id": 1203,"Record function id": 0, "Ev Idx": 1202 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381113014.203, "dur": 6.352 , "args": { "External id": 1204,"Record function id": 0, "Ev Idx": 1203 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113023.000, "dur": 10.690 , "args": { "External id": 1205,"Record function id": 0, "Ev Idx": 1204 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113023.521, "dur": 0.781 , "args": { "External id": 1206,"Record function id": 0, "Ev Idx": 1205 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113023.691, "dur": 0.511 , "args": { "External id": 1207,"Record function id": 0, "Ev Idx": 1206 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113023.982, "dur": 0.110 , "args": { "External id": 1208,"Record function id": 0, "Ev Idx": 1207 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113024.583, "dur": 0.641 , "args": { "External id": 1209,"Record function id": 0, "Ev Idx": 1208 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113024.753, "dur": 0.381 , "args": { "External id": 1210,"Record function id": 0, "Ev Idx": 1209 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113024.943, "dur": 0.091 , "args": { "External id": 1211,"Record function id": 0, "Ev Idx": 1210 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113025.474, "dur": 0.782 , "args": { "External id": 1212,"Record function id": 0, "Ev Idx": 1211 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113025.625, "dur": 0.541 , "args": { "External id": 1213,"Record function id": 0, "Ev Idx": 1212 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113025.885, "dur": 0.181 , "args": { "External id": 1214,"Record function id": 0, "Ev Idx": 1213 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113027.378, "dur": 1.292 , "args": { "External id": 1215,"Record function id": 0, "Ev Idx": 1214 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113027.558, "dur": 1.002 , "args": { "External id": 1216,"Record function id": 0, "Ev Idx": 1215 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113036.906, "dur": 3.827 , "args": { "External id": 1217,"Record function id": 0, "Ev Idx": 1216 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113041.835, "dur": 0.621 , "args": { "External id": 1218,"Record function id": 0, "Ev Idx": 1217 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113043.859, "dur": 0.551 , "args": { "External id": 1219,"Record function id": 0, "Ev Idx": 1218 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381113048.277, "dur": 0.341 , "args": { "External id": 1220,"Record function id": 0, "Ev Idx": 1219 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381113048.467, "dur": 0.050 , "args": { "External id": 1221,"Record function id": 0, "Ev Idx": 1220 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381113051.112, "dur": 5.951 , "args": { "External id": 1222,"Record function id": 0, "Ev Idx": 1221 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113059.237, "dur": 10.460 , "args": { "External id": 1223,"Record function id": 0, "Ev Idx": 1222 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113059.778, "dur": 0.752 , "args": { "External id": 1224,"Record function id": 0, "Ev Idx": 1223 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113059.979, "dur": 0.461 , "args": { "External id": 1225,"Record function id": 0, "Ev Idx": 1224 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113060.219, "dur": 0.110 , "args": { "External id": 1226,"Record function id": 0, "Ev Idx": 1225 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113060.800, "dur": 0.651 , "args": { "External id": 1227,"Record function id": 0, "Ev Idx": 1226 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113060.971, "dur": 0.390 , "args": { "External id": 1228,"Record function id": 0, "Ev Idx": 1227 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113061.181, "dur": 0.080 , "args": { "External id": 1229,"Record function id": 0, "Ev Idx": 1228 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113061.712, "dur": 0.691 , "args": { "External id": 1230,"Record function id": 0, "Ev Idx": 1229 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113061.862, "dur": 0.451 , "args": { "External id": 1231,"Record function id": 0, "Ev Idx": 1230 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113062.053, "dur": 0.150 , "args": { "External id": 1232,"Record function id": 0, "Ev Idx": 1231 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113063.465, "dur": 1.303 , "args": { "External id": 1233,"Record function id": 0, "Ev Idx": 1232 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113063.645, "dur": 1.012 , "args": { "External id": 1234,"Record function id": 0, "Ev Idx": 1233 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381113074.406, "dur": 4.217 , "args": { "External id": 1235,"Record function id": 0, "Ev Idx": 1234 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381113083.863, "dur": 4.669 , "args": { "External id": 1236,"Record function id": 0, "Ev Idx": 1235 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381113093.441, "dur": 4.298 , "args": { "External id": 1237,"Record function id": 0, "Ev Idx": 1236 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381113100.153, "dur": 17.573 , "args": { "External id": 1238,"Record function id": 0, "Ev Idx": 1237 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113100.674, "dur": 0.592 , "args": { "External id": 1239,"Record function id": 0, "Ev Idx": 1238 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113100.925, "dur": 0.220 , "args": { "External id": 1240,"Record function id": 0, "Ev Idx": 1239 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113101.516, "dur": 0.471 , "args": { "External id": 1241,"Record function id": 0, "Ev Idx": 1240 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113101.796, "dur": 0.081 , "args": { "External id": 1242,"Record function id": 0, "Ev Idx": 1241 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113102.227, "dur": 0.381 , "args": { "External id": 1243,"Record function id": 0, "Ev Idx": 1242 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113102.418, "dur": 0.090 , "args": { "External id": 1244,"Record function id": 0, "Ev Idx": 1243 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381113102.949, "dur": 13.334 , "args": { "External id": 1245,"Record function id": 0, "Ev Idx": 1244 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381113103.920, "dur": 1.824 , "args": { "External id": 1246,"Record function id": 0, "Ev Idx": 1245 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113104.181, "dur": 1.443 , "args": { "External id": 1247,"Record function id": 0, "Ev Idx": 1246 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113106.185, "dur": 1.282 , "args": { "External id": 1248,"Record function id": 0, "Ev Idx": 1247 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113107.908, "dur": 1.222 , "args": { "External id": 1249,"Record function id": 0, "Ev Idx": 1248 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113109.411, "dur": 1.422 , "args": { "External id": 1250,"Record function id": 0, "Ev Idx": 1249 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113114.901, "dur": 1.052 , "args": { "External id": 1251,"Record function id": 0, "Ev Idx": 1250 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113116.564, "dur": 0.461 , "args": { "External id": 1252,"Record function id": 0, "Ev Idx": 1251 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113116.784, "dur": 0.131 , "args": { "External id": 1253,"Record function id": 0, "Ev Idx": 1252 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113124.789, "dur": 6.733 , "args": { "External id": 1254,"Record function id": 0, "Ev Idx": 1253 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113133.986, "dur": 10.720 , "args": { "External id": 1255,"Record function id": 0, "Ev Idx": 1254 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113134.547, "dur": 0.772 , "args": { "External id": 1256,"Record function id": 0, "Ev Idx": 1255 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113134.798, "dur": 0.421 , "args": { "External id": 1257,"Record function id": 0, "Ev Idx": 1256 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113135.028, "dur": 0.081 , "args": { "External id": 1258,"Record function id": 0, "Ev Idx": 1257 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113135.609, "dur": 0.622 , "args": { "External id": 1259,"Record function id": 0, "Ev Idx": 1258 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113135.770, "dur": 0.370 , "args": { "External id": 1260,"Record function id": 0, "Ev Idx": 1259 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113135.970, "dur": 0.070 , "args": { "External id": 1261,"Record function id": 0, "Ev Idx": 1260 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113136.501, "dur": 0.631 , "args": { "External id": 1262,"Record function id": 0, "Ev Idx": 1261 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113136.661, "dur": 0.381 , "args": { "External id": 1263,"Record function id": 0, "Ev Idx": 1262 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113136.862, "dur": 0.070 , "args": { "External id": 1264,"Record function id": 0, "Ev Idx": 1263 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113138.254, "dur": 1.263 , "args": { "External id": 1265,"Record function id": 0, "Ev Idx": 1264 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113138.455, "dur": 0.962 , "args": { "External id": 1266,"Record function id": 0, "Ev Idx": 1265 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381113149.095, "dur": 5.329 , "args": { "External id": 1267,"Record function id": 0, "Ev Idx": 1266 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113157.460, "dur": 5.300 , "args": { "External id": 1268,"Record function id": 0, "Ev Idx": 1267 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113166.016, "dur": 5.140 , "args": { "External id": 1269,"Record function id": 0, "Ev Idx": 1268 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113174.832, "dur": 10.179 , "args": { "External id": 1270,"Record function id": 0, "Ev Idx": 1269 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113175.333, "dur": 0.672 , "args": { "External id": 1271,"Record function id": 0, "Ev Idx": 1270 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113175.504, "dur": 0.400 , "args": { "External id": 1272,"Record function id": 0, "Ev Idx": 1271 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113175.724, "dur": 0.080 , "args": { "External id": 1273,"Record function id": 0, "Ev Idx": 1272 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113176.255, "dur": 0.661 , "args": { "External id": 1274,"Record function id": 0, "Ev Idx": 1273 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113176.425, "dur": 0.401 , "args": { "External id": 1275,"Record function id": 0, "Ev Idx": 1274 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113176.656, "dur": 0.070 , "args": { "External id": 1276,"Record function id": 0, "Ev Idx": 1275 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113177.177, "dur": 0.651 , "args": { "External id": 1277,"Record function id": 0, "Ev Idx": 1276 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113177.347, "dur": 0.391 , "args": { "External id": 1278,"Record function id": 0, "Ev Idx": 1277 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113177.548, "dur": 0.090 , "args": { "External id": 1279,"Record function id": 0, "Ev Idx": 1278 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113178.870, "dur": 1.262 , "args": { "External id": 1280,"Record function id": 0, "Ev Idx": 1279 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113179.070, "dur": 0.962 , "args": { "External id": 1281,"Record function id": 0, "Ev Idx": 1280 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113188.668, "dur": 10.009 , "args": { "External id": 1282,"Record function id": 0, "Ev Idx": 1281 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113189.209, "dur": 0.641 , "args": { "External id": 1283,"Record function id": 0, "Ev Idx": 1282 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113189.370, "dur": 0.390 , "args": { "External id": 1284,"Record function id": 0, "Ev Idx": 1283 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113189.570, "dur": 0.080 , "args": { "External id": 1285,"Record function id": 0, "Ev Idx": 1284 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113190.131, "dur": 0.641 , "args": { "External id": 1286,"Record function id": 0, "Ev Idx": 1285 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113190.301, "dur": 0.371 , "args": { "External id": 1287,"Record function id": 0, "Ev Idx": 1286 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113190.502, "dur": 0.070 , "args": { "External id": 1288,"Record function id": 0, "Ev Idx": 1287 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113191.013, "dur": 0.631 , "args": { "External id": 1289,"Record function id": 0, "Ev Idx": 1288 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113191.163, "dur": 0.391 , "args": { "External id": 1290,"Record function id": 0, "Ev Idx": 1289 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113191.383, "dur": 0.070 , "args": { "External id": 1291,"Record function id": 0, "Ev Idx": 1290 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113192.666, "dur": 1.232 , "args": { "External id": 1292,"Record function id": 0, "Ev Idx": 1291 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113192.846, "dur": 0.952 , "args": { "External id": 1293,"Record function id": 0, "Ev Idx": 1292 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381113203.356, "dur": 4.167 , "args": { "External id": 1294,"Record function id": 0, "Ev Idx": 1293 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113210.870, "dur": 5.289 , "args": { "External id": 1295,"Record function id": 0, "Ev Idx": 1294 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113219.375, "dur": 10.119 , "args": { "External id": 1296,"Record function id": 0, "Ev Idx": 1295 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113219.886, "dur": 0.672 , "args": { "External id": 1297,"Record function id": 0, "Ev Idx": 1296 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113220.047, "dur": 0.420 , "args": { "External id": 1298,"Record function id": 0, "Ev Idx": 1297 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113220.277, "dur": 0.070 , "args": { "External id": 1299,"Record function id": 0, "Ev Idx": 1298 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113220.848, "dur": 0.681 , "args": { "External id": 1300,"Record function id": 0, "Ev Idx": 1299 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113221.018, "dur": 0.421 , "args": { "External id": 1301,"Record function id": 0, "Ev Idx": 1300 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113221.259, "dur": 0.070 , "args": { "External id": 1302,"Record function id": 0, "Ev Idx": 1301 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113221.780, "dur": 0.611 , "args": { "External id": 1303,"Record function id": 0, "Ev Idx": 1302 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113221.930, "dur": 0.371 , "args": { "External id": 1304,"Record function id": 0, "Ev Idx": 1303 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113222.120, "dur": 0.071 , "args": { "External id": 1305,"Record function id": 0, "Ev Idx": 1304 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113223.423, "dur": 1.232 , "args": { "External id": 1306,"Record function id": 0, "Ev Idx": 1305 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113223.603, "dur": 0.952 , "args": { "External id": 1307,"Record function id": 0, "Ev Idx": 1306 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381113234.333, "dur": 4.919 , "args": { "External id": 1308,"Record function id": 0, "Ev Idx": 1307 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381113259.129, "dur": 25.157 , "args": { "External id": 1309,"Record function id": 0, "Ev Idx": 1308 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381113284.637, "dur": 347.246 , "args": { "External id": 1310,"Record function id": 0, "Ev Idx": 1309 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381113289.566, "dur": 10.490 , "args": { "External id": 1311,"Record function id": 0, "Ev Idx": 1310 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113292.932, "dur": 0.431 , "args": { "External id": 1312,"Record function id": 0, "Ev Idx": 1311 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381113293.844, "dur": 0.040 , "args": { "External id": 1313,"Record function id": 0, "Ev Idx": 1312 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381113294.295, "dur": 0.040 , "args": { "External id": 1314,"Record function id": 0, "Ev Idx": 1313 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113295.567, "dur": 0.201 , "args": { "External id": 1315,"Record function id": 0, "Ev Idx": 1314 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381113296.098, "dur": 0.030 , "args": { "External id": 1316,"Record function id": 0, "Ev Idx": 1315 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381113296.449, "dur": 0.020 , "args": { "External id": 1317,"Record function id": 0, "Ev Idx": 1316 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381113302.861, "dur": 1.032 , "args": { "External id": 1318,"Record function id": 0, "Ev Idx": 1317 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381113306.217, "dur": 322.540 , "args": { "External id": 1319,"Record function id": 0, "Ev Idx": 1318 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113314.833, "dur": 6.522 , "args": { "External id": 1320,"Record function id": 0, "Ev Idx": 1319 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113328.018, "dur": 7.273 , "args": { "External id": 1321,"Record function id": 0, "Ev Idx": 1320 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113341.202, "dur": 5.480 , "args": { "External id": 1322,"Record function id": 0, "Ev Idx": 1321 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113352.213, "dur": 5.019 , "args": { "External id": 1323,"Record function id": 0, "Ev Idx": 1322 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113359.677, "dur": 12.082 , "args": { "External id": 1324,"Record function id": 0, "Ev Idx": 1323 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113360.288, "dur": 1.042 , "args": { "External id": 1325,"Record function id": 0, "Ev Idx": 1324 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113360.498, "dur": 0.721 , "args": { "External id": 1326,"Record function id": 0, "Ev Idx": 1325 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113360.799, "dur": 0.280 , "args": { "External id": 1327,"Record function id": 0, "Ev Idx": 1326 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113361.630, "dur": 0.641 , "args": { "External id": 1328,"Record function id": 0, "Ev Idx": 1327 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113361.790, "dur": 0.391 , "args": { "External id": 1329,"Record function id": 0, "Ev Idx": 1328 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113361.991, "dur": 0.080 , "args": { "External id": 1330,"Record function id": 0, "Ev Idx": 1329 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113362.522, "dur": 0.631 , "args": { "External id": 1331,"Record function id": 0, "Ev Idx": 1330 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113362.682, "dur": 0.381 , "args": { "External id": 1332,"Record function id": 0, "Ev Idx": 1331 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113362.872, "dur": 0.081 , "args": { "External id": 1333,"Record function id": 0, "Ev Idx": 1332 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113364.395, "dur": 1.393 , "args": { "External id": 1334,"Record function id": 0, "Ev Idx": 1333 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113364.586, "dur": 1.092 , "args": { "External id": 1335,"Record function id": 0, "Ev Idx": 1334 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113377.149, "dur": 4.549 , "args": { "External id": 1336,"Record function id": 0, "Ev Idx": 1335 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113383.942, "dur": 0.831 , "args": { "External id": 1337,"Record function id": 0, "Ev Idx": 1336 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113386.527, "dur": 0.551 , "args": { "External id": 1338,"Record function id": 0, "Ev Idx": 1337 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381113390.945, "dur": 0.451 , "args": { "External id": 1339,"Record function id": 0, "Ev Idx": 1338 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381113391.225, "dur": 0.070 , "args": { "External id": 1340,"Record function id": 0, "Ev Idx": 1339 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381113394.411, "dur": 6.122 , "args": { "External id": 1341,"Record function id": 0, "Ev Idx": 1340 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113402.787, "dur": 10.780 , "args": { "External id": 1342,"Record function id": 0, "Ev Idx": 1341 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113403.318, "dur": 0.731 , "args": { "External id": 1343,"Record function id": 0, "Ev Idx": 1342 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113403.508, "dur": 0.451 , "args": { "External id": 1344,"Record function id": 0, "Ev Idx": 1343 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113403.749, "dur": 0.100 , "args": { "External id": 1345,"Record function id": 0, "Ev Idx": 1344 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113404.320, "dur": 0.631 , "args": { "External id": 1346,"Record function id": 0, "Ev Idx": 1345 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113404.480, "dur": 0.381 , "args": { "External id": 1347,"Record function id": 0, "Ev Idx": 1346 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113404.670, "dur": 0.080 , "args": { "External id": 1348,"Record function id": 0, "Ev Idx": 1347 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113405.201, "dur": 0.792 , "args": { "External id": 1349,"Record function id": 0, "Ev Idx": 1348 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113405.362, "dur": 0.541 , "args": { "External id": 1350,"Record function id": 0, "Ev Idx": 1349 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113405.632, "dur": 0.170 , "args": { "External id": 1351,"Record function id": 0, "Ev Idx": 1350 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113407.105, "dur": 1.282 , "args": { "External id": 1352,"Record function id": 0, "Ev Idx": 1351 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113407.305, "dur": 0.982 , "args": { "External id": 1353,"Record function id": 0, "Ev Idx": 1352 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113416.773, "dur": 3.867 , "args": { "External id": 1354,"Record function id": 0, "Ev Idx": 1353 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113421.712, "dur": 0.591 , "args": { "External id": 1355,"Record function id": 0, "Ev Idx": 1354 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113423.616, "dur": 0.571 , "args": { "External id": 1356,"Record function id": 0, "Ev Idx": 1355 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381113428.174, "dur": 0.381 , "args": { "External id": 1357,"Record function id": 0, "Ev Idx": 1356 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381113428.354, "dur": 0.101 , "args": { "External id": 1358,"Record function id": 0, "Ev Idx": 1357 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381113431.240, "dur": 6.562 , "args": { "External id": 1359,"Record function id": 0, "Ev Idx": 1358 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113439.996, "dur": 10.720 , "args": { "External id": 1360,"Record function id": 0, "Ev Idx": 1359 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113440.557, "dur": 0.751 , "args": { "External id": 1361,"Record function id": 0, "Ev Idx": 1360 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113440.757, "dur": 0.461 , "args": { "External id": 1362,"Record function id": 0, "Ev Idx": 1361 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113441.008, "dur": 0.110 , "args": { "External id": 1363,"Record function id": 0, "Ev Idx": 1362 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113441.579, "dur": 0.962 , "args": { "External id": 1364,"Record function id": 0, "Ev Idx": 1363 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113441.729, "dur": 0.722 , "args": { "External id": 1365,"Record function id": 0, "Ev Idx": 1364 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113442.200, "dur": 0.150 , "args": { "External id": 1366,"Record function id": 0, "Ev Idx": 1365 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113442.801, "dur": 0.692 , "args": { "External id": 1367,"Record function id": 0, "Ev Idx": 1366 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113442.962, "dur": 0.440 , "args": { "External id": 1368,"Record function id": 0, "Ev Idx": 1367 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113443.142, "dur": 0.160 , "args": { "External id": 1369,"Record function id": 0, "Ev Idx": 1368 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113444.534, "dur": 1.333 , "args": { "External id": 1370,"Record function id": 0, "Ev Idx": 1369 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113444.725, "dur": 1.032 , "args": { "External id": 1371,"Record function id": 0, "Ev Idx": 1370 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381113455.635, "dur": 4.138 , "args": { "External id": 1372,"Record function id": 0, "Ev Idx": 1371 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381113465.113, "dur": 5.189 , "args": { "External id": 1373,"Record function id": 0, "Ev Idx": 1372 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381113475.412, "dur": 4.959 , "args": { "External id": 1374,"Record function id": 0, "Ev Idx": 1373 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381113482.756, "dur": 17.352 , "args": { "External id": 1375,"Record function id": 0, "Ev Idx": 1374 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113483.216, "dur": 0.491 , "args": { "External id": 1376,"Record function id": 0, "Ev Idx": 1375 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113483.447, "dur": 0.150 , "args": { "External id": 1377,"Record function id": 0, "Ev Idx": 1376 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113483.948, "dur": 0.381 , "args": { "External id": 1378,"Record function id": 0, "Ev Idx": 1377 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113484.138, "dur": 0.090 , "args": { "External id": 1379,"Record function id": 0, "Ev Idx": 1378 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113484.569, "dur": 0.381 , "args": { "External id": 1380,"Record function id": 0, "Ev Idx": 1379 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113484.759, "dur": 0.091 , "args": { "External id": 1381,"Record function id": 0, "Ev Idx": 1380 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381113485.290, "dur": 13.165 , "args": { "External id": 1382,"Record function id": 0, "Ev Idx": 1381 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381113486.192, "dur": 1.934 , "args": { "External id": 1383,"Record function id": 0, "Ev Idx": 1382 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113486.483, "dur": 1.532 , "args": { "External id": 1384,"Record function id": 0, "Ev Idx": 1383 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113488.607, "dur": 1.242 , "args": { "External id": 1385,"Record function id": 0, "Ev Idx": 1384 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113490.280, "dur": 1.222 , "args": { "External id": 1386,"Record function id": 0, "Ev Idx": 1385 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113491.782, "dur": 1.383 , "args": { "External id": 1387,"Record function id": 0, "Ev Idx": 1386 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113497.132, "dur": 1.002 , "args": { "External id": 1388,"Record function id": 0, "Ev Idx": 1387 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113498.765, "dur": 0.612 , "args": { "External id": 1389,"Record function id": 0, "Ev Idx": 1388 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113498.986, "dur": 0.280 , "args": { "External id": 1390,"Record function id": 0, "Ev Idx": 1389 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113507.221, "dur": 6.322 , "args": { "External id": 1391,"Record function id": 0, "Ev Idx": 1390 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113515.897, "dur": 10.700 , "args": { "External id": 1392,"Record function id": 0, "Ev Idx": 1391 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113516.518, "dur": 0.712 , "args": { "External id": 1393,"Record function id": 0, "Ev Idx": 1392 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113516.709, "dur": 0.431 , "args": { "External id": 1394,"Record function id": 0, "Ev Idx": 1393 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113516.919, "dur": 0.100 , "args": { "External id": 1395,"Record function id": 0, "Ev Idx": 1394 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113517.510, "dur": 0.642 , "args": { "External id": 1396,"Record function id": 0, "Ev Idx": 1395 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113517.681, "dur": 0.380 , "args": { "External id": 1397,"Record function id": 0, "Ev Idx": 1396 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113517.891, "dur": 0.070 , "args": { "External id": 1398,"Record function id": 0, "Ev Idx": 1397 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113518.402, "dur": 0.661 , "args": { "External id": 1399,"Record function id": 0, "Ev Idx": 1398 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113518.582, "dur": 0.391 , "args": { "External id": 1400,"Record function id": 0, "Ev Idx": 1399 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113518.793, "dur": 0.070 , "args": { "External id": 1401,"Record function id": 0, "Ev Idx": 1400 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113520.155, "dur": 1.233 , "args": { "External id": 1402,"Record function id": 0, "Ev Idx": 1401 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113520.336, "dur": 0.951 , "args": { "External id": 1403,"Record function id": 0, "Ev Idx": 1402 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381113531.416, "dur": 5.240 , "args": { "External id": 1404,"Record function id": 0, "Ev Idx": 1403 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113543.178, "dur": 5.180 , "args": { "External id": 1405,"Record function id": 0, "Ev Idx": 1404 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113551.914, "dur": 4.910 , "args": { "External id": 1406,"Record function id": 0, "Ev Idx": 1405 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113560.751, "dur": 10.620 , "args": { "External id": 1407,"Record function id": 0, "Ev Idx": 1406 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113561.302, "dur": 0.691 , "args": { "External id": 1408,"Record function id": 0, "Ev Idx": 1407 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113561.482, "dur": 0.421 , "args": { "External id": 1409,"Record function id": 0, "Ev Idx": 1408 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113561.713, "dur": 0.090 , "args": { "External id": 1410,"Record function id": 0, "Ev Idx": 1409 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113562.264, "dur": 0.651 , "args": { "External id": 1411,"Record function id": 0, "Ev Idx": 1410 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113562.434, "dur": 0.391 , "args": { "External id": 1412,"Record function id": 0, "Ev Idx": 1411 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113562.644, "dur": 0.080 , "args": { "External id": 1413,"Record function id": 0, "Ev Idx": 1412 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113563.165, "dur": 0.631 , "args": { "External id": 1414,"Record function id": 0, "Ev Idx": 1413 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113563.346, "dur": 0.360 , "args": { "External id": 1415,"Record function id": 0, "Ev Idx": 1414 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113563.526, "dur": 0.080 , "args": { "External id": 1416,"Record function id": 0, "Ev Idx": 1415 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113564.909, "dur": 1.302 , "args": { "External id": 1417,"Record function id": 0, "Ev Idx": 1416 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113565.109, "dur": 1.002 , "args": { "External id": 1418,"Record function id": 0, "Ev Idx": 1417 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113575.208, "dur": 9.878 , "args": { "External id": 1419,"Record function id": 0, "Ev Idx": 1418 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113575.689, "dur": 0.641 , "args": { "External id": 1420,"Record function id": 0, "Ev Idx": 1419 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113575.859, "dur": 0.381 , "args": { "External id": 1421,"Record function id": 0, "Ev Idx": 1420 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113576.059, "dur": 0.070 , "args": { "External id": 1422,"Record function id": 0, "Ev Idx": 1421 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113576.590, "dur": 0.621 , "args": { "External id": 1423,"Record function id": 0, "Ev Idx": 1422 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113576.741, "dur": 0.380 , "args": { "External id": 1424,"Record function id": 0, "Ev Idx": 1423 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113576.951, "dur": 0.070 , "args": { "External id": 1425,"Record function id": 0, "Ev Idx": 1424 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113577.472, "dur": 0.631 , "args": { "External id": 1426,"Record function id": 0, "Ev Idx": 1425 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113577.612, "dur": 0.401 , "args": { "External id": 1427,"Record function id": 0, "Ev Idx": 1426 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113577.833, "dur": 0.070 , "args": { "External id": 1428,"Record function id": 0, "Ev Idx": 1427 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113579.105, "dur": 1.182 , "args": { "External id": 1429,"Record function id": 0, "Ev Idx": 1428 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113579.265, "dur": 0.912 , "args": { "External id": 1430,"Record function id": 0, "Ev Idx": 1429 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381113589.705, "dur": 4.458 , "args": { "External id": 1431,"Record function id": 0, "Ev Idx": 1430 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113597.369, "dur": 6.001 , "args": { "External id": 1432,"Record function id": 0, "Ev Idx": 1431 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113606.526, "dur": 10.179 , "args": { "External id": 1433,"Record function id": 0, "Ev Idx": 1432 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113607.037, "dur": 0.671 , "args": { "External id": 1434,"Record function id": 0, "Ev Idx": 1433 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113607.207, "dur": 0.411 , "args": { "External id": 1435,"Record function id": 0, "Ev Idx": 1434 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113607.418, "dur": 0.090 , "args": { "External id": 1436,"Record function id": 0, "Ev Idx": 1435 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113607.969, "dur": 0.651 , "args": { "External id": 1437,"Record function id": 0, "Ev Idx": 1436 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113608.139, "dur": 0.391 , "args": { "External id": 1438,"Record function id": 0, "Ev Idx": 1437 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113608.359, "dur": 0.071 , "args": { "External id": 1439,"Record function id": 0, "Ev Idx": 1438 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113608.870, "dur": 0.642 , "args": { "External id": 1440,"Record function id": 0, "Ev Idx": 1439 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113609.031, "dur": 0.390 , "args": { "External id": 1441,"Record function id": 0, "Ev Idx": 1440 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113609.231, "dur": 0.090 , "args": { "External id": 1442,"Record function id": 0, "Ev Idx": 1441 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113610.543, "dur": 1.243 , "args": { "External id": 1443,"Record function id": 0, "Ev Idx": 1442 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113610.724, "dur": 0.962 , "args": { "External id": 1444,"Record function id": 0, "Ev Idx": 1443 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381113621.484, "dur": 5.640 , "args": { "External id": 1445,"Record function id": 0, "Ev Idx": 1444 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381113648.023, "dur": 24.696 , "args": { "External id": 1446,"Record function id": 0, "Ev Idx": 1445 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381113673.110, "dur": 343.980 , "args": { "External id": 1447,"Record function id": 0, "Ev Idx": 1446 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381113678.159, "dur": 10.420 , "args": { "External id": 1448,"Record function id": 0, "Ev Idx": 1447 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113681.576, "dur": 0.451 , "args": { "External id": 1449,"Record function id": 0, "Ev Idx": 1448 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381113682.477, "dur": 0.050 , "args": { "External id": 1450,"Record function id": 0, "Ev Idx": 1449 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381113682.908, "dur": 0.060 , "args": { "External id": 1451,"Record function id": 0, "Ev Idx": 1450 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113684.211, "dur": 0.210 , "args": { "External id": 1452,"Record function id": 0, "Ev Idx": 1451 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381113684.762, "dur": 0.030 , "args": { "External id": 1453,"Record function id": 0, "Ev Idx": 1452 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381113685.102, "dur": 0.030 , "args": { "External id": 1454,"Record function id": 0, "Ev Idx": 1453 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381113691.374, "dur": 1.132 , "args": { "External id": 1455,"Record function id": 0, "Ev Idx": 1454 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381113694.630, "dur": 319.234 , "args": { "External id": 1456,"Record function id": 0, "Ev Idx": 1455 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113703.236, "dur": 6.512 , "args": { "External id": 1457,"Record function id": 0, "Ev Idx": 1456 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113716.421, "dur": 7.203 , "args": { "External id": 1458,"Record function id": 0, "Ev Idx": 1457 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113729.545, "dur": 5.310 , "args": { "External id": 1459,"Record function id": 0, "Ev Idx": 1458 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113740.495, "dur": 5.060 , "args": { "External id": 1460,"Record function id": 0, "Ev Idx": 1459 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113747.929, "dur": 12.483 , "args": { "External id": 1461,"Record function id": 0, "Ev Idx": 1460 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113748.540, "dur": 0.932 , "args": { "External id": 1462,"Record function id": 0, "Ev Idx": 1461 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113748.751, "dur": 0.611 , "args": { "External id": 1463,"Record function id": 0, "Ev Idx": 1462 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113749.051, "dur": 0.171 , "args": { "External id": 1464,"Record function id": 0, "Ev Idx": 1463 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113749.773, "dur": 0.641 , "args": { "External id": 1465,"Record function id": 0, "Ev Idx": 1464 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113749.943, "dur": 0.381 , "args": { "External id": 1466,"Record function id": 0, "Ev Idx": 1465 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113750.153, "dur": 0.071 , "args": { "External id": 1467,"Record function id": 0, "Ev Idx": 1466 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113750.664, "dur": 0.812 , "args": { "External id": 1468,"Record function id": 0, "Ev Idx": 1467 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113750.825, "dur": 0.561 , "args": { "External id": 1469,"Record function id": 0, "Ev Idx": 1468 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113751.045, "dur": 0.240 , "args": { "External id": 1470,"Record function id": 0, "Ev Idx": 1469 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113752.798, "dur": 1.493 , "args": { "External id": 1471,"Record function id": 0, "Ev Idx": 1470 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113753.009, "dur": 1.172 , "args": { "External id": 1472,"Record function id": 0, "Ev Idx": 1471 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113765.933, "dur": 4.598 , "args": { "External id": 1473,"Record function id": 0, "Ev Idx": 1472 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113772.936, "dur": 0.911 , "args": { "External id": 1474,"Record function id": 0, "Ev Idx": 1473 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113775.671, "dur": 0.561 , "args": { "External id": 1475,"Record function id": 0, "Ev Idx": 1474 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381113780.079, "dur": 0.461 , "args": { "External id": 1476,"Record function id": 0, "Ev Idx": 1475 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381113780.360, "dur": 0.060 , "args": { "External id": 1477,"Record function id": 0, "Ev Idx": 1476 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381113783.696, "dur": 7.123 , "args": { "External id": 1478,"Record function id": 0, "Ev Idx": 1477 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113793.143, "dur": 10.590 , "args": { "External id": 1479,"Record function id": 0, "Ev Idx": 1478 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113793.674, "dur": 0.842 , "args": { "External id": 1480,"Record function id": 0, "Ev Idx": 1479 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113793.855, "dur": 0.561 , "args": { "External id": 1481,"Record function id": 0, "Ev Idx": 1480 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113794.135, "dur": 0.181 , "args": { "External id": 1482,"Record function id": 0, "Ev Idx": 1481 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113794.796, "dur": 0.622 , "args": { "External id": 1483,"Record function id": 0, "Ev Idx": 1482 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113794.957, "dur": 0.370 , "args": { "External id": 1484,"Record function id": 0, "Ev Idx": 1483 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113795.157, "dur": 0.070 , "args": { "External id": 1485,"Record function id": 0, "Ev Idx": 1484 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113795.678, "dur": 0.651 , "args": { "External id": 1486,"Record function id": 0, "Ev Idx": 1485 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113795.838, "dur": 0.401 , "args": { "External id": 1487,"Record function id": 0, "Ev Idx": 1486 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113796.029, "dur": 0.110 , "args": { "External id": 1488,"Record function id": 0, "Ev Idx": 1487 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113797.411, "dur": 1.303 , "args": { "External id": 1489,"Record function id": 0, "Ev Idx": 1488 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113797.612, "dur": 0.992 , "args": { "External id": 1490,"Record function id": 0, "Ev Idx": 1489 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113806.969, "dur": 3.777 , "args": { "External id": 1491,"Record function id": 0, "Ev Idx": 1490 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113811.808, "dur": 0.601 , "args": { "External id": 1492,"Record function id": 0, "Ev Idx": 1491 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381113813.722, "dur": 0.551 , "args": { "External id": 1493,"Record function id": 0, "Ev Idx": 1492 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381113818.140, "dur": 0.351 , "args": { "External id": 1494,"Record function id": 0, "Ev Idx": 1493 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381113818.330, "dur": 0.050 , "args": { "External id": 1495,"Record function id": 0, "Ev Idx": 1494 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381113821.136, "dur": 6.081 , "args": { "External id": 1496,"Record function id": 0, "Ev Idx": 1495 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113829.431, "dur": 10.540 , "args": { "External id": 1497,"Record function id": 0, "Ev Idx": 1496 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113829.982, "dur": 0.711 , "args": { "External id": 1498,"Record function id": 0, "Ev Idx": 1497 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113830.172, "dur": 0.421 , "args": { "External id": 1499,"Record function id": 0, "Ev Idx": 1498 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113830.413, "dur": 0.080 , "args": { "External id": 1500,"Record function id": 0, "Ev Idx": 1499 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113830.954, "dur": 0.621 , "args": { "External id": 1501,"Record function id": 0, "Ev Idx": 1500 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113831.114, "dur": 0.371 , "args": { "External id": 1502,"Record function id": 0, "Ev Idx": 1501 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113831.304, "dur": 0.081 , "args": { "External id": 1503,"Record function id": 0, "Ev Idx": 1502 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113831.825, "dur": 0.692 , "args": { "External id": 1504,"Record function id": 0, "Ev Idx": 1503 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113831.976, "dur": 0.451 , "args": { "External id": 1505,"Record function id": 0, "Ev Idx": 1504 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113832.176, "dur": 0.150 , "args": { "External id": 1506,"Record function id": 0, "Ev Idx": 1505 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113833.569, "dur": 1.272 , "args": { "External id": 1507,"Record function id": 0, "Ev Idx": 1506 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113833.749, "dur": 0.992 , "args": { "External id": 1508,"Record function id": 0, "Ev Idx": 1507 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381113844.900, "dur": 4.228 , "args": { "External id": 1509,"Record function id": 0, "Ev Idx": 1508 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381113854.387, "dur": 4.709 , "args": { "External id": 1510,"Record function id": 0, "Ev Idx": 1509 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381113864.075, "dur": 4.779 , "args": { "External id": 1511,"Record function id": 0, "Ev Idx": 1510 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381113871.349, "dur": 18.014 , "args": { "External id": 1512,"Record function id": 0, "Ev Idx": 1511 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113871.860, "dur": 0.521 , "args": { "External id": 1513,"Record function id": 0, "Ev Idx": 1512 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113872.100, "dur": 0.161 , "args": { "External id": 1514,"Record function id": 0, "Ev Idx": 1513 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113872.611, "dur": 0.411 , "args": { "External id": 1515,"Record function id": 0, "Ev Idx": 1514 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113872.822, "dur": 0.090 , "args": { "External id": 1516,"Record function id": 0, "Ev Idx": 1515 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113873.263, "dur": 0.651 , "args": { "External id": 1517,"Record function id": 0, "Ev Idx": 1516 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113873.713, "dur": 0.091 , "args": { "External id": 1518,"Record function id": 0, "Ev Idx": 1517 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381113874.304, "dur": 13.456 , "args": { "External id": 1519,"Record function id": 0, "Ev Idx": 1518 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381113875.246, "dur": 1.814 , "args": { "External id": 1520,"Record function id": 0, "Ev Idx": 1519 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113875.507, "dur": 1.442 , "args": { "External id": 1521,"Record function id": 0, "Ev Idx": 1520 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113877.490, "dur": 1.223 , "args": { "External id": 1522,"Record function id": 0, "Ev Idx": 1521 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113879.143, "dur": 1.273 , "args": { "External id": 1523,"Record function id": 0, "Ev Idx": 1522 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113880.686, "dur": 1.473 , "args": { "External id": 1524,"Record function id": 0, "Ev Idx": 1523 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113886.397, "dur": 1.002 , "args": { "External id": 1525,"Record function id": 0, "Ev Idx": 1524 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113888.050, "dur": 0.551 , "args": { "External id": 1526,"Record function id": 0, "Ev Idx": 1525 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113888.270, "dur": 0.231 , "args": { "External id": 1527,"Record function id": 0, "Ev Idx": 1526 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113896.115, "dur": 6.342 , "args": { "External id": 1528,"Record function id": 0, "Ev Idx": 1527 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113904.851, "dur": 13.906 , "args": { "External id": 1529,"Record function id": 0, "Ev Idx": 1528 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113905.462, "dur": 0.682 , "args": { "External id": 1530,"Record function id": 0, "Ev Idx": 1529 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113905.633, "dur": 0.411 , "args": { "External id": 1531,"Record function id": 0, "Ev Idx": 1530 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113905.853, "dur": 0.080 , "args": { "External id": 1532,"Record function id": 0, "Ev Idx": 1531 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113906.404, "dur": 0.651 , "args": { "External id": 1533,"Record function id": 0, "Ev Idx": 1532 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113906.575, "dur": 0.390 , "args": { "External id": 1534,"Record function id": 0, "Ev Idx": 1533 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113906.785, "dur": 0.080 , "args": { "External id": 1535,"Record function id": 0, "Ev Idx": 1534 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113907.326, "dur": 3.697 , "args": { "External id": 1536,"Record function id": 0, "Ev Idx": 1535 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113910.512, "dur": 0.411 , "args": { "External id": 1537,"Record function id": 0, "Ev Idx": 1536 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113910.732, "dur": 0.090 , "args": { "External id": 1538,"Record function id": 0, "Ev Idx": 1537 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113912.195, "dur": 1.292 , "args": { "External id": 1539,"Record function id": 0, "Ev Idx": 1538 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113912.395, "dur": 0.982 , "args": { "External id": 1540,"Record function id": 0, "Ev Idx": 1539 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381113922.955, "dur": 5.400 , "args": { "External id": 1541,"Record function id": 0, "Ev Idx": 1540 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113931.521, "dur": 5.220 , "args": { "External id": 1542,"Record function id": 0, "Ev Idx": 1541 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381113940.087, "dur": 4.919 , "args": { "External id": 1543,"Record function id": 0, "Ev Idx": 1542 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113948.553, "dur": 10.159 , "args": { "External id": 1544,"Record function id": 0, "Ev Idx": 1543 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113949.064, "dur": 0.661 , "args": { "External id": 1545,"Record function id": 0, "Ev Idx": 1544 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113949.244, "dur": 0.391 , "args": { "External id": 1546,"Record function id": 0, "Ev Idx": 1545 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113949.454, "dur": 0.081 , "args": { "External id": 1547,"Record function id": 0, "Ev Idx": 1546 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113950.025, "dur": 0.662 , "args": { "External id": 1548,"Record function id": 0, "Ev Idx": 1547 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113950.206, "dur": 0.390 , "args": { "External id": 1549,"Record function id": 0, "Ev Idx": 1548 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113950.426, "dur": 0.070 , "args": { "External id": 1550,"Record function id": 0, "Ev Idx": 1549 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113950.937, "dur": 0.631 , "args": { "External id": 1551,"Record function id": 0, "Ev Idx": 1550 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113951.097, "dur": 0.381 , "args": { "External id": 1552,"Record function id": 0, "Ev Idx": 1551 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113951.298, "dur": 0.070 , "args": { "External id": 1553,"Record function id": 0, "Ev Idx": 1552 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113952.580, "dur": 1.263 , "args": { "External id": 1554,"Record function id": 0, "Ev Idx": 1553 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113952.771, "dur": 0.971 , "args": { "External id": 1555,"Record function id": 0, "Ev Idx": 1554 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113962.509, "dur": 9.938 , "args": { "External id": 1556,"Record function id": 0, "Ev Idx": 1555 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113963.010, "dur": 0.621 , "args": { "External id": 1557,"Record function id": 0, "Ev Idx": 1556 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113963.160, "dur": 0.381 , "args": { "External id": 1558,"Record function id": 0, "Ev Idx": 1557 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113963.360, "dur": 0.080 , "args": { "External id": 1559,"Record function id": 0, "Ev Idx": 1558 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113963.891, "dur": 0.621 , "args": { "External id": 1560,"Record function id": 0, "Ev Idx": 1559 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113964.042, "dur": 0.380 , "args": { "External id": 1561,"Record function id": 0, "Ev Idx": 1560 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113964.242, "dur": 0.070 , "args": { "External id": 1562,"Record function id": 0, "Ev Idx": 1561 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113964.763, "dur": 0.631 , "args": { "External id": 1563,"Record function id": 0, "Ev Idx": 1562 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113964.923, "dur": 0.381 , "args": { "External id": 1564,"Record function id": 0, "Ev Idx": 1563 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113965.124, "dur": 0.070 , "args": { "External id": 1565,"Record function id": 0, "Ev Idx": 1564 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113966.406, "dur": 1.232 , "args": { "External id": 1566,"Record function id": 0, "Ev Idx": 1565 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113966.576, "dur": 0.962 , "args": { "External id": 1567,"Record function id": 0, "Ev Idx": 1566 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381113977.006, "dur": 4.127 , "args": { "External id": 1568,"Record function id": 0, "Ev Idx": 1567 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381113984.269, "dur": 5.160 , "args": { "External id": 1569,"Record function id": 0, "Ev Idx": 1568 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381113992.635, "dur": 10.319 , "args": { "External id": 1570,"Record function id": 0, "Ev Idx": 1569 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113993.146, "dur": 0.671 , "args": { "External id": 1571,"Record function id": 0, "Ev Idx": 1570 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113993.326, "dur": 0.401 , "args": { "External id": 1572,"Record function id": 0, "Ev Idx": 1571 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113993.546, "dur": 0.081 , "args": { "External id": 1573,"Record function id": 0, "Ev Idx": 1572 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113994.087, "dur": 0.672 , "args": { "External id": 1574,"Record function id": 0, "Ev Idx": 1573 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113994.248, "dur": 0.421 , "args": { "External id": 1575,"Record function id": 0, "Ev Idx": 1574 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113994.488, "dur": 0.070 , "args": { "External id": 1576,"Record function id": 0, "Ev Idx": 1575 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381113995.019, "dur": 0.621 , "args": { "External id": 1577,"Record function id": 0, "Ev Idx": 1576 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381113995.169, "dur": 0.381 , "args": { "External id": 1578,"Record function id": 0, "Ev Idx": 1577 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381113995.360, "dur": 0.080 , "args": { "External id": 1579,"Record function id": 0, "Ev Idx": 1578 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113996.692, "dur": 1.263 , "args": { "External id": 1580,"Record function id": 0, "Ev Idx": 1579 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381113996.893, "dur": 0.951 , "args": { "External id": 1581,"Record function id": 0, "Ev Idx": 1580 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381114007.593, "dur": 4.648 , "args": { "External id": 1582,"Record function id": 0, "Ev Idx": 1581 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381114032.789, "dur": 23.514 , "args": { "External id": 1583,"Record function id": 0, "Ev Idx": 1582 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381114056.644, "dur": 340.924 , "args": { "External id": 1584,"Record function id": 0, "Ev Idx": 1583 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381114061.523, "dur": 10.940 , "args": { "External id": 1585,"Record function id": 0, "Ev Idx": 1584 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114065.380, "dur": 0.351 , "args": { "External id": 1586,"Record function id": 0, "Ev Idx": 1585 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381114066.192, "dur": 0.040 , "args": { "External id": 1587,"Record function id": 0, "Ev Idx": 1586 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381114066.612, "dur": 0.051 , "args": { "External id": 1588,"Record function id": 0, "Ev Idx": 1587 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114067.875, "dur": 0.260 , "args": { "External id": 1589,"Record function id": 0, "Ev Idx": 1588 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381114068.496, "dur": 0.020 , "args": { "External id": 1590,"Record function id": 0, "Ev Idx": 1589 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381114068.867, "dur": 0.030 , "args": { "External id": 1591,"Record function id": 0, "Ev Idx": 1590 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381114075.208, "dur": 1.002 , "args": { "External id": 1592,"Record function id": 0, "Ev Idx": 1591 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381114078.485, "dur": 316.088 , "args": { "External id": 1593,"Record function id": 0, "Ev Idx": 1592 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114087.491, "dur": 6.683 , "args": { "External id": 1594,"Record function id": 0, "Ev Idx": 1593 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114100.646, "dur": 7.223 , "args": { "External id": 1595,"Record function id": 0, "Ev Idx": 1594 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114113.900, "dur": 5.861 , "args": { "External id": 1596,"Record function id": 0, "Ev Idx": 1595 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114125.662, "dur": 5.120 , "args": { "External id": 1597,"Record function id": 0, "Ev Idx": 1596 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114133.397, "dur": 12.283 , "args": { "External id": 1598,"Record function id": 0, "Ev Idx": 1597 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114133.988, "dur": 1.002 , "args": { "External id": 1599,"Record function id": 0, "Ev Idx": 1598 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114134.208, "dur": 0.672 , "args": { "External id": 1600,"Record function id": 0, "Ev Idx": 1599 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114134.499, "dur": 0.230 , "args": { "External id": 1601,"Record function id": 0, "Ev Idx": 1600 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114135.270, "dur": 0.631 , "args": { "External id": 1602,"Record function id": 0, "Ev Idx": 1601 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114135.441, "dur": 0.370 , "args": { "External id": 1603,"Record function id": 0, "Ev Idx": 1602 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114135.631, "dur": 0.080 , "args": { "External id": 1604,"Record function id": 0, "Ev Idx": 1603 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114136.152, "dur": 0.711 , "args": { "External id": 1605,"Record function id": 0, "Ev Idx": 1604 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114136.322, "dur": 0.451 , "args": { "External id": 1606,"Record function id": 0, "Ev Idx": 1605 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114136.533, "dur": 0.140 , "args": { "External id": 1607,"Record function id": 0, "Ev Idx": 1606 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114138.106, "dur": 1.462 , "args": { "External id": 1608,"Record function id": 0, "Ev Idx": 1607 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114138.316, "dur": 1.152 , "args": { "External id": 1609,"Record function id": 0, "Ev Idx": 1608 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114151.200, "dur": 4.669 , "args": { "External id": 1610,"Record function id": 0, "Ev Idx": 1609 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114158.033, "dur": 0.841 , "args": { "External id": 1611,"Record function id": 0, "Ev Idx": 1610 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114160.688, "dur": 0.531 , "args": { "External id": 1612,"Record function id": 0, "Ev Idx": 1611 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381114165.116, "dur": 0.421 , "args": { "External id": 1613,"Record function id": 0, "Ev Idx": 1612 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381114165.376, "dur": 0.060 , "args": { "External id": 1614,"Record function id": 0, "Ev Idx": 1613 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381114168.562, "dur": 6.402 , "args": { "External id": 1615,"Record function id": 0, "Ev Idx": 1614 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114177.008, "dur": 10.910 , "args": { "External id": 1616,"Record function id": 0, "Ev Idx": 1615 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114177.539, "dur": 0.731 , "args": { "External id": 1617,"Record function id": 0, "Ev Idx": 1616 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114177.709, "dur": 0.471 , "args": { "External id": 1618,"Record function id": 0, "Ev Idx": 1617 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114177.950, "dur": 0.120 , "args": { "External id": 1619,"Record function id": 0, "Ev Idx": 1618 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114178.551, "dur": 0.651 , "args": { "External id": 1620,"Record function id": 0, "Ev Idx": 1619 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114178.711, "dur": 0.401 , "args": { "External id": 1621,"Record function id": 0, "Ev Idx": 1620 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114178.921, "dur": 0.081 , "args": { "External id": 1622,"Record function id": 0, "Ev Idx": 1621 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114179.463, "dur": 0.951 , "args": { "External id": 1623,"Record function id": 0, "Ev Idx": 1622 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114179.633, "dur": 0.691 , "args": { "External id": 1624,"Record function id": 0, "Ev Idx": 1623 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114180.064, "dur": 0.160 , "args": { "External id": 1625,"Record function id": 0, "Ev Idx": 1624 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114181.516, "dur": 1.343 , "args": { "External id": 1626,"Record function id": 0, "Ev Idx": 1625 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114181.767, "dur": 0.992 , "args": { "External id": 1627,"Record function id": 0, "Ev Idx": 1626 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114191.174, "dur": 3.827 , "args": { "External id": 1628,"Record function id": 0, "Ev Idx": 1627 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114196.063, "dur": 0.622 , "args": { "External id": 1629,"Record function id": 0, "Ev Idx": 1628 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114197.937, "dur": 0.571 , "args": { "External id": 1630,"Record function id": 0, "Ev Idx": 1629 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381114202.185, "dur": 0.330 , "args": { "External id": 1631,"Record function id": 0, "Ev Idx": 1630 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381114202.375, "dur": 0.040 , "args": { "External id": 1632,"Record function id": 0, "Ev Idx": 1631 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381114205.190, "dur": 5.941 , "args": { "External id": 1633,"Record function id": 0, "Ev Idx": 1632 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114213.265, "dur": 10.560 , "args": { "External id": 1634,"Record function id": 0, "Ev Idx": 1633 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114213.836, "dur": 0.712 , "args": { "External id": 1635,"Record function id": 0, "Ev Idx": 1634 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114214.017, "dur": 0.441 , "args": { "External id": 1636,"Record function id": 0, "Ev Idx": 1635 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114214.257, "dur": 0.090 , "args": { "External id": 1637,"Record function id": 0, "Ev Idx": 1636 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114214.818, "dur": 0.611 , "args": { "External id": 1638,"Record function id": 0, "Ev Idx": 1637 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114214.979, "dur": 0.360 , "args": { "External id": 1639,"Record function id": 0, "Ev Idx": 1638 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114215.169, "dur": 0.070 , "args": { "External id": 1640,"Record function id": 0, "Ev Idx": 1639 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114215.690, "dur": 0.701 , "args": { "External id": 1641,"Record function id": 0, "Ev Idx": 1640 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114215.840, "dur": 0.461 , "args": { "External id": 1642,"Record function id": 0, "Ev Idx": 1641 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114216.041, "dur": 0.150 , "args": { "External id": 1643,"Record function id": 0, "Ev Idx": 1642 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114217.453, "dur": 1.313 , "args": { "External id": 1644,"Record function id": 0, "Ev Idx": 1643 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114217.654, "dur": 1.001 , "args": { "External id": 1645,"Record function id": 0, "Ev Idx": 1644 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381114228.854, "dur": 4.248 , "args": { "External id": 1646,"Record function id": 0, "Ev Idx": 1645 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381114238.603, "dur": 4.829 , "args": { "External id": 1647,"Record function id": 0, "Ev Idx": 1646 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381114248.441, "dur": 4.338 , "args": { "External id": 1648,"Record function id": 0, "Ev Idx": 1647 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381114255.013, "dur": 18.064 , "args": { "External id": 1649,"Record function id": 0, "Ev Idx": 1648 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114255.534, "dur": 0.511 , "args": { "External id": 1650,"Record function id": 0, "Ev Idx": 1649 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114255.765, "dur": 0.160 , "args": { "External id": 1651,"Record function id": 0, "Ev Idx": 1650 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114256.296, "dur": 0.430 , "args": { "External id": 1652,"Record function id": 0, "Ev Idx": 1651 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114256.506, "dur": 0.110 , "args": { "External id": 1653,"Record function id": 0, "Ev Idx": 1652 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114256.967, "dur": 0.401 , "args": { "External id": 1654,"Record function id": 0, "Ev Idx": 1653 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114257.177, "dur": 0.080 , "args": { "External id": 1655,"Record function id": 0, "Ev Idx": 1654 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381114257.718, "dur": 13.816 , "args": { "External id": 1656,"Record function id": 0, "Ev Idx": 1655 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381114258.840, "dur": 2.014 , "args": { "External id": 1657,"Record function id": 0, "Ev Idx": 1656 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114259.181, "dur": 1.563 , "args": { "External id": 1658,"Record function id": 0, "Ev Idx": 1657 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114261.295, "dur": 1.222 , "args": { "External id": 1659,"Record function id": 0, "Ev Idx": 1658 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114262.948, "dur": 1.242 , "args": { "External id": 1660,"Record function id": 0, "Ev Idx": 1659 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114264.461, "dur": 1.543 , "args": { "External id": 1661,"Record function id": 0, "Ev Idx": 1660 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114270.171, "dur": 1.012 , "args": { "External id": 1662,"Record function id": 0, "Ev Idx": 1661 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114271.804, "dur": 0.531 , "args": { "External id": 1663,"Record function id": 0, "Ev Idx": 1662 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114272.015, "dur": 0.210 , "args": { "External id": 1664,"Record function id": 0, "Ev Idx": 1663 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114280.010, "dur": 6.231 , "args": { "External id": 1665,"Record function id": 0, "Ev Idx": 1664 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114288.666, "dur": 10.600 , "args": { "External id": 1666,"Record function id": 0, "Ev Idx": 1665 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114289.277, "dur": 0.671 , "args": { "External id": 1667,"Record function id": 0, "Ev Idx": 1666 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114289.447, "dur": 0.411 , "args": { "External id": 1668,"Record function id": 0, "Ev Idx": 1667 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114289.658, "dur": 0.080 , "args": { "External id": 1669,"Record function id": 0, "Ev Idx": 1668 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114290.229, "dur": 0.651 , "args": { "External id": 1670,"Record function id": 0, "Ev Idx": 1669 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114290.389, "dur": 0.401 , "args": { "External id": 1671,"Record function id": 0, "Ev Idx": 1670 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114290.609, "dur": 0.081 , "args": { "External id": 1672,"Record function id": 0, "Ev Idx": 1671 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114291.120, "dur": 0.652 , "args": { "External id": 1673,"Record function id": 0, "Ev Idx": 1672 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114291.291, "dur": 0.390 , "args": { "External id": 1674,"Record function id": 0, "Ev Idx": 1673 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114291.501, "dur": 0.080 , "args": { "External id": 1675,"Record function id": 0, "Ev Idx": 1674 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114292.854, "dur": 1.262 , "args": { "External id": 1676,"Record function id": 0, "Ev Idx": 1675 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114293.024, "dur": 0.992 , "args": { "External id": 1677,"Record function id": 0, "Ev Idx": 1676 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381114303.173, "dur": 5.179 , "args": { "External id": 1678,"Record function id": 0, "Ev Idx": 1677 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114311.198, "dur": 5.149 , "args": { "External id": 1679,"Record function id": 0, "Ev Idx": 1678 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114319.563, "dur": 4.929 , "args": { "External id": 1680,"Record function id": 0, "Ev Idx": 1679 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114328.099, "dur": 10.239 , "args": { "External id": 1681,"Record function id": 0, "Ev Idx": 1680 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114328.640, "dur": 0.671 , "args": { "External id": 1682,"Record function id": 0, "Ev Idx": 1681 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114328.821, "dur": 0.400 , "args": { "External id": 1683,"Record function id": 0, "Ev Idx": 1682 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114329.031, "dur": 0.080 , "args": { "External id": 1684,"Record function id": 0, "Ev Idx": 1683 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114329.592, "dur": 0.631 , "args": { "External id": 1685,"Record function id": 0, "Ev Idx": 1684 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114329.752, "dur": 0.381 , "args": { "External id": 1686,"Record function id": 0, "Ev Idx": 1685 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114329.963, "dur": 0.070 , "args": { "External id": 1687,"Record function id": 0, "Ev Idx": 1686 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114330.484, "dur": 0.651 , "args": { "External id": 1688,"Record function id": 0, "Ev Idx": 1687 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114330.644, "dur": 0.401 , "args": { "External id": 1689,"Record function id": 0, "Ev Idx": 1688 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114330.844, "dur": 0.090 , "args": { "External id": 1690,"Record function id": 0, "Ev Idx": 1689 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114332.177, "dur": 1.232 , "args": { "External id": 1691,"Record function id": 0, "Ev Idx": 1690 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114332.357, "dur": 0.952 , "args": { "External id": 1692,"Record function id": 0, "Ev Idx": 1691 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114342.075, "dur": 9.798 , "args": { "External id": 1693,"Record function id": 0, "Ev Idx": 1692 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114342.596, "dur": 0.621 , "args": { "External id": 1694,"Record function id": 0, "Ev Idx": 1693 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114342.746, "dur": 0.381 , "args": { "External id": 1695,"Record function id": 0, "Ev Idx": 1694 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114342.957, "dur": 0.070 , "args": { "External id": 1696,"Record function id": 0, "Ev Idx": 1695 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114343.488, "dur": 0.611 , "args": { "External id": 1697,"Record function id": 0, "Ev Idx": 1696 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114343.648, "dur": 0.361 , "args": { "External id": 1698,"Record function id": 0, "Ev Idx": 1697 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114343.838, "dur": 0.071 , "args": { "External id": 1699,"Record function id": 0, "Ev Idx": 1698 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114344.359, "dur": 0.622 , "args": { "External id": 1700,"Record function id": 0, "Ev Idx": 1699 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114344.510, "dur": 0.380 , "args": { "External id": 1701,"Record function id": 0, "Ev Idx": 1700 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114344.710, "dur": 0.080 , "args": { "External id": 1702,"Record function id": 0, "Ev Idx": 1701 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114345.992, "dur": 1.223 , "args": { "External id": 1703,"Record function id": 0, "Ev Idx": 1702 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114346.173, "dur": 0.942 , "args": { "External id": 1704,"Record function id": 0, "Ev Idx": 1703 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381114356.532, "dur": 4.118 , "args": { "External id": 1705,"Record function id": 0, "Ev Idx": 1704 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114364.006, "dur": 5.350 , "args": { "External id": 1706,"Record function id": 0, "Ev Idx": 1705 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114372.782, "dur": 10.540 , "args": { "External id": 1707,"Record function id": 0, "Ev Idx": 1706 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114373.303, "dur": 0.752 , "args": { "External id": 1708,"Record function id": 0, "Ev Idx": 1707 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114373.474, "dur": 0.461 , "args": { "External id": 1709,"Record function id": 0, "Ev Idx": 1708 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114373.684, "dur": 0.110 , "args": { "External id": 1710,"Record function id": 0, "Ev Idx": 1709 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114374.335, "dur": 0.671 , "args": { "External id": 1711,"Record function id": 0, "Ev Idx": 1710 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114374.496, "dur": 0.420 , "args": { "External id": 1712,"Record function id": 0, "Ev Idx": 1711 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114374.736, "dur": 0.070 , "args": { "External id": 1713,"Record function id": 0, "Ev Idx": 1712 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114375.267, "dur": 0.611 , "args": { "External id": 1714,"Record function id": 0, "Ev Idx": 1713 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114375.427, "dur": 0.361 , "args": { "External id": 1715,"Record function id": 0, "Ev Idx": 1714 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114375.608, "dur": 0.080 , "args": { "External id": 1716,"Record function id": 0, "Ev Idx": 1715 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114376.920, "dur": 1.292 , "args": { "External id": 1717,"Record function id": 0, "Ev Idx": 1716 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114377.110, "dur": 0.992 , "args": { "External id": 1718,"Record function id": 0, "Ev Idx": 1717 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381114388.141, "dur": 4.809 , "args": { "External id": 1719,"Record function id": 0, "Ev Idx": 1718 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381114412.837, "dur": 23.955 , "args": { "External id": 1720,"Record function id": 0, "Ev Idx": 1719 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381114437.142, "dur": 343.379 , "args": { "External id": 1721,"Record function id": 0, "Ev Idx": 1720 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381114442.011, "dur": 10.740 , "args": { "External id": 1722,"Record function id": 0, "Ev Idx": 1721 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114445.367, "dur": 0.401 , "args": { "External id": 1723,"Record function id": 0, "Ev Idx": 1722 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381114446.259, "dur": 0.040 , "args": { "External id": 1724,"Record function id": 0, "Ev Idx": 1723 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381114446.690, "dur": 0.030 , "args": { "External id": 1725,"Record function id": 0, "Ev Idx": 1724 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114448.143, "dur": 0.270 , "args": { "External id": 1726,"Record function id": 0, "Ev Idx": 1725 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381114448.744, "dur": 0.030 , "args": { "External id": 1727,"Record function id": 0, "Ev Idx": 1726 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381114449.104, "dur": 0.020 , "args": { "External id": 1728,"Record function id": 0, "Ev Idx": 1727 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381114455.617, "dur": 1.142 , "args": { "External id": 1729,"Record function id": 0, "Ev Idx": 1728 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381114458.843, "dur": 318.553 , "args": { "External id": 1730,"Record function id": 0, "Ev Idx": 1729 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114467.298, "dur": 6.472 , "args": { "External id": 1731,"Record function id": 0, "Ev Idx": 1730 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114480.473, "dur": 7.243 , "args": { "External id": 1732,"Record function id": 0, "Ev Idx": 1731 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114493.597, "dur": 5.450 , "args": { "External id": 1733,"Record function id": 0, "Ev Idx": 1732 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114504.538, "dur": 5.179 , "args": { "External id": 1734,"Record function id": 0, "Ev Idx": 1733 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114512.342, "dur": 12.313 , "args": { "External id": 1735,"Record function id": 0, "Ev Idx": 1734 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114512.963, "dur": 0.982 , "args": { "External id": 1736,"Record function id": 0, "Ev Idx": 1735 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114513.184, "dur": 0.661 , "args": { "External id": 1737,"Record function id": 0, "Ev Idx": 1736 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114513.484, "dur": 0.211 , "args": { "External id": 1738,"Record function id": 0, "Ev Idx": 1737 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114514.236, "dur": 0.631 , "args": { "External id": 1739,"Record function id": 0, "Ev Idx": 1738 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114514.396, "dur": 0.381 , "args": { "External id": 1740,"Record function id": 0, "Ev Idx": 1739 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114514.596, "dur": 0.070 , "args": { "External id": 1741,"Record function id": 0, "Ev Idx": 1740 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114515.137, "dur": 0.712 , "args": { "External id": 1742,"Record function id": 0, "Ev Idx": 1741 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114515.288, "dur": 0.470 , "args": { "External id": 1743,"Record function id": 0, "Ev Idx": 1742 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114515.488, "dur": 0.170 , "args": { "External id": 1744,"Record function id": 0, "Ev Idx": 1743 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114517.111, "dur": 1.433 , "args": { "External id": 1745,"Record function id": 0, "Ev Idx": 1744 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114517.311, "dur": 1.122 , "args": { "External id": 1746,"Record function id": 0, "Ev Idx": 1745 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114531.047, "dur": 4.418 , "args": { "External id": 1747,"Record function id": 0, "Ev Idx": 1746 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114540.324, "dur": 0.892 , "args": { "External id": 1748,"Record function id": 0, "Ev Idx": 1747 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114543.069, "dur": 0.611 , "args": { "External id": 1749,"Record function id": 0, "Ev Idx": 1748 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381114547.407, "dur": 0.481 , "args": { "External id": 1750,"Record function id": 0, "Ev Idx": 1749 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381114547.688, "dur": 0.080 , "args": { "External id": 1751,"Record function id": 0, "Ev Idx": 1750 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381114550.814, "dur": 6.121 , "args": { "External id": 1752,"Record function id": 0, "Ev Idx": 1751 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114559.149, "dur": 10.951 , "args": { "External id": 1753,"Record function id": 0, "Ev Idx": 1752 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114559.700, "dur": 0.762 , "args": { "External id": 1754,"Record function id": 0, "Ev Idx": 1753 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114559.871, "dur": 0.491 , "args": { "External id": 1755,"Record function id": 0, "Ev Idx": 1754 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114560.151, "dur": 0.100 , "args": { "External id": 1756,"Record function id": 0, "Ev Idx": 1755 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114560.742, "dur": 0.641 , "args": { "External id": 1757,"Record function id": 0, "Ev Idx": 1756 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114560.913, "dur": 0.380 , "args": { "External id": 1758,"Record function id": 0, "Ev Idx": 1757 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114561.103, "dur": 0.090 , "args": { "External id": 1759,"Record function id": 0, "Ev Idx": 1758 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114561.644, "dur": 0.641 , "args": { "External id": 1760,"Record function id": 0, "Ev Idx": 1759 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114561.814, "dur": 0.381 , "args": { "External id": 1761,"Record function id": 0, "Ev Idx": 1760 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114562.005, "dur": 0.090 , "args": { "External id": 1762,"Record function id": 0, "Ev Idx": 1761 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114563.407, "dur": 1.363 , "args": { "External id": 1763,"Record function id": 0, "Ev Idx": 1762 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114563.608, "dur": 1.062 , "args": { "External id": 1764,"Record function id": 0, "Ev Idx": 1763 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114573.326, "dur": 3.767 , "args": { "External id": 1765,"Record function id": 0, "Ev Idx": 1764 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114578.195, "dur": 0.601 , "args": { "External id": 1766,"Record function id": 0, "Ev Idx": 1765 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114580.128, "dur": 0.561 , "args": { "External id": 1767,"Record function id": 0, "Ev Idx": 1766 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381114584.557, "dur": 0.330 , "args": { "External id": 1768,"Record function id": 0, "Ev Idx": 1767 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381114584.737, "dur": 0.040 , "args": { "External id": 1769,"Record function id": 0, "Ev Idx": 1768 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381114587.402, "dur": 5.420 , "args": { "External id": 1770,"Record function id": 0, "Ev Idx": 1769 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114595.487, "dur": 10.369 , "args": { "External id": 1771,"Record function id": 0, "Ev Idx": 1770 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114596.068, "dur": 0.761 , "args": { "External id": 1772,"Record function id": 0, "Ev Idx": 1771 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114596.278, "dur": 0.461 , "args": { "External id": 1773,"Record function id": 0, "Ev Idx": 1772 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114596.519, "dur": 0.110 , "args": { "External id": 1774,"Record function id": 0, "Ev Idx": 1773 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114597.110, "dur": 0.621 , "args": { "External id": 1775,"Record function id": 0, "Ev Idx": 1774 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114597.260, "dur": 0.381 , "args": { "External id": 1776,"Record function id": 0, "Ev Idx": 1775 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114597.461, "dur": 0.080 , "args": { "External id": 1777,"Record function id": 0, "Ev Idx": 1776 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114597.992, "dur": 0.631 , "args": { "External id": 1778,"Record function id": 0, "Ev Idx": 1777 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114598.152, "dur": 0.381 , "args": { "External id": 1779,"Record function id": 0, "Ev Idx": 1778 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114598.352, "dur": 0.080 , "args": { "External id": 1780,"Record function id": 0, "Ev Idx": 1779 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114599.685, "dur": 1.302 , "args": { "External id": 1781,"Record function id": 0, "Ev Idx": 1780 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114599.905, "dur": 0.982 , "args": { "External id": 1782,"Record function id": 0, "Ev Idx": 1781 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381114610.555, "dur": 4.448 , "args": { "External id": 1783,"Record function id": 0, "Ev Idx": 1782 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381114620.534, "dur": 5.059 , "args": { "External id": 1784,"Record function id": 0, "Ev Idx": 1783 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381114630.502, "dur": 4.328 , "args": { "External id": 1785,"Record function id": 0, "Ev Idx": 1784 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381114637.175, "dur": 17.332 , "args": { "External id": 1786,"Record function id": 0, "Ev Idx": 1785 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114637.695, "dur": 0.521 , "args": { "External id": 1787,"Record function id": 0, "Ev Idx": 1786 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114637.926, "dur": 0.160 , "args": { "External id": 1788,"Record function id": 0, "Ev Idx": 1787 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114638.447, "dur": 0.391 , "args": { "External id": 1789,"Record function id": 0, "Ev Idx": 1788 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114638.647, "dur": 0.090 , "args": { "External id": 1790,"Record function id": 0, "Ev Idx": 1789 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114639.098, "dur": 0.381 , "args": { "External id": 1791,"Record function id": 0, "Ev Idx": 1790 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114639.298, "dur": 0.081 , "args": { "External id": 1792,"Record function id": 0, "Ev Idx": 1791 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381114639.799, "dur": 13.205 , "args": { "External id": 1793,"Record function id": 0, "Ev Idx": 1792 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381114640.721, "dur": 1.844 , "args": { "External id": 1794,"Record function id": 0, "Ev Idx": 1793 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114641.072, "dur": 1.392 , "args": { "External id": 1795,"Record function id": 0, "Ev Idx": 1794 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114643.025, "dur": 1.323 , "args": { "External id": 1796,"Record function id": 0, "Ev Idx": 1795 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114644.779, "dur": 1.262 , "args": { "External id": 1797,"Record function id": 0, "Ev Idx": 1796 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114646.322, "dur": 1.292 , "args": { "External id": 1798,"Record function id": 0, "Ev Idx": 1797 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114651.712, "dur": 0.941 , "args": { "External id": 1799,"Record function id": 0, "Ev Idx": 1798 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114653.305, "dur": 0.480 , "args": { "External id": 1800,"Record function id": 0, "Ev Idx": 1799 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114653.515, "dur": 0.160 , "args": { "External id": 1801,"Record function id": 0, "Ev Idx": 1800 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114661.440, "dur": 6.682 , "args": { "External id": 1802,"Record function id": 0, "Ev Idx": 1801 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114670.707, "dur": 10.760 , "args": { "External id": 1803,"Record function id": 0, "Ev Idx": 1802 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114671.318, "dur": 0.701 , "args": { "External id": 1804,"Record function id": 0, "Ev Idx": 1803 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114671.508, "dur": 0.421 , "args": { "External id": 1805,"Record function id": 0, "Ev Idx": 1804 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114671.719, "dur": 0.100 , "args": { "External id": 1806,"Record function id": 0, "Ev Idx": 1805 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114672.330, "dur": 0.621 , "args": { "External id": 1807,"Record function id": 0, "Ev Idx": 1806 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114672.480, "dur": 0.381 , "args": { "External id": 1808,"Record function id": 0, "Ev Idx": 1807 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114672.681, "dur": 0.080 , "args": { "External id": 1809,"Record function id": 0, "Ev Idx": 1808 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114673.202, "dur": 0.651 , "args": { "External id": 1810,"Record function id": 0, "Ev Idx": 1809 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114673.372, "dur": 0.381 , "args": { "External id": 1811,"Record function id": 0, "Ev Idx": 1810 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114673.582, "dur": 0.070 , "args": { "External id": 1812,"Record function id": 0, "Ev Idx": 1811 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114674.975, "dur": 1.292 , "args": { "External id": 1813,"Record function id": 0, "Ev Idx": 1812 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114675.185, "dur": 0.972 , "args": { "External id": 1814,"Record function id": 0, "Ev Idx": 1813 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381114685.484, "dur": 5.390 , "args": { "External id": 1815,"Record function id": 0, "Ev Idx": 1814 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114693.850, "dur": 5.280 , "args": { "External id": 1816,"Record function id": 0, "Ev Idx": 1815 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114702.426, "dur": 4.969 , "args": { "External id": 1817,"Record function id": 0, "Ev Idx": 1816 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114710.942, "dur": 10.459 , "args": { "External id": 1818,"Record function id": 0, "Ev Idx": 1817 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114711.463, "dur": 0.701 , "args": { "External id": 1819,"Record function id": 0, "Ev Idx": 1818 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114711.633, "dur": 0.441 , "args": { "External id": 1820,"Record function id": 0, "Ev Idx": 1819 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114711.854, "dur": 0.100 , "args": { "External id": 1821,"Record function id": 0, "Ev Idx": 1820 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114712.435, "dur": 0.681 , "args": { "External id": 1822,"Record function id": 0, "Ev Idx": 1821 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114712.605, "dur": 0.411 , "args": { "External id": 1823,"Record function id": 0, "Ev Idx": 1822 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114712.835, "dur": 0.080 , "args": { "External id": 1824,"Record function id": 0, "Ev Idx": 1823 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114713.366, "dur": 0.632 , "args": { "External id": 1825,"Record function id": 0, "Ev Idx": 1824 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114713.527, "dur": 0.380 , "args": { "External id": 1826,"Record function id": 0, "Ev Idx": 1825 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114713.727, "dur": 0.080 , "args": { "External id": 1827,"Record function id": 0, "Ev Idx": 1826 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114715.049, "dur": 1.293 , "args": { "External id": 1828,"Record function id": 0, "Ev Idx": 1827 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114715.260, "dur": 0.972 , "args": { "External id": 1829,"Record function id": 0, "Ev Idx": 1828 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114725.369, "dur": 10.139 , "args": { "External id": 1830,"Record function id": 0, "Ev Idx": 1829 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114725.870, "dur": 0.661 , "args": { "External id": 1831,"Record function id": 0, "Ev Idx": 1830 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114726.050, "dur": 0.381 , "args": { "External id": 1832,"Record function id": 0, "Ev Idx": 1831 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114726.260, "dur": 0.070 , "args": { "External id": 1833,"Record function id": 0, "Ev Idx": 1832 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114726.791, "dur": 0.611 , "args": { "External id": 1834,"Record function id": 0, "Ev Idx": 1833 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114726.952, "dur": 0.360 , "args": { "External id": 1835,"Record function id": 0, "Ev Idx": 1834 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114727.142, "dur": 0.070 , "args": { "External id": 1836,"Record function id": 0, "Ev Idx": 1835 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114727.663, "dur": 0.641 , "args": { "External id": 1837,"Record function id": 0, "Ev Idx": 1836 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114727.813, "dur": 0.401 , "args": { "External id": 1838,"Record function id": 0, "Ev Idx": 1837 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114728.034, "dur": 0.070 , "args": { "External id": 1839,"Record function id": 0, "Ev Idx": 1838 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114729.326, "dur": 1.222 , "args": { "External id": 1840,"Record function id": 0, "Ev Idx": 1839 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114729.486, "dur": 0.962 , "args": { "External id": 1841,"Record function id": 0, "Ev Idx": 1840 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381114740.106, "dur": 4.238 , "args": { "External id": 1842,"Record function id": 0, "Ev Idx": 1841 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114747.700, "dur": 5.270 , "args": { "External id": 1843,"Record function id": 0, "Ev Idx": 1842 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114756.156, "dur": 10.229 , "args": { "External id": 1844,"Record function id": 0, "Ev Idx": 1843 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114756.667, "dur": 0.711 , "args": { "External id": 1845,"Record function id": 0, "Ev Idx": 1844 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114756.877, "dur": 0.411 , "args": { "External id": 1846,"Record function id": 0, "Ev Idx": 1845 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114757.088, "dur": 0.090 , "args": { "External id": 1847,"Record function id": 0, "Ev Idx": 1846 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114757.659, "dur": 0.631 , "args": { "External id": 1848,"Record function id": 0, "Ev Idx": 1847 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114757.809, "dur": 0.381 , "args": { "External id": 1849,"Record function id": 0, "Ev Idx": 1848 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114758.019, "dur": 0.071 , "args": { "External id": 1850,"Record function id": 0, "Ev Idx": 1849 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114758.550, "dur": 0.632 , "args": { "External id": 1851,"Record function id": 0, "Ev Idx": 1850 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114758.711, "dur": 0.380 , "args": { "External id": 1852,"Record function id": 0, "Ev Idx": 1851 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114758.911, "dur": 0.080 , "args": { "External id": 1853,"Record function id": 0, "Ev Idx": 1852 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114760.214, "dur": 1.262 , "args": { "External id": 1854,"Record function id": 0, "Ev Idx": 1853 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114760.394, "dur": 0.982 , "args": { "External id": 1855,"Record function id": 0, "Ev Idx": 1854 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381114770.974, "dur": 4.849 , "args": { "External id": 1856,"Record function id": 0, "Ev Idx": 1855 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381114795.489, "dur": 23.394 , "args": { "External id": 1857,"Record function id": 0, "Ev Idx": 1856 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381114819.243, "dur": 340.164 , "args": { "External id": 1858,"Record function id": 0, "Ev Idx": 1857 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381114824.143, "dur": 10.800 , "args": { "External id": 1859,"Record function id": 0, "Ev Idx": 1858 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114827.689, "dur": 0.411 , "args": { "External id": 1860,"Record function id": 0, "Ev Idx": 1859 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381114828.591, "dur": 0.050 , "args": { "External id": 1861,"Record function id": 0, "Ev Idx": 1860 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381114829.032, "dur": 0.040 , "args": { "External id": 1862,"Record function id": 0, "Ev Idx": 1861 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114830.404, "dur": 0.271 , "args": { "External id": 1863,"Record function id": 0, "Ev Idx": 1862 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381114831.035, "dur": 0.030 , "args": { "External id": 1864,"Record function id": 0, "Ev Idx": 1863 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381114831.376, "dur": 0.030 , "args": { "External id": 1865,"Record function id": 0, "Ev Idx": 1864 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381114837.768, "dur": 1.082 , "args": { "External id": 1866,"Record function id": 0, "Ev Idx": 1865 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381114841.064, "dur": 315.197 , "args": { "External id": 1867,"Record function id": 0, "Ev Idx": 1866 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114849.810, "dur": 6.272 , "args": { "External id": 1868,"Record function id": 0, "Ev Idx": 1867 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114862.674, "dur": 7.093 , "args": { "External id": 1869,"Record function id": 0, "Ev Idx": 1868 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114875.598, "dur": 5.581 , "args": { "External id": 1870,"Record function id": 0, "Ev Idx": 1869 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381114886.559, "dur": 4.909 , "args": { "External id": 1871,"Record function id": 0, "Ev Idx": 1870 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114894.063, "dur": 12.283 , "args": { "External id": 1872,"Record function id": 0, "Ev Idx": 1871 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114894.674, "dur": 0.982 , "args": { "External id": 1873,"Record function id": 0, "Ev Idx": 1872 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114894.904, "dur": 0.651 , "args": { "External id": 1874,"Record function id": 0, "Ev Idx": 1873 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114895.195, "dur": 0.220 , "args": { "External id": 1875,"Record function id": 0, "Ev Idx": 1874 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114895.956, "dur": 0.671 , "args": { "External id": 1876,"Record function id": 0, "Ev Idx": 1875 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114896.137, "dur": 0.400 , "args": { "External id": 1877,"Record function id": 0, "Ev Idx": 1876 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114896.357, "dur": 0.070 , "args": { "External id": 1878,"Record function id": 0, "Ev Idx": 1877 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114896.888, "dur": 0.801 , "args": { "External id": 1879,"Record function id": 0, "Ev Idx": 1878 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114897.048, "dur": 0.551 , "args": { "External id": 1880,"Record function id": 0, "Ev Idx": 1879 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114897.239, "dur": 0.250 , "args": { "External id": 1881,"Record function id": 0, "Ev Idx": 1880 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114899.002, "dur": 1.433 , "args": { "External id": 1882,"Record function id": 0, "Ev Idx": 1881 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114899.232, "dur": 1.082 , "args": { "External id": 1883,"Record function id": 0, "Ev Idx": 1882 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114911.655, "dur": 4.659 , "args": { "External id": 1884,"Record function id": 0, "Ev Idx": 1883 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114918.558, "dur": 0.932 , "args": { "External id": 1885,"Record function id": 0, "Ev Idx": 1884 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114921.223, "dur": 0.551 , "args": { "External id": 1886,"Record function id": 0, "Ev Idx": 1885 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381114925.571, "dur": 0.401 , "args": { "External id": 1887,"Record function id": 0, "Ev Idx": 1886 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381114925.812, "dur": 0.040 , "args": { "External id": 1888,"Record function id": 0, "Ev Idx": 1887 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381114929.278, "dur": 6.001 , "args": { "External id": 1889,"Record function id": 0, "Ev Idx": 1888 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114937.484, "dur": 10.499 , "args": { "External id": 1890,"Record function id": 0, "Ev Idx": 1889 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114938.035, "dur": 0.741 , "args": { "External id": 1891,"Record function id": 0, "Ev Idx": 1890 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114938.215, "dur": 0.471 , "args": { "External id": 1892,"Record function id": 0, "Ev Idx": 1891 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114938.455, "dur": 0.101 , "args": { "External id": 1893,"Record function id": 0, "Ev Idx": 1892 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114939.046, "dur": 0.632 , "args": { "External id": 1894,"Record function id": 0, "Ev Idx": 1893 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114939.207, "dur": 0.380 , "args": { "External id": 1895,"Record function id": 0, "Ev Idx": 1894 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114939.407, "dur": 0.070 , "args": { "External id": 1896,"Record function id": 0, "Ev Idx": 1895 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114939.928, "dur": 0.621 , "args": { "External id": 1897,"Record function id": 0, "Ev Idx": 1896 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114940.088, "dur": 0.371 , "args": { "External id": 1898,"Record function id": 0, "Ev Idx": 1897 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114940.279, "dur": 0.080 , "args": { "External id": 1899,"Record function id": 0, "Ev Idx": 1898 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114941.651, "dur": 1.293 , "args": { "External id": 1900,"Record function id": 0, "Ev Idx": 1899 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114941.832, "dur": 1.001 , "args": { "External id": 1901,"Record function id": 0, "Ev Idx": 1900 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381114951.229, "dur": 3.827 , "args": { "External id": 1902,"Record function id": 0, "Ev Idx": 1901 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114956.118, "dur": 0.661 , "args": { "External id": 1903,"Record function id": 0, "Ev Idx": 1902 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381114958.162, "dur": 0.581 , "args": { "External id": 1904,"Record function id": 0, "Ev Idx": 1903 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381114962.741, "dur": 0.320 , "args": { "External id": 1905,"Record function id": 0, "Ev Idx": 1904 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381114962.921, "dur": 0.040 , "args": { "External id": 1906,"Record function id": 0, "Ev Idx": 1905 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381114965.576, "dur": 5.670 , "args": { "External id": 1907,"Record function id": 0, "Ev Idx": 1906 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381114973.470, "dur": 10.570 , "args": { "External id": 1908,"Record function id": 0, "Ev Idx": 1907 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114974.032, "dur": 0.711 , "args": { "External id": 1909,"Record function id": 0, "Ev Idx": 1908 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114974.222, "dur": 0.421 , "args": { "External id": 1910,"Record function id": 0, "Ev Idx": 1909 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114974.452, "dur": 0.090 , "args": { "External id": 1911,"Record function id": 0, "Ev Idx": 1910 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114975.013, "dur": 0.642 , "args": { "External id": 1912,"Record function id": 0, "Ev Idx": 1911 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114975.184, "dur": 0.370 , "args": { "External id": 1913,"Record function id": 0, "Ev Idx": 1912 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114975.384, "dur": 0.070 , "args": { "External id": 1914,"Record function id": 0, "Ev Idx": 1913 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381114975.905, "dur": 0.631 , "args": { "External id": 1915,"Record function id": 0, "Ev Idx": 1914 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381114976.065, "dur": 0.381 , "args": { "External id": 1916,"Record function id": 0, "Ev Idx": 1915 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381114976.256, "dur": 0.090 , "args": { "External id": 1917,"Record function id": 0, "Ev Idx": 1916 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114977.598, "dur": 1.283 , "args": { "External id": 1918,"Record function id": 0, "Ev Idx": 1917 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381114977.779, "dur": 0.991 , "args": { "External id": 1919,"Record function id": 0, "Ev Idx": 1918 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381114988.529, "dur": 4.227 , "args": { "External id": 1920,"Record function id": 0, "Ev Idx": 1919 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381114998.297, "dur": 4.909 , "args": { "External id": 1921,"Record function id": 0, "Ev Idx": 1920 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381115008.135, "dur": 4.729 , "args": { "External id": 1922,"Record function id": 0, "Ev Idx": 1921 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381115015.348, "dur": 17.513 , "args": { "External id": 1923,"Record function id": 0, "Ev Idx": 1922 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115015.839, "dur": 0.511 , "args": { "External id": 1924,"Record function id": 0, "Ev Idx": 1923 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115016.070, "dur": 0.160 , "args": { "External id": 1925,"Record function id": 0, "Ev Idx": 1924 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115016.581, "dur": 0.400 , "args": { "External id": 1926,"Record function id": 0, "Ev Idx": 1925 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115016.791, "dur": 0.090 , "args": { "External id": 1927,"Record function id": 0, "Ev Idx": 1926 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115017.232, "dur": 0.381 , "args": { "External id": 1928,"Record function id": 0, "Ev Idx": 1927 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115017.422, "dur": 0.090 , "args": { "External id": 1929,"Record function id": 0, "Ev Idx": 1928 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381115017.953, "dur": 13.315 , "args": { "External id": 1930,"Record function id": 0, "Ev Idx": 1929 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381115018.865, "dur": 1.884 , "args": { "External id": 1931,"Record function id": 0, "Ev Idx": 1930 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115019.156, "dur": 1.462 , "args": { "External id": 1932,"Record function id": 0, "Ev Idx": 1931 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115021.229, "dur": 1.243 , "args": { "External id": 1933,"Record function id": 0, "Ev Idx": 1932 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115022.943, "dur": 1.232 , "args": { "External id": 1934,"Record function id": 0, "Ev Idx": 1933 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115024.455, "dur": 1.283 , "args": { "External id": 1935,"Record function id": 0, "Ev Idx": 1934 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115029.996, "dur": 0.931 , "args": { "External id": 1936,"Record function id": 0, "Ev Idx": 1935 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115031.599, "dur": 0.551 , "args": { "External id": 1937,"Record function id": 0, "Ev Idx": 1936 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115031.819, "dur": 0.231 , "args": { "External id": 1938,"Record function id": 0, "Ev Idx": 1937 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115040.255, "dur": 6.532 , "args": { "External id": 1939,"Record function id": 0, "Ev Idx": 1938 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115049.201, "dur": 11.061 , "args": { "External id": 1940,"Record function id": 0, "Ev Idx": 1939 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115049.833, "dur": 0.691 , "args": { "External id": 1941,"Record function id": 0, "Ev Idx": 1940 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115050.013, "dur": 0.411 , "args": { "External id": 1942,"Record function id": 0, "Ev Idx": 1941 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115050.233, "dur": 0.081 , "args": { "External id": 1943,"Record function id": 0, "Ev Idx": 1942 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115050.814, "dur": 0.682 , "args": { "External id": 1944,"Record function id": 0, "Ev Idx": 1943 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115050.975, "dur": 0.431 , "args": { "External id": 1945,"Record function id": 0, "Ev Idx": 1944 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115051.225, "dur": 0.080 , "args": { "External id": 1946,"Record function id": 0, "Ev Idx": 1945 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115051.756, "dur": 0.651 , "args": { "External id": 1947,"Record function id": 0, "Ev Idx": 1946 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115051.907, "dur": 0.400 , "args": { "External id": 1948,"Record function id": 0, "Ev Idx": 1947 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115052.117, "dur": 0.090 , "args": { "External id": 1949,"Record function id": 0, "Ev Idx": 1948 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115053.530, "dur": 1.272 , "args": { "External id": 1950,"Record function id": 0, "Ev Idx": 1949 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115053.740, "dur": 0.952 , "args": { "External id": 1951,"Record function id": 0, "Ev Idx": 1950 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381115064.249, "dur": 5.451 , "args": { "External id": 1952,"Record function id": 0, "Ev Idx": 1951 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115072.745, "dur": 5.040 , "args": { "External id": 1953,"Record function id": 0, "Ev Idx": 1952 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115081.091, "dur": 4.989 , "args": { "External id": 1954,"Record function id": 0, "Ev Idx": 1953 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115089.737, "dur": 10.499 , "args": { "External id": 1955,"Record function id": 0, "Ev Idx": 1954 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115090.278, "dur": 0.701 , "args": { "External id": 1956,"Record function id": 0, "Ev Idx": 1955 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115090.458, "dur": 0.421 , "args": { "External id": 1957,"Record function id": 0, "Ev Idx": 1956 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115090.679, "dur": 0.090 , "args": { "External id": 1958,"Record function id": 0, "Ev Idx": 1957 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115091.250, "dur": 0.701 , "args": { "External id": 1959,"Record function id": 0, "Ev Idx": 1958 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115091.430, "dur": 0.431 , "args": { "External id": 1960,"Record function id": 0, "Ev Idx": 1959 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115091.681, "dur": 0.070 , "args": { "External id": 1961,"Record function id": 0, "Ev Idx": 1960 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115092.202, "dur": 0.671 , "args": { "External id": 1962,"Record function id": 0, "Ev Idx": 1961 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115092.362, "dur": 0.421 , "args": { "External id": 1963,"Record function id": 0, "Ev Idx": 1962 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115092.592, "dur": 0.080 , "args": { "External id": 1964,"Record function id": 0, "Ev Idx": 1963 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115093.965, "dur": 1.282 , "args": { "External id": 1965,"Record function id": 0, "Ev Idx": 1964 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115094.155, "dur": 0.992 , "args": { "External id": 1966,"Record function id": 0, "Ev Idx": 1965 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115103.923, "dur": 10.109 , "args": { "External id": 1967,"Record function id": 0, "Ev Idx": 1966 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115104.424, "dur": 0.651 , "args": { "External id": 1968,"Record function id": 0, "Ev Idx": 1967 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115104.595, "dur": 0.390 , "args": { "External id": 1969,"Record function id": 0, "Ev Idx": 1968 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115104.785, "dur": 0.100 , "args": { "External id": 1970,"Record function id": 0, "Ev Idx": 1969 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115105.336, "dur": 0.621 , "args": { "External id": 1971,"Record function id": 0, "Ev Idx": 1970 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115105.496, "dur": 0.371 , "args": { "External id": 1972,"Record function id": 0, "Ev Idx": 1971 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115105.697, "dur": 0.070 , "args": { "External id": 1973,"Record function id": 0, "Ev Idx": 1972 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115106.218, "dur": 0.651 , "args": { "External id": 1974,"Record function id": 0, "Ev Idx": 1973 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115106.388, "dur": 0.391 , "args": { "External id": 1975,"Record function id": 0, "Ev Idx": 1974 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115106.598, "dur": 0.070 , "args": { "External id": 1976,"Record function id": 0, "Ev Idx": 1975 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115107.891, "dur": 1.242 , "args": { "External id": 1977,"Record function id": 0, "Ev Idx": 1976 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115108.081, "dur": 0.952 , "args": { "External id": 1978,"Record function id": 0, "Ev Idx": 1977 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381115118.531, "dur": 4.077 , "args": { "External id": 1979,"Record function id": 0, "Ev Idx": 1978 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115125.934, "dur": 4.819 , "args": { "External id": 1980,"Record function id": 0, "Ev Idx": 1979 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115134.019, "dur": 11.291 , "args": { "External id": 1981,"Record function id": 0, "Ev Idx": 1980 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115134.540, "dur": 0.672 , "args": { "External id": 1982,"Record function id": 0, "Ev Idx": 1981 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115134.721, "dur": 0.400 , "args": { "External id": 1983,"Record function id": 0, "Ev Idx": 1982 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115134.941, "dur": 0.070 , "args": { "External id": 1984,"Record function id": 0, "Ev Idx": 1983 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115135.482, "dur": 0.651 , "args": { "External id": 1985,"Record function id": 0, "Ev Idx": 1984 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115135.632, "dur": 0.411 , "args": { "External id": 1986,"Record function id": 0, "Ev Idx": 1985 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115135.863, "dur": 0.070 , "args": { "External id": 1987,"Record function id": 0, "Ev Idx": 1986 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115136.394, "dur": 0.631 , "args": { "External id": 1988,"Record function id": 0, "Ev Idx": 1987 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115136.554, "dur": 0.381 , "args": { "External id": 1989,"Record function id": 0, "Ev Idx": 1988 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115136.754, "dur": 0.071 , "args": { "External id": 1990,"Record function id": 0, "Ev Idx": 1989 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115138.077, "dur": 1.342 , "args": { "External id": 1991,"Record function id": 0, "Ev Idx": 1990 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115138.257, "dur": 1.052 , "args": { "External id": 1992,"Record function id": 0, "Ev Idx": 1991 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381115149.899, "dur": 4.749 , "args": { "External id": 1993,"Record function id": 0, "Ev Idx": 1992 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381115173.984, "dur": 24.145 , "args": { "External id": 1994,"Record function id": 0, "Ev Idx": 1993 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381115198.479, "dur": 343.470 , "args": { "External id": 1995,"Record function id": 0, "Ev Idx": 1994 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381115203.409, "dur": 10.319 , "args": { "External id": 1996,"Record function id": 0, "Ev Idx": 1995 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115206.695, "dur": 0.350 , "args": { "External id": 1997,"Record function id": 0, "Ev Idx": 1996 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381115207.536, "dur": 0.050 , "args": { "External id": 1998,"Record function id": 0, "Ev Idx": 1997 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381115207.967, "dur": 0.040 , "args": { "External id": 1999,"Record function id": 0, "Ev Idx": 1998 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115209.149, "dur": 0.271 , "args": { "External id": 2000,"Record function id": 0, "Ev Idx": 1999 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381115209.770, "dur": 0.020 , "args": { "External id": 2001,"Record function id": 0, "Ev Idx": 2000 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381115210.131, "dur": 0.020 , "args": { "External id": 2002,"Record function id": 0, "Ev Idx": 2001 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381115216.573, "dur": 1.092 , "args": { "External id": 2003,"Record function id": 0, "Ev Idx": 2002 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381115219.799, "dur": 316.579 , "args": { "External id": 2004,"Record function id": 0, "Ev Idx": 2003 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115228.215, "dur": 6.201 , "args": { "External id": 2005,"Record function id": 0, "Ev Idx": 2004 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115241.209, "dur": 7.193 , "args": { "External id": 2006,"Record function id": 0, "Ev Idx": 2005 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115254.474, "dur": 5.580 , "args": { "External id": 2007,"Record function id": 0, "Ev Idx": 2006 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115265.755, "dur": 5.079 , "args": { "External id": 2008,"Record function id": 0, "Ev Idx": 2007 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115273.289, "dur": 12.343 , "args": { "External id": 2009,"Record function id": 0, "Ev Idx": 2008 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115273.900, "dur": 0.951 , "args": { "External id": 2010,"Record function id": 0, "Ev Idx": 2009 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115274.130, "dur": 0.611 , "args": { "External id": 2011,"Record function id": 0, "Ev Idx": 2010 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115274.461, "dur": 0.140 , "args": { "External id": 2012,"Record function id": 0, "Ev Idx": 2011 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115275.162, "dur": 0.661 , "args": { "External id": 2013,"Record function id": 0, "Ev Idx": 2012 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115275.332, "dur": 0.401 , "args": { "External id": 2014,"Record function id": 0, "Ev Idx": 2013 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115275.543, "dur": 0.070 , "args": { "External id": 2015,"Record function id": 0, "Ev Idx": 2014 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115276.074, "dur": 0.651 , "args": { "External id": 2016,"Record function id": 0, "Ev Idx": 2015 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115276.244, "dur": 0.391 , "args": { "External id": 2017,"Record function id": 0, "Ev Idx": 2016 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115276.444, "dur": 0.081 , "args": { "External id": 2018,"Record function id": 0, "Ev Idx": 2017 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115278.057, "dur": 1.503 , "args": { "External id": 2019,"Record function id": 0, "Ev Idx": 2018 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115278.328, "dur": 1.102 , "args": { "External id": 2020,"Record function id": 0, "Ev Idx": 2019 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115291.092, "dur": 4.518 , "args": { "External id": 2021,"Record function id": 0, "Ev Idx": 2020 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115297.824, "dur": 0.862 , "args": { "External id": 2022,"Record function id": 0, "Ev Idx": 2021 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115300.559, "dur": 0.591 , "args": { "External id": 2023,"Record function id": 0, "Ev Idx": 2022 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381115304.817, "dur": 0.411 , "args": { "External id": 2024,"Record function id": 0, "Ev Idx": 2023 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381115305.058, "dur": 0.050 , "args": { "External id": 2025,"Record function id": 0, "Ev Idx": 2024 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381115308.184, "dur": 6.251 , "args": { "External id": 2026,"Record function id": 0, "Ev Idx": 2025 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115316.649, "dur": 10.760 , "args": { "External id": 2027,"Record function id": 0, "Ev Idx": 2026 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115317.200, "dur": 0.742 , "args": { "External id": 2028,"Record function id": 0, "Ev Idx": 2027 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115317.391, "dur": 0.451 , "args": { "External id": 2029,"Record function id": 0, "Ev Idx": 2028 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115317.641, "dur": 0.090 , "args": { "External id": 2030,"Record function id": 0, "Ev Idx": 2029 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115318.212, "dur": 0.611 , "args": { "External id": 2031,"Record function id": 0, "Ev Idx": 2030 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115318.373, "dur": 0.360 , "args": { "External id": 2032,"Record function id": 0, "Ev Idx": 2031 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115318.553, "dur": 0.070 , "args": { "External id": 2033,"Record function id": 0, "Ev Idx": 2032 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115319.084, "dur": 0.631 , "args": { "External id": 2034,"Record function id": 0, "Ev Idx": 2033 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115319.244, "dur": 0.381 , "args": { "External id": 2035,"Record function id": 0, "Ev Idx": 2034 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115319.434, "dur": 0.091 , "args": { "External id": 2036,"Record function id": 0, "Ev Idx": 2035 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115320.827, "dur": 1.353 , "args": { "External id": 2037,"Record function id": 0, "Ev Idx": 2036 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115321.037, "dur": 1.042 , "args": { "External id": 2038,"Record function id": 0, "Ev Idx": 2037 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115330.575, "dur": 3.737 , "args": { "External id": 2039,"Record function id": 0, "Ev Idx": 2038 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115335.394, "dur": 0.631 , "args": { "External id": 2040,"Record function id": 0, "Ev Idx": 2039 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115337.398, "dur": 0.571 , "args": { "External id": 2041,"Record function id": 0, "Ev Idx": 2040 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381115341.706, "dur": 0.341 , "args": { "External id": 2042,"Record function id": 0, "Ev Idx": 2041 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381115341.896, "dur": 0.030 , "args": { "External id": 2043,"Record function id": 0, "Ev Idx": 2042 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381115344.601, "dur": 5.581 , "args": { "External id": 2044,"Record function id": 0, "Ev Idx": 2043 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115352.426, "dur": 13.645 , "args": { "External id": 2045,"Record function id": 0, "Ev Idx": 2044 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115352.997, "dur": 0.721 , "args": { "External id": 2046,"Record function id": 0, "Ev Idx": 2045 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115353.187, "dur": 0.441 , "args": { "External id": 2047,"Record function id": 0, "Ev Idx": 2046 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115353.438, "dur": 0.090 , "args": { "External id": 2048,"Record function id": 0, "Ev Idx": 2047 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115357.075, "dur": 0.751 , "args": { "External id": 2049,"Record function id": 0, "Ev Idx": 2048 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115357.255, "dur": 0.481 , "args": { "External id": 2050,"Record function id": 0, "Ev Idx": 2049 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115357.545, "dur": 0.081 , "args": { "External id": 2051,"Record function id": 0, "Ev Idx": 2050 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115358.096, "dur": 0.622 , "args": { "External id": 2052,"Record function id": 0, "Ev Idx": 2051 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115358.257, "dur": 0.370 , "args": { "External id": 2053,"Record function id": 0, "Ev Idx": 2052 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115358.447, "dur": 0.080 , "args": { "External id": 2054,"Record function id": 0, "Ev Idx": 2053 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115359.800, "dur": 1.302 , "args": { "External id": 2055,"Record function id": 0, "Ev Idx": 2054 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115360.000, "dur": 0.992 , "args": { "External id": 2056,"Record function id": 0, "Ev Idx": 2055 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381115370.540, "dur": 4.177 , "args": { "External id": 2057,"Record function id": 0, "Ev Idx": 2056 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381115380.188, "dur": 4.708 , "args": { "External id": 2058,"Record function id": 0, "Ev Idx": 2057 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381115389.886, "dur": 4.388 , "args": { "External id": 2059,"Record function id": 0, "Ev Idx": 2058 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381115396.578, "dur": 17.172 , "args": { "External id": 2060,"Record function id": 0, "Ev Idx": 2059 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115397.109, "dur": 0.501 , "args": { "External id": 2061,"Record function id": 0, "Ev Idx": 2060 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115397.350, "dur": 0.140 , "args": { "External id": 2062,"Record function id": 0, "Ev Idx": 2061 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115397.840, "dur": 0.401 , "args": { "External id": 2063,"Record function id": 0, "Ev Idx": 2062 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115398.051, "dur": 0.080 , "args": { "External id": 2064,"Record function id": 0, "Ev Idx": 2063 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115398.492, "dur": 0.380 , "args": { "External id": 2065,"Record function id": 0, "Ev Idx": 2064 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115398.682, "dur": 0.090 , "args": { "External id": 2066,"Record function id": 0, "Ev Idx": 2065 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381115399.203, "dur": 13.054 , "args": { "External id": 2067,"Record function id": 0, "Ev Idx": 2066 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381115400.085, "dur": 1.833 , "args": { "External id": 2068,"Record function id": 0, "Ev Idx": 2067 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115400.375, "dur": 1.423 , "args": { "External id": 2069,"Record function id": 0, "Ev Idx": 2068 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115402.349, "dur": 1.252 , "args": { "External id": 2070,"Record function id": 0, "Ev Idx": 2069 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115404.032, "dur": 1.232 , "args": { "External id": 2071,"Record function id": 0, "Ev Idx": 2070 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115405.545, "dur": 1.262 , "args": { "External id": 2072,"Record function id": 0, "Ev Idx": 2071 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115410.925, "dur": 0.972 , "args": { "External id": 2073,"Record function id": 0, "Ev Idx": 2072 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115412.528, "dur": 0.471 , "args": { "External id": 2074,"Record function id": 0, "Ev Idx": 2073 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115412.738, "dur": 0.150 , "args": { "External id": 2075,"Record function id": 0, "Ev Idx": 2074 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115420.953, "dur": 6.523 , "args": { "External id": 2076,"Record function id": 0, "Ev Idx": 2075 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115429.920, "dur": 10.670 , "args": { "External id": 2077,"Record function id": 0, "Ev Idx": 2076 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115430.521, "dur": 0.762 , "args": { "External id": 2078,"Record function id": 0, "Ev Idx": 2077 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115430.712, "dur": 0.481 , "args": { "External id": 2079,"Record function id": 0, "Ev Idx": 2078 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115430.932, "dur": 0.140 , "args": { "External id": 2080,"Record function id": 0, "Ev Idx": 2079 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115431.563, "dur": 0.661 , "args": { "External id": 2081,"Record function id": 0, "Ev Idx": 2080 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115431.724, "dur": 0.410 , "args": { "External id": 2082,"Record function id": 0, "Ev Idx": 2081 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115431.944, "dur": 0.080 , "args": { "External id": 2083,"Record function id": 0, "Ev Idx": 2082 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115432.475, "dur": 0.631 , "args": { "External id": 2084,"Record function id": 0, "Ev Idx": 2083 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115432.645, "dur": 0.371 , "args": { "External id": 2085,"Record function id": 0, "Ev Idx": 2084 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115432.836, "dur": 0.080 , "args": { "External id": 2086,"Record function id": 0, "Ev Idx": 2085 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115434.188, "dur": 1.252 , "args": { "External id": 2087,"Record function id": 0, "Ev Idx": 2086 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115434.368, "dur": 0.972 , "args": { "External id": 2088,"Record function id": 0, "Ev Idx": 2087 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381115444.628, "dur": 5.189 , "args": { "External id": 2089,"Record function id": 0, "Ev Idx": 2088 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115452.783, "dur": 5.199 , "args": { "External id": 2090,"Record function id": 0, "Ev Idx": 2089 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115461.399, "dur": 5.179 , "args": { "External id": 2091,"Record function id": 0, "Ev Idx": 2090 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115470.095, "dur": 10.229 , "args": { "External id": 2092,"Record function id": 0, "Ev Idx": 2091 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115470.606, "dur": 0.681 , "args": { "External id": 2093,"Record function id": 0, "Ev Idx": 2092 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115470.776, "dur": 0.421 , "args": { "External id": 2094,"Record function id": 0, "Ev Idx": 2093 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115470.997, "dur": 0.090 , "args": { "External id": 2095,"Record function id": 0, "Ev Idx": 2094 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115471.558, "dur": 0.631 , "args": { "External id": 2096,"Record function id": 0, "Ev Idx": 2095 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115471.718, "dur": 0.381 , "args": { "External id": 2097,"Record function id": 0, "Ev Idx": 2096 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115471.928, "dur": 0.070 , "args": { "External id": 2098,"Record function id": 0, "Ev Idx": 2097 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115472.439, "dur": 0.621 , "args": { "External id": 2099,"Record function id": 0, "Ev Idx": 2098 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115472.600, "dur": 0.370 , "args": { "External id": 2100,"Record function id": 0, "Ev Idx": 2099 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115472.790, "dur": 0.080 , "args": { "External id": 2101,"Record function id": 0, "Ev Idx": 2100 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115474.132, "dur": 1.253 , "args": { "External id": 2102,"Record function id": 0, "Ev Idx": 2101 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115474.313, "dur": 0.972 , "args": { "External id": 2103,"Record function id": 0, "Ev Idx": 2102 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115484.131, "dur": 9.818 , "args": { "External id": 2104,"Record function id": 0, "Ev Idx": 2103 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115484.632, "dur": 0.631 , "args": { "External id": 2105,"Record function id": 0, "Ev Idx": 2104 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115484.792, "dur": 0.381 , "args": { "External id": 2106,"Record function id": 0, "Ev Idx": 2105 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115484.993, "dur": 0.080 , "args": { "External id": 2107,"Record function id": 0, "Ev Idx": 2106 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115485.534, "dur": 0.611 , "args": { "External id": 2108,"Record function id": 0, "Ev Idx": 2107 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115485.694, "dur": 0.361 , "args": { "External id": 2109,"Record function id": 0, "Ev Idx": 2108 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115485.884, "dur": 0.070 , "args": { "External id": 2110,"Record function id": 0, "Ev Idx": 2109 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115486.405, "dur": 0.621 , "args": { "External id": 2111,"Record function id": 0, "Ev Idx": 2110 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115486.566, "dur": 0.370 , "args": { "External id": 2112,"Record function id": 0, "Ev Idx": 2111 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115486.756, "dur": 0.080 , "args": { "External id": 2113,"Record function id": 0, "Ev Idx": 2112 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115488.018, "dur": 1.243 , "args": { "External id": 2114,"Record function id": 0, "Ev Idx": 2113 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115488.189, "dur": 0.961 , "args": { "External id": 2115,"Record function id": 0, "Ev Idx": 2114 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381115498.608, "dur": 4.058 , "args": { "External id": 2116,"Record function id": 0, "Ev Idx": 2115 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115506.062, "dur": 5.250 , "args": { "External id": 2117,"Record function id": 0, "Ev Idx": 2116 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115514.568, "dur": 10.078 , "args": { "External id": 2118,"Record function id": 0, "Ev Idx": 2117 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115515.079, "dur": 0.651 , "args": { "External id": 2119,"Record function id": 0, "Ev Idx": 2118 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115515.249, "dur": 0.391 , "args": { "External id": 2120,"Record function id": 0, "Ev Idx": 2119 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115515.459, "dur": 0.081 , "args": { "External id": 2121,"Record function id": 0, "Ev Idx": 2120 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115516.000, "dur": 0.662 , "args": { "External id": 2122,"Record function id": 0, "Ev Idx": 2121 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115516.161, "dur": 0.410 , "args": { "External id": 2123,"Record function id": 0, "Ev Idx": 2122 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115516.401, "dur": 0.070 , "args": { "External id": 2124,"Record function id": 0, "Ev Idx": 2123 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115516.922, "dur": 0.611 , "args": { "External id": 2125,"Record function id": 0, "Ev Idx": 2124 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115517.072, "dur": 0.371 , "args": { "External id": 2126,"Record function id": 0, "Ev Idx": 2125 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115517.263, "dur": 0.080 , "args": { "External id": 2127,"Record function id": 0, "Ev Idx": 2126 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115518.545, "dur": 1.242 , "args": { "External id": 2128,"Record function id": 0, "Ev Idx": 2127 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115518.725, "dur": 0.952 , "args": { "External id": 2129,"Record function id": 0, "Ev Idx": 2128 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381115530.026, "dur": 4.749 , "args": { "External id": 2130,"Record function id": 0, "Ev Idx": 2129 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381115557.157, "dur": 23.965 , "args": { "External id": 2131,"Record function id": 0, "Ev Idx": 2130 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381115581.532, "dur": 336.897 , "args": { "External id": 2132,"Record function id": 0, "Ev Idx": 2131 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381115586.472, "dur": 10.429 , "args": { "External id": 2133,"Record function id": 0, "Ev Idx": 2132 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115589.778, "dur": 0.451 , "args": { "External id": 2134,"Record function id": 0, "Ev Idx": 2133 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381115590.729, "dur": 0.051 , "args": { "External id": 2135,"Record function id": 0, "Ev Idx": 2134 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381115591.170, "dur": 0.040 , "args": { "External id": 2136,"Record function id": 0, "Ev Idx": 2135 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115592.393, "dur": 0.240 , "args": { "External id": 2137,"Record function id": 0, "Ev Idx": 2136 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381115592.974, "dur": 0.040 , "args": { "External id": 2138,"Record function id": 0, "Ev Idx": 2137 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381115593.314, "dur": 0.030 , "args": { "External id": 2139,"Record function id": 0, "Ev Idx": 2138 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381115599.666, "dur": 1.082 , "args": { "External id": 2140,"Record function id": 0, "Ev Idx": 2139 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381115602.962, "dur": 312.562 , "args": { "External id": 2141,"Record function id": 0, "Ev Idx": 2140 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115611.919, "dur": 6.572 , "args": { "External id": 2142,"Record function id": 0, "Ev Idx": 2141 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115625.184, "dur": 7.133 , "args": { "External id": 2143,"Record function id": 0, "Ev Idx": 2142 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115638.208, "dur": 5.410 , "args": { "External id": 2144,"Record function id": 0, "Ev Idx": 2143 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115649.048, "dur": 4.969 , "args": { "External id": 2145,"Record function id": 0, "Ev Idx": 2144 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115656.582, "dur": 12.393 , "args": { "External id": 2146,"Record function id": 0, "Ev Idx": 2145 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115657.193, "dur": 1.052 , "args": { "External id": 2147,"Record function id": 0, "Ev Idx": 2146 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115657.404, "dur": 0.741 , "args": { "External id": 2148,"Record function id": 0, "Ev Idx": 2147 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115657.714, "dur": 0.261 , "args": { "External id": 2149,"Record function id": 0, "Ev Idx": 2148 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115658.536, "dur": 0.641 , "args": { "External id": 2150,"Record function id": 0, "Ev Idx": 2149 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115658.706, "dur": 0.381 , "args": { "External id": 2151,"Record function id": 0, "Ev Idx": 2150 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115658.906, "dur": 0.071 , "args": { "External id": 2152,"Record function id": 0, "Ev Idx": 2151 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115659.447, "dur": 0.642 , "args": { "External id": 2153,"Record function id": 0, "Ev Idx": 2152 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115659.618, "dur": 0.380 , "args": { "External id": 2154,"Record function id": 0, "Ev Idx": 2153 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115659.808, "dur": 0.090 , "args": { "External id": 2155,"Record function id": 0, "Ev Idx": 2154 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115661.441, "dur": 1.453 , "args": { "External id": 2156,"Record function id": 0, "Ev Idx": 2155 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115661.692, "dur": 1.082 , "args": { "External id": 2157,"Record function id": 0, "Ev Idx": 2156 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115674.375, "dur": 4.509 , "args": { "External id": 2158,"Record function id": 0, "Ev Idx": 2157 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115681.198, "dur": 0.892 , "args": { "External id": 2159,"Record function id": 0, "Ev Idx": 2158 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115683.853, "dur": 0.591 , "args": { "External id": 2160,"Record function id": 0, "Ev Idx": 2159 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381115688.311, "dur": 0.471 , "args": { "External id": 2161,"Record function id": 0, "Ev Idx": 2160 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381115688.592, "dur": 0.070 , "args": { "External id": 2162,"Record function id": 0, "Ev Idx": 2161 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381115691.547, "dur": 5.951 , "args": { "External id": 2163,"Record function id": 0, "Ev Idx": 2162 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115699.722, "dur": 10.730 , "args": { "External id": 2164,"Record function id": 0, "Ev Idx": 2163 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115700.273, "dur": 0.732 , "args": { "External id": 2165,"Record function id": 0, "Ev Idx": 2164 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115700.444, "dur": 0.471 , "args": { "External id": 2166,"Record function id": 0, "Ev Idx": 2165 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115700.684, "dur": 0.110 , "args": { "External id": 2167,"Record function id": 0, "Ev Idx": 2166 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115701.285, "dur": 0.631 , "args": { "External id": 2168,"Record function id": 0, "Ev Idx": 2167 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115701.446, "dur": 0.380 , "args": { "External id": 2169,"Record function id": 0, "Ev Idx": 2168 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115701.646, "dur": 0.070 , "args": { "External id": 2170,"Record function id": 0, "Ev Idx": 2169 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115702.167, "dur": 0.661 , "args": { "External id": 2171,"Record function id": 0, "Ev Idx": 2170 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115702.327, "dur": 0.411 , "args": { "External id": 2172,"Record function id": 0, "Ev Idx": 2171 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115702.528, "dur": 0.100 , "args": { "External id": 2173,"Record function id": 0, "Ev Idx": 2172 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115703.930, "dur": 1.343 , "args": { "External id": 2174,"Record function id": 0, "Ev Idx": 2173 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115704.131, "dur": 1.031 , "args": { "External id": 2175,"Record function id": 0, "Ev Idx": 2174 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115713.688, "dur": 3.807 , "args": { "External id": 2176,"Record function id": 0, "Ev Idx": 2175 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115718.537, "dur": 0.612 , "args": { "External id": 2177,"Record function id": 0, "Ev Idx": 2176 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381115720.421, "dur": 0.571 , "args": { "External id": 2178,"Record function id": 0, "Ev Idx": 2177 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381115724.659, "dur": 0.330 , "args": { "External id": 2179,"Record function id": 0, "Ev Idx": 2178 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381115724.829, "dur": 0.050 , "args": { "External id": 2180,"Record function id": 0, "Ev Idx": 2179 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381115727.444, "dur": 5.550 , "args": { "External id": 2181,"Record function id": 0, "Ev Idx": 2180 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115735.289, "dur": 10.479 , "args": { "External id": 2182,"Record function id": 0, "Ev Idx": 2181 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115735.820, "dur": 0.801 , "args": { "External id": 2183,"Record function id": 0, "Ev Idx": 2182 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115736.010, "dur": 0.511 , "args": { "External id": 2184,"Record function id": 0, "Ev Idx": 2183 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115736.260, "dur": 0.151 , "args": { "External id": 2185,"Record function id": 0, "Ev Idx": 2184 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115736.882, "dur": 0.641 , "args": { "External id": 2186,"Record function id": 0, "Ev Idx": 2185 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115737.062, "dur": 0.371 , "args": { "External id": 2187,"Record function id": 0, "Ev Idx": 2186 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115737.252, "dur": 0.070 , "args": { "External id": 2188,"Record function id": 0, "Ev Idx": 2187 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115737.783, "dur": 0.641 , "args": { "External id": 2189,"Record function id": 0, "Ev Idx": 2188 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115737.944, "dur": 0.390 , "args": { "External id": 2190,"Record function id": 0, "Ev Idx": 2189 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115738.144, "dur": 0.090 , "args": { "External id": 2191,"Record function id": 0, "Ev Idx": 2190 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115739.476, "dur": 1.293 , "args": { "External id": 2192,"Record function id": 0, "Ev Idx": 2191 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115739.667, "dur": 1.002 , "args": { "External id": 2193,"Record function id": 0, "Ev Idx": 2192 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381115750.116, "dur": 4.208 , "args": { "External id": 2194,"Record function id": 0, "Ev Idx": 2193 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381115759.584, "dur": 4.679 , "args": { "External id": 2195,"Record function id": 0, "Ev Idx": 2194 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381115769.112, "dur": 4.889 , "args": { "External id": 2196,"Record function id": 0, "Ev Idx": 2195 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381115776.305, "dur": 17.703 , "args": { "External id": 2197,"Record function id": 0, "Ev Idx": 2196 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115776.806, "dur": 0.501 , "args": { "External id": 2198,"Record function id": 0, "Ev Idx": 2197 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115777.036, "dur": 0.161 , "args": { "External id": 2199,"Record function id": 0, "Ev Idx": 2198 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115777.547, "dur": 0.471 , "args": { "External id": 2200,"Record function id": 0, "Ev Idx": 2199 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115777.738, "dur": 0.180 , "args": { "External id": 2201,"Record function id": 0, "Ev Idx": 2200 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115778.269, "dur": 0.420 , "args": { "External id": 2202,"Record function id": 0, "Ev Idx": 2201 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115778.479, "dur": 0.110 , "args": { "External id": 2203,"Record function id": 0, "Ev Idx": 2202 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381115779.060, "dur": 13.515 , "args": { "External id": 2204,"Record function id": 0, "Ev Idx": 2203 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381115779.942, "dur": 1.943 , "args": { "External id": 2205,"Record function id": 0, "Ev Idx": 2204 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115780.272, "dur": 1.483 , "args": { "External id": 2206,"Record function id": 0, "Ev Idx": 2205 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115782.306, "dur": 1.303 , "args": { "External id": 2207,"Record function id": 0, "Ev Idx": 2206 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115784.059, "dur": 1.193 , "args": { "External id": 2208,"Record function id": 0, "Ev Idx": 2207 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115785.532, "dur": 1.212 , "args": { "External id": 2209,"Record function id": 0, "Ev Idx": 2208 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115791.213, "dur": 1.002 , "args": { "External id": 2210,"Record function id": 0, "Ev Idx": 2209 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115792.846, "dur": 0.441 , "args": { "External id": 2211,"Record function id": 0, "Ev Idx": 2210 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115793.046, "dur": 0.140 , "args": { "External id": 2212,"Record function id": 0, "Ev Idx": 2211 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115800.730, "dur": 6.482 , "args": { "External id": 2213,"Record function id": 0, "Ev Idx": 2212 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115809.567, "dur": 10.750 , "args": { "External id": 2214,"Record function id": 0, "Ev Idx": 2213 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115810.158, "dur": 0.681 , "args": { "External id": 2215,"Record function id": 0, "Ev Idx": 2214 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115810.328, "dur": 0.421 , "args": { "External id": 2216,"Record function id": 0, "Ev Idx": 2215 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115810.559, "dur": 0.080 , "args": { "External id": 2217,"Record function id": 0, "Ev Idx": 2216 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115811.130, "dur": 0.661 , "args": { "External id": 2218,"Record function id": 0, "Ev Idx": 2217 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115811.310, "dur": 0.391 , "args": { "External id": 2219,"Record function id": 0, "Ev Idx": 2218 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115811.520, "dur": 0.071 , "args": { "External id": 2220,"Record function id": 0, "Ev Idx": 2219 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115812.051, "dur": 0.652 , "args": { "External id": 2221,"Record function id": 0, "Ev Idx": 2220 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115812.212, "dur": 0.390 , "args": { "External id": 2222,"Record function id": 0, "Ev Idx": 2221 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115812.422, "dur": 0.080 , "args": { "External id": 2223,"Record function id": 0, "Ev Idx": 2222 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115813.845, "dur": 1.312 , "args": { "External id": 2224,"Record function id": 0, "Ev Idx": 2223 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115814.055, "dur": 0.992 , "args": { "External id": 2225,"Record function id": 0, "Ev Idx": 2224 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381115824.294, "dur": 5.430 , "args": { "External id": 2226,"Record function id": 0, "Ev Idx": 2225 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115833.041, "dur": 5.289 , "args": { "External id": 2227,"Record function id": 0, "Ev Idx": 2226 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115841.546, "dur": 5.280 , "args": { "External id": 2228,"Record function id": 0, "Ev Idx": 2227 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115850.423, "dur": 10.379 , "args": { "External id": 2229,"Record function id": 0, "Ev Idx": 2228 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115850.944, "dur": 0.671 , "args": { "External id": 2230,"Record function id": 0, "Ev Idx": 2229 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115851.114, "dur": 0.411 , "args": { "External id": 2231,"Record function id": 0, "Ev Idx": 2230 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115851.325, "dur": 0.090 , "args": { "External id": 2232,"Record function id": 0, "Ev Idx": 2231 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115851.896, "dur": 0.641 , "args": { "External id": 2233,"Record function id": 0, "Ev Idx": 2232 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115852.056, "dur": 0.391 , "args": { "External id": 2234,"Record function id": 0, "Ev Idx": 2233 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115852.266, "dur": 0.070 , "args": { "External id": 2235,"Record function id": 0, "Ev Idx": 2234 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115852.787, "dur": 0.621 , "args": { "External id": 2236,"Record function id": 0, "Ev Idx": 2235 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115852.948, "dur": 0.370 , "args": { "External id": 2237,"Record function id": 0, "Ev Idx": 2236 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115853.138, "dur": 0.080 , "args": { "External id": 2238,"Record function id": 0, "Ev Idx": 2237 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115854.480, "dur": 1.273 , "args": { "External id": 2239,"Record function id": 0, "Ev Idx": 2238 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115854.661, "dur": 0.982 , "args": { "External id": 2240,"Record function id": 0, "Ev Idx": 2239 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115864.339, "dur": 10.008 , "args": { "External id": 2241,"Record function id": 0, "Ev Idx": 2240 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115864.830, "dur": 0.641 , "args": { "External id": 2242,"Record function id": 0, "Ev Idx": 2241 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115864.990, "dur": 0.391 , "args": { "External id": 2243,"Record function id": 0, "Ev Idx": 2242 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115865.200, "dur": 0.081 , "args": { "External id": 2244,"Record function id": 0, "Ev Idx": 2243 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115865.741, "dur": 0.622 , "args": { "External id": 2245,"Record function id": 0, "Ev Idx": 2244 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115865.902, "dur": 0.370 , "args": { "External id": 2246,"Record function id": 0, "Ev Idx": 2245 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115866.102, "dur": 0.070 , "args": { "External id": 2247,"Record function id": 0, "Ev Idx": 2246 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115866.623, "dur": 0.671 , "args": { "External id": 2248,"Record function id": 0, "Ev Idx": 2247 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115866.793, "dur": 0.411 , "args": { "External id": 2249,"Record function id": 0, "Ev Idx": 2248 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115867.014, "dur": 0.090 , "args": { "External id": 2250,"Record function id": 0, "Ev Idx": 2249 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115868.286, "dur": 1.273 , "args": { "External id": 2251,"Record function id": 0, "Ev Idx": 2250 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115868.466, "dur": 0.982 , "args": { "External id": 2252,"Record function id": 0, "Ev Idx": 2251 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381115878.846, "dur": 4.137 , "args": { "External id": 2253,"Record function id": 0, "Ev Idx": 2252 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381115886.210, "dur": 5.199 , "args": { "External id": 2254,"Record function id": 0, "Ev Idx": 2253 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381115894.615, "dur": 10.119 , "args": { "External id": 2255,"Record function id": 0, "Ev Idx": 2254 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115895.126, "dur": 0.671 , "args": { "External id": 2256,"Record function id": 0, "Ev Idx": 2255 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115895.286, "dur": 0.411 , "args": { "External id": 2257,"Record function id": 0, "Ev Idx": 2256 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115895.517, "dur": 0.080 , "args": { "External id": 2258,"Record function id": 0, "Ev Idx": 2257 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115896.068, "dur": 0.661 , "args": { "External id": 2259,"Record function id": 0, "Ev Idx": 2258 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115896.228, "dur": 0.401 , "args": { "External id": 2260,"Record function id": 0, "Ev Idx": 2259 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115896.459, "dur": 0.070 , "args": { "External id": 2261,"Record function id": 0, "Ev Idx": 2260 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381115896.980, "dur": 0.641 , "args": { "External id": 2262,"Record function id": 0, "Ev Idx": 2261 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381115897.150, "dur": 0.381 , "args": { "External id": 2263,"Record function id": 0, "Ev Idx": 2262 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381115897.340, "dur": 0.080 , "args": { "External id": 2264,"Record function id": 0, "Ev Idx": 2263 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115898.663, "dur": 1.242 , "args": { "External id": 2265,"Record function id": 0, "Ev Idx": 2264 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115898.843, "dur": 0.962 , "args": { "External id": 2266,"Record function id": 0, "Ev Idx": 2265 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381115909.333, "dur": 4.618 , "args": { "External id": 2267,"Record function id": 0, "Ev Idx": 2266 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381115933.518, "dur": 23.764 , "args": { "External id": 2268,"Record function id": 0, "Ev Idx": 2267 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381115957.663, "dur": 334.783 , "args": { "External id": 2269,"Record function id": 0, "Ev Idx": 2268 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381115962.271, "dur": 10.830 , "args": { "External id": 2270,"Record function id": 0, "Ev Idx": 2269 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115965.667, "dur": 0.431 , "args": { "External id": 2271,"Record function id": 0, "Ev Idx": 2270 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381115966.599, "dur": 0.040 , "args": { "External id": 2272,"Record function id": 0, "Ev Idx": 2271 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381115967.040, "dur": 0.040 , "args": { "External id": 2273,"Record function id": 0, "Ev Idx": 2272 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381115968.463, "dur": 0.240 , "args": { "External id": 2274,"Record function id": 0, "Ev Idx": 2273 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381115969.054, "dur": 0.030 , "args": { "External id": 2275,"Record function id": 0, "Ev Idx": 2274 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381115969.404, "dur": 0.030 , "args": { "External id": 2276,"Record function id": 0, "Ev Idx": 2275 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381115975.856, "dur": 1.092 , "args": { "External id": 2277,"Record function id": 0, "Ev Idx": 2276 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381115979.122, "dur": 310.338 , "args": { "External id": 2278,"Record function id": 0, "Ev Idx": 2277 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381115987.618, "dur": 6.212 , "args": { "External id": 2279,"Record function id": 0, "Ev Idx": 2278 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116000.552, "dur": 6.743 , "args": { "External id": 2280,"Record function id": 0, "Ev Idx": 2279 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116013.116, "dur": 5.470 , "args": { "External id": 2281,"Record function id": 0, "Ev Idx": 2280 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116023.886, "dur": 4.909 , "args": { "External id": 2282,"Record function id": 0, "Ev Idx": 2281 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116031.239, "dur": 12.073 , "args": { "External id": 2283,"Record function id": 0, "Ev Idx": 2282 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116031.841, "dur": 1.082 , "args": { "External id": 2284,"Record function id": 0, "Ev Idx": 2283 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116032.061, "dur": 0.751 , "args": { "External id": 2285,"Record function id": 0, "Ev Idx": 2284 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116032.462, "dur": 0.210 , "args": { "External id": 2286,"Record function id": 0, "Ev Idx": 2285 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116033.223, "dur": 0.651 , "args": { "External id": 2287,"Record function id": 0, "Ev Idx": 2286 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116033.393, "dur": 0.391 , "args": { "External id": 2288,"Record function id": 0, "Ev Idx": 2287 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116033.604, "dur": 0.070 , "args": { "External id": 2289,"Record function id": 0, "Ev Idx": 2288 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116034.135, "dur": 0.621 , "args": { "External id": 2290,"Record function id": 0, "Ev Idx": 2289 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116034.285, "dur": 0.381 , "args": { "External id": 2291,"Record function id": 0, "Ev Idx": 2290 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116034.475, "dur": 0.091 , "args": { "External id": 2292,"Record function id": 0, "Ev Idx": 2291 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116036.028, "dur": 1.413 , "args": { "External id": 2293,"Record function id": 0, "Ev Idx": 2292 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116036.239, "dur": 1.092 , "args": { "External id": 2294,"Record function id": 0, "Ev Idx": 2293 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116048.702, "dur": 4.478 , "args": { "External id": 2295,"Record function id": 0, "Ev Idx": 2294 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116055.435, "dur": 0.851 , "args": { "External id": 2296,"Record function id": 0, "Ev Idx": 2295 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116058.009, "dur": 0.551 , "args": { "External id": 2297,"Record function id": 0, "Ev Idx": 2296 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381116062.387, "dur": 0.421 , "args": { "External id": 2298,"Record function id": 0, "Ev Idx": 2297 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381116062.638, "dur": 0.050 , "args": { "External id": 2299,"Record function id": 0, "Ev Idx": 2298 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381116065.613, "dur": 6.132 , "args": { "External id": 2300,"Record function id": 0, "Ev Idx": 2299 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116074.260, "dur": 10.629 , "args": { "External id": 2301,"Record function id": 0, "Ev Idx": 2300 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116074.821, "dur": 0.761 , "args": { "External id": 2302,"Record function id": 0, "Ev Idx": 2301 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116075.001, "dur": 0.491 , "args": { "External id": 2303,"Record function id": 0, "Ev Idx": 2302 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116075.271, "dur": 0.111 , "args": { "External id": 2304,"Record function id": 0, "Ev Idx": 2303 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116075.853, "dur": 0.611 , "args": { "External id": 2305,"Record function id": 0, "Ev Idx": 2304 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116076.003, "dur": 0.370 , "args": { "External id": 2306,"Record function id": 0, "Ev Idx": 2305 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116076.193, "dur": 0.070 , "args": { "External id": 2307,"Record function id": 0, "Ev Idx": 2306 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116076.714, "dur": 0.671 , "args": { "External id": 2308,"Record function id": 0, "Ev Idx": 2307 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116076.884, "dur": 0.411 , "args": { "External id": 2309,"Record function id": 0, "Ev Idx": 2308 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116077.095, "dur": 0.090 , "args": { "External id": 2310,"Record function id": 0, "Ev Idx": 2309 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116078.457, "dur": 1.343 , "args": { "External id": 2311,"Record function id": 0, "Ev Idx": 2310 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116078.658, "dur": 1.042 , "args": { "External id": 2312,"Record function id": 0, "Ev Idx": 2311 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116088.015, "dur": 3.857 , "args": { "External id": 2313,"Record function id": 0, "Ev Idx": 2312 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116092.934, "dur": 0.621 , "args": { "External id": 2314,"Record function id": 0, "Ev Idx": 2313 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116094.878, "dur": 0.561 , "args": { "External id": 2315,"Record function id": 0, "Ev Idx": 2314 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381116099.066, "dur": 0.330 , "args": { "External id": 2316,"Record function id": 0, "Ev Idx": 2315 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381116099.256, "dur": 0.040 , "args": { "External id": 2317,"Record function id": 0, "Ev Idx": 2316 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381116101.791, "dur": 5.490 , "args": { "External id": 2318,"Record function id": 0, "Ev Idx": 2317 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116109.445, "dur": 10.319 , "args": { "External id": 2319,"Record function id": 0, "Ev Idx": 2318 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116109.976, "dur": 0.731 , "args": { "External id": 2320,"Record function id": 0, "Ev Idx": 2319 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116110.176, "dur": 0.441 , "args": { "External id": 2321,"Record function id": 0, "Ev Idx": 2320 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116110.417, "dur": 0.080 , "args": { "External id": 2322,"Record function id": 0, "Ev Idx": 2321 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116110.978, "dur": 0.621 , "args": { "External id": 2323,"Record function id": 0, "Ev Idx": 2322 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116111.148, "dur": 0.361 , "args": { "External id": 2324,"Record function id": 0, "Ev Idx": 2323 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116111.329, "dur": 0.070 , "args": { "External id": 2325,"Record function id": 0, "Ev Idx": 2324 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116111.860, "dur": 0.621 , "args": { "External id": 2326,"Record function id": 0, "Ev Idx": 2325 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116112.020, "dur": 0.371 , "args": { "External id": 2327,"Record function id": 0, "Ev Idx": 2326 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116112.200, "dur": 0.090 , "args": { "External id": 2328,"Record function id": 0, "Ev Idx": 2327 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116113.533, "dur": 1.302 , "args": { "External id": 2329,"Record function id": 0, "Ev Idx": 2328 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116113.733, "dur": 1.002 , "args": { "External id": 2330,"Record function id": 0, "Ev Idx": 2329 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381116124.243, "dur": 4.157 , "args": { "External id": 2331,"Record function id": 0, "Ev Idx": 2330 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381116134.081, "dur": 4.849 , "args": { "External id": 2332,"Record function id": 0, "Ev Idx": 2331 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381116144.160, "dur": 4.829 , "args": { "External id": 2333,"Record function id": 0, "Ev Idx": 2332 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381116151.113, "dur": 16.961 , "args": { "External id": 2334,"Record function id": 0, "Ev Idx": 2333 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116151.593, "dur": 0.521 , "args": { "External id": 2335,"Record function id": 0, "Ev Idx": 2334 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116151.834, "dur": 0.170 , "args": { "External id": 2336,"Record function id": 0, "Ev Idx": 2335 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116152.365, "dur": 0.391 , "args": { "External id": 2337,"Record function id": 0, "Ev Idx": 2336 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116152.555, "dur": 0.090 , "args": { "External id": 2338,"Record function id": 0, "Ev Idx": 2337 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116153.006, "dur": 0.391 , "args": { "External id": 2339,"Record function id": 0, "Ev Idx": 2338 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116153.206, "dur": 0.081 , "args": { "External id": 2340,"Record function id": 0, "Ev Idx": 2339 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381116153.758, "dur": 12.914 , "args": { "External id": 2341,"Record function id": 0, "Ev Idx": 2340 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381116154.669, "dur": 1.844 , "args": { "External id": 2342,"Record function id": 0, "Ev Idx": 2341 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116154.960, "dur": 1.442 , "args": { "External id": 2343,"Record function id": 0, "Ev Idx": 2342 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116156.963, "dur": 1.213 , "args": { "External id": 2344,"Record function id": 0, "Ev Idx": 2343 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116158.587, "dur": 1.212 , "args": { "External id": 2345,"Record function id": 0, "Ev Idx": 2344 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116160.079, "dur": 1.203 , "args": { "External id": 2346,"Record function id": 0, "Ev Idx": 2345 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116165.359, "dur": 0.972 , "args": { "External id": 2347,"Record function id": 0, "Ev Idx": 2346 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116166.942, "dur": 0.461 , "args": { "External id": 2348,"Record function id": 0, "Ev Idx": 2347 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116167.162, "dur": 0.131 , "args": { "External id": 2349,"Record function id": 0, "Ev Idx": 2348 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116174.797, "dur": 6.191 , "args": { "External id": 2350,"Record function id": 0, "Ev Idx": 2349 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116183.403, "dur": 10.640 , "args": { "External id": 2351,"Record function id": 0, "Ev Idx": 2350 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116184.034, "dur": 0.701 , "args": { "External id": 2352,"Record function id": 0, "Ev Idx": 2351 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116184.224, "dur": 0.421 , "args": { "External id": 2353,"Record function id": 0, "Ev Idx": 2352 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116184.445, "dur": 0.090 , "args": { "External id": 2354,"Record function id": 0, "Ev Idx": 2353 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116185.006, "dur": 0.641 , "args": { "External id": 2355,"Record function id": 0, "Ev Idx": 2354 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116185.176, "dur": 0.381 , "args": { "External id": 2356,"Record function id": 0, "Ev Idx": 2355 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116185.386, "dur": 0.071 , "args": { "External id": 2357,"Record function id": 0, "Ev Idx": 2356 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116185.917, "dur": 0.652 , "args": { "External id": 2358,"Record function id": 0, "Ev Idx": 2357 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116186.078, "dur": 0.400 , "args": { "External id": 2359,"Record function id": 0, "Ev Idx": 2358 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116186.298, "dur": 0.080 , "args": { "External id": 2360,"Record function id": 0, "Ev Idx": 2359 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116187.681, "dur": 1.262 , "args": { "External id": 2361,"Record function id": 0, "Ev Idx": 2360 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116187.881, "dur": 0.952 , "args": { "External id": 2362,"Record function id": 0, "Ev Idx": 2361 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381116198.180, "dur": 5.230 , "args": { "External id": 2363,"Record function id": 0, "Ev Idx": 2362 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116206.165, "dur": 5.520 , "args": { "External id": 2364,"Record function id": 0, "Ev Idx": 2363 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116215.032, "dur": 5.119 , "args": { "External id": 2365,"Record function id": 0, "Ev Idx": 2364 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116223.678, "dur": 10.449 , "args": { "External id": 2366,"Record function id": 0, "Ev Idx": 2365 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116224.239, "dur": 0.721 , "args": { "External id": 2367,"Record function id": 0, "Ev Idx": 2366 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116224.429, "dur": 0.441 , "args": { "External id": 2368,"Record function id": 0, "Ev Idx": 2367 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116224.660, "dur": 0.100 , "args": { "External id": 2369,"Record function id": 0, "Ev Idx": 2368 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116225.261, "dur": 0.671 , "args": { "External id": 2370,"Record function id": 0, "Ev Idx": 2369 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116225.431, "dur": 0.411 , "args": { "External id": 2371,"Record function id": 0, "Ev Idx": 2370 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116225.661, "dur": 0.070 , "args": { "External id": 2372,"Record function id": 0, "Ev Idx": 2371 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116226.182, "dur": 0.621 , "args": { "External id": 2373,"Record function id": 0, "Ev Idx": 2372 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116226.343, "dur": 0.370 , "args": { "External id": 2374,"Record function id": 0, "Ev Idx": 2373 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116226.533, "dur": 0.080 , "args": { "External id": 2375,"Record function id": 0, "Ev Idx": 2374 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116227.865, "dur": 1.263 , "args": { "External id": 2376,"Record function id": 0, "Ev Idx": 2375 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116228.056, "dur": 0.972 , "args": { "External id": 2377,"Record function id": 0, "Ev Idx": 2376 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116237.894, "dur": 9.929 , "args": { "External id": 2378,"Record function id": 0, "Ev Idx": 2377 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116238.425, "dur": 0.651 , "args": { "External id": 2379,"Record function id": 0, "Ev Idx": 2378 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116238.595, "dur": 0.391 , "args": { "External id": 2380,"Record function id": 0, "Ev Idx": 2379 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116238.806, "dur": 0.070 , "args": { "External id": 2381,"Record function id": 0, "Ev Idx": 2380 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116239.337, "dur": 0.621 , "args": { "External id": 2382,"Record function id": 0, "Ev Idx": 2381 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116239.497, "dur": 0.361 , "args": { "External id": 2383,"Record function id": 0, "Ev Idx": 2382 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116239.687, "dur": 0.071 , "args": { "External id": 2384,"Record function id": 0, "Ev Idx": 2383 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116240.208, "dur": 0.622 , "args": { "External id": 2385,"Record function id": 0, "Ev Idx": 2384 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116240.369, "dur": 0.370 , "args": { "External id": 2386,"Record function id": 0, "Ev Idx": 2385 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116240.559, "dur": 0.080 , "args": { "External id": 2387,"Record function id": 0, "Ev Idx": 2386 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116241.841, "dur": 1.233 , "args": { "External id": 2388,"Record function id": 0, "Ev Idx": 2387 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116242.022, "dur": 0.942 , "args": { "External id": 2389,"Record function id": 0, "Ev Idx": 2388 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381116252.511, "dur": 3.988 , "args": { "External id": 2390,"Record function id": 0, "Ev Idx": 2389 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116259.775, "dur": 5.079 , "args": { "External id": 2391,"Record function id": 0, "Ev Idx": 2390 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116268.110, "dur": 10.129 , "args": { "External id": 2392,"Record function id": 0, "Ev Idx": 2391 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116268.621, "dur": 0.682 , "args": { "External id": 2393,"Record function id": 0, "Ev Idx": 2392 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116268.802, "dur": 0.410 , "args": { "External id": 2394,"Record function id": 0, "Ev Idx": 2393 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116269.022, "dur": 0.080 , "args": { "External id": 2395,"Record function id": 0, "Ev Idx": 2394 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116269.593, "dur": 0.631 , "args": { "External id": 2396,"Record function id": 0, "Ev Idx": 2395 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116269.743, "dur": 0.391 , "args": { "External id": 2397,"Record function id": 0, "Ev Idx": 2396 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116269.964, "dur": 0.070 , "args": { "External id": 2398,"Record function id": 0, "Ev Idx": 2397 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116270.485, "dur": 0.621 , "args": { "External id": 2399,"Record function id": 0, "Ev Idx": 2398 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116270.635, "dur": 0.381 , "args": { "External id": 2400,"Record function id": 0, "Ev Idx": 2399 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116270.845, "dur": 0.071 , "args": { "External id": 2401,"Record function id": 0, "Ev Idx": 2400 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116272.118, "dur": 1.252 , "args": { "External id": 2402,"Record function id": 0, "Ev Idx": 2401 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116272.298, "dur": 0.962 , "args": { "External id": 2403,"Record function id": 0, "Ev Idx": 2402 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381116283.108, "dur": 4.679 , "args": { "External id": 2404,"Record function id": 0, "Ev Idx": 2403 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381116307.714, "dur": 24.325 , "args": { "External id": 2405,"Record function id": 0, "Ev Idx": 2404 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381116332.400, "dur": 345.373 , "args": { "External id": 2406,"Record function id": 0, "Ev Idx": 2405 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381116336.989, "dur": 10.579 , "args": { "External id": 2407,"Record function id": 0, "Ev Idx": 2406 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116340.485, "dur": 0.391 , "args": { "External id": 2408,"Record function id": 0, "Ev Idx": 2407 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381116341.377, "dur": 0.040 , "args": { "External id": 2409,"Record function id": 0, "Ev Idx": 2408 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381116341.828, "dur": 0.050 , "args": { "External id": 2410,"Record function id": 0, "Ev Idx": 2409 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116343.020, "dur": 0.270 , "args": { "External id": 2411,"Record function id": 0, "Ev Idx": 2410 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381116343.631, "dur": 0.030 , "args": { "External id": 2412,"Record function id": 0, "Ev Idx": 2411 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381116343.992, "dur": 0.020 , "args": { "External id": 2413,"Record function id": 0, "Ev Idx": 2412 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381116350.113, "dur": 1.102 , "args": { "External id": 2414,"Record function id": 0, "Ev Idx": 2413 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381116353.399, "dur": 321.228 , "args": { "External id": 2415,"Record function id": 0, "Ev Idx": 2414 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116362.206, "dur": 6.171 , "args": { "External id": 2416,"Record function id": 0, "Ev Idx": 2415 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116374.849, "dur": 6.913 , "args": { "External id": 2417,"Record function id": 0, "Ev Idx": 2416 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116387.783, "dur": 5.571 , "args": { "External id": 2418,"Record function id": 0, "Ev Idx": 2417 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116398.994, "dur": 4.889 , "args": { "External id": 2419,"Record function id": 0, "Ev Idx": 2418 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116406.448, "dur": 12.273 , "args": { "External id": 2420,"Record function id": 0, "Ev Idx": 2419 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116407.039, "dur": 1.022 , "args": { "External id": 2421,"Record function id": 0, "Ev Idx": 2420 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116407.269, "dur": 0.682 , "args": { "External id": 2422,"Record function id": 0, "Ev Idx": 2421 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116407.560, "dur": 0.230 , "args": { "External id": 2423,"Record function id": 0, "Ev Idx": 2422 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116408.351, "dur": 0.622 , "args": { "External id": 2424,"Record function id": 0, "Ev Idx": 2423 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116408.502, "dur": 0.380 , "args": { "External id": 2425,"Record function id": 0, "Ev Idx": 2424 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116408.702, "dur": 0.080 , "args": { "External id": 2426,"Record function id": 0, "Ev Idx": 2425 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116409.243, "dur": 0.671 , "args": { "External id": 2427,"Record function id": 0, "Ev Idx": 2426 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116409.403, "dur": 0.421 , "args": { "External id": 2428,"Record function id": 0, "Ev Idx": 2427 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116409.624, "dur": 0.100 , "args": { "External id": 2429,"Record function id": 0, "Ev Idx": 2428 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116411.247, "dur": 1.402 , "args": { "External id": 2430,"Record function id": 0, "Ev Idx": 2429 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116411.467, "dur": 1.072 , "args": { "External id": 2431,"Record function id": 0, "Ev Idx": 2430 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116424.161, "dur": 4.458 , "args": { "External id": 2432,"Record function id": 0, "Ev Idx": 2431 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116430.723, "dur": 0.922 , "args": { "External id": 2433,"Record function id": 0, "Ev Idx": 2432 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116433.488, "dur": 0.571 , "args": { "External id": 2434,"Record function id": 0, "Ev Idx": 2433 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381116437.846, "dur": 0.461 , "args": { "External id": 2435,"Record function id": 0, "Ev Idx": 2434 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381116438.117, "dur": 0.050 , "args": { "External id": 2436,"Record function id": 0, "Ev Idx": 2435 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381116441.273, "dur": 6.031 , "args": { "External id": 2437,"Record function id": 0, "Ev Idx": 2436 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116449.508, "dur": 11.061 , "args": { "External id": 2438,"Record function id": 0, "Ev Idx": 2437 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116450.129, "dur": 0.732 , "args": { "External id": 2439,"Record function id": 0, "Ev Idx": 2438 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116450.300, "dur": 0.460 , "args": { "External id": 2440,"Record function id": 0, "Ev Idx": 2439 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116450.550, "dur": 0.100 , "args": { "External id": 2441,"Record function id": 0, "Ev Idx": 2440 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116451.141, "dur": 0.631 , "args": { "External id": 2442,"Record function id": 0, "Ev Idx": 2441 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116451.311, "dur": 0.371 , "args": { "External id": 2443,"Record function id": 0, "Ev Idx": 2442 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116451.502, "dur": 0.070 , "args": { "External id": 2444,"Record function id": 0, "Ev Idx": 2443 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116452.033, "dur": 0.631 , "args": { "External id": 2445,"Record function id": 0, "Ev Idx": 2444 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116452.193, "dur": 0.381 , "args": { "External id": 2446,"Record function id": 0, "Ev Idx": 2445 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116452.383, "dur": 0.091 , "args": { "External id": 2447,"Record function id": 0, "Ev Idx": 2446 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116453.776, "dur": 1.342 , "args": { "External id": 2448,"Record function id": 0, "Ev Idx": 2447 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116453.976, "dur": 1.032 , "args": { "External id": 2449,"Record function id": 0, "Ev Idx": 2448 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116463.895, "dur": 3.917 , "args": { "External id": 2450,"Record function id": 0, "Ev Idx": 2449 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116468.834, "dur": 0.621 , "args": { "External id": 2451,"Record function id": 0, "Ev Idx": 2450 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116470.788, "dur": 0.591 , "args": { "External id": 2452,"Record function id": 0, "Ev Idx": 2451 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381116475.406, "dur": 0.341 , "args": { "External id": 2453,"Record function id": 0, "Ev Idx": 2452 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381116475.597, "dur": 0.040 , "args": { "External id": 2454,"Record function id": 0, "Ev Idx": 2453 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381116478.302, "dur": 5.720 , "args": { "External id": 2455,"Record function id": 0, "Ev Idx": 2454 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116486.196, "dur": 10.480 , "args": { "External id": 2456,"Record function id": 0, "Ev Idx": 2455 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116486.727, "dur": 0.742 , "args": { "External id": 2457,"Record function id": 0, "Ev Idx": 2456 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116486.918, "dur": 0.461 , "args": { "External id": 2458,"Record function id": 0, "Ev Idx": 2457 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116487.158, "dur": 0.110 , "args": { "External id": 2459,"Record function id": 0, "Ev Idx": 2458 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116487.739, "dur": 0.621 , "args": { "External id": 2460,"Record function id": 0, "Ev Idx": 2459 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116487.900, "dur": 0.370 , "args": { "External id": 2461,"Record function id": 0, "Ev Idx": 2460 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116488.090, "dur": 0.070 , "args": { "External id": 2462,"Record function id": 0, "Ev Idx": 2461 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116488.621, "dur": 0.651 , "args": { "External id": 2463,"Record function id": 0, "Ev Idx": 2462 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116488.781, "dur": 0.401 , "args": { "External id": 2464,"Record function id": 0, "Ev Idx": 2463 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116488.982, "dur": 0.090 , "args": { "External id": 2465,"Record function id": 0, "Ev Idx": 2464 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116490.344, "dur": 1.313 , "args": { "External id": 2466,"Record function id": 0, "Ev Idx": 2465 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116490.524, "dur": 1.022 , "args": { "External id": 2467,"Record function id": 0, "Ev Idx": 2466 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381116501.194, "dur": 4.258 , "args": { "External id": 2468,"Record function id": 0, "Ev Idx": 2467 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381116510.712, "dur": 4.769 , "args": { "External id": 2469,"Record function id": 0, "Ev Idx": 2468 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381116520.570, "dur": 4.198 , "args": { "External id": 2470,"Record function id": 0, "Ev Idx": 2469 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381116527.042, "dur": 23.975 , "args": { "External id": 2471,"Record function id": 0, "Ev Idx": 2470 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116533.625, "dur": 0.591 , "args": { "External id": 2472,"Record function id": 0, "Ev Idx": 2471 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116533.915, "dur": 0.180 , "args": { "External id": 2473,"Record function id": 0, "Ev Idx": 2472 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116534.456, "dur": 0.411 , "args": { "External id": 2474,"Record function id": 0, "Ev Idx": 2473 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116534.667, "dur": 0.090 , "args": { "External id": 2475,"Record function id": 0, "Ev Idx": 2474 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116535.107, "dur": 0.401 , "args": { "External id": 2476,"Record function id": 0, "Ev Idx": 2475 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116535.328, "dur": 0.080 , "args": { "External id": 2477,"Record function id": 0, "Ev Idx": 2476 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381116535.869, "dur": 13.645 , "args": { "External id": 2478,"Record function id": 0, "Ev Idx": 2477 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381116536.861, "dur": 2.063 , "args": { "External id": 2479,"Record function id": 0, "Ev Idx": 2478 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116537.171, "dur": 1.653 , "args": { "External id": 2480,"Record function id": 0, "Ev Idx": 2479 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116539.395, "dur": 1.293 , "args": { "External id": 2481,"Record function id": 0, "Ev Idx": 2480 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116541.119, "dur": 1.222 , "args": { "External id": 2482,"Record function id": 0, "Ev Idx": 2481 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116542.621, "dur": 1.273 , "args": { "External id": 2483,"Record function id": 0, "Ev Idx": 2482 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116548.202, "dur": 0.931 , "args": { "External id": 2484,"Record function id": 0, "Ev Idx": 2483 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116549.805, "dur": 0.471 , "args": { "External id": 2485,"Record function id": 0, "Ev Idx": 2484 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116550.005, "dur": 0.160 , "args": { "External id": 2486,"Record function id": 0, "Ev Idx": 2485 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116558.230, "dur": 6.573 , "args": { "External id": 2487,"Record function id": 0, "Ev Idx": 2486 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116567.307, "dur": 10.971 , "args": { "External id": 2488,"Record function id": 0, "Ev Idx": 2487 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116567.989, "dur": 0.791 , "args": { "External id": 2489,"Record function id": 0, "Ev Idx": 2488 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116568.169, "dur": 0.511 , "args": { "External id": 2490,"Record function id": 0, "Ev Idx": 2489 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116568.379, "dur": 0.181 , "args": { "External id": 2491,"Record function id": 0, "Ev Idx": 2490 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116569.061, "dur": 0.671 , "args": { "External id": 2492,"Record function id": 0, "Ev Idx": 2491 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116569.241, "dur": 0.401 , "args": { "External id": 2493,"Record function id": 0, "Ev Idx": 2492 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116569.461, "dur": 0.080 , "args": { "External id": 2494,"Record function id": 0, "Ev Idx": 2493 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116570.002, "dur": 0.632 , "args": { "External id": 2495,"Record function id": 0, "Ev Idx": 2494 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116570.163, "dur": 0.380 , "args": { "External id": 2496,"Record function id": 0, "Ev Idx": 2495 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116570.373, "dur": 0.070 , "args": { "External id": 2497,"Record function id": 0, "Ev Idx": 2496 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116571.796, "dur": 1.282 , "args": { "External id": 2498,"Record function id": 0, "Ev Idx": 2497 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116571.986, "dur": 0.992 , "args": { "External id": 2499,"Record function id": 0, "Ev Idx": 2498 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381116582.486, "dur": 5.239 , "args": { "External id": 2500,"Record function id": 0, "Ev Idx": 2499 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116590.601, "dur": 5.079 , "args": { "External id": 2501,"Record function id": 0, "Ev Idx": 2500 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116598.926, "dur": 4.829 , "args": { "External id": 2502,"Record function id": 0, "Ev Idx": 2501 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116608.063, "dur": 10.470 , "args": { "External id": 2503,"Record function id": 0, "Ev Idx": 2502 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116608.604, "dur": 0.691 , "args": { "External id": 2504,"Record function id": 0, "Ev Idx": 2503 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116608.785, "dur": 0.420 , "args": { "External id": 2505,"Record function id": 0, "Ev Idx": 2504 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116609.005, "dur": 0.100 , "args": { "External id": 2506,"Record function id": 0, "Ev Idx": 2505 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116609.596, "dur": 0.661 , "args": { "External id": 2507,"Record function id": 0, "Ev Idx": 2506 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116609.756, "dur": 0.411 , "args": { "External id": 2508,"Record function id": 0, "Ev Idx": 2507 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116609.987, "dur": 0.070 , "args": { "External id": 2509,"Record function id": 0, "Ev Idx": 2508 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116610.518, "dur": 0.691 , "args": { "External id": 2510,"Record function id": 0, "Ev Idx": 2509 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116610.678, "dur": 0.441 , "args": { "External id": 2511,"Record function id": 0, "Ev Idx": 2510 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116610.868, "dur": 0.141 , "args": { "External id": 2512,"Record function id": 0, "Ev Idx": 2511 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116612.261, "dur": 1.272 , "args": { "External id": 2513,"Record function id": 0, "Ev Idx": 2512 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116612.451, "dur": 0.972 , "args": { "External id": 2514,"Record function id": 0, "Ev Idx": 2513 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116622.210, "dur": 10.158 , "args": { "External id": 2515,"Record function id": 0, "Ev Idx": 2514 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116622.730, "dur": 0.632 , "args": { "External id": 2516,"Record function id": 0, "Ev Idx": 2515 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116622.901, "dur": 0.370 , "args": { "External id": 2517,"Record function id": 0, "Ev Idx": 2516 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116623.091, "dur": 0.070 , "args": { "External id": 2518,"Record function id": 0, "Ev Idx": 2517 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116623.652, "dur": 0.631 , "args": { "External id": 2519,"Record function id": 0, "Ev Idx": 2518 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116623.813, "dur": 0.380 , "args": { "External id": 2520,"Record function id": 0, "Ev Idx": 2519 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116624.013, "dur": 0.070 , "args": { "External id": 2521,"Record function id": 0, "Ev Idx": 2520 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116624.534, "dur": 0.631 , "args": { "External id": 2522,"Record function id": 0, "Ev Idx": 2521 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116624.704, "dur": 0.371 , "args": { "External id": 2523,"Record function id": 0, "Ev Idx": 2522 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116624.895, "dur": 0.080 , "args": { "External id": 2524,"Record function id": 0, "Ev Idx": 2523 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116626.207, "dur": 1.262 , "args": { "External id": 2525,"Record function id": 0, "Ev Idx": 2524 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116626.397, "dur": 0.962 , "args": { "External id": 2526,"Record function id": 0, "Ev Idx": 2525 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381116636.937, "dur": 4.278 , "args": { "External id": 2527,"Record function id": 0, "Ev Idx": 2526 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116644.461, "dur": 5.440 , "args": { "External id": 2528,"Record function id": 0, "Ev Idx": 2527 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116653.337, "dur": 10.340 , "args": { "External id": 2529,"Record function id": 0, "Ev Idx": 2528 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116653.878, "dur": 0.682 , "args": { "External id": 2530,"Record function id": 0, "Ev Idx": 2529 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116654.059, "dur": 0.401 , "args": { "External id": 2531,"Record function id": 0, "Ev Idx": 2530 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116654.269, "dur": 0.090 , "args": { "External id": 2532,"Record function id": 0, "Ev Idx": 2531 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116654.860, "dur": 0.692 , "args": { "External id": 2533,"Record function id": 0, "Ev Idx": 2532 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116655.031, "dur": 0.430 , "args": { "External id": 2534,"Record function id": 0, "Ev Idx": 2533 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116655.281, "dur": 0.070 , "args": { "External id": 2535,"Record function id": 0, "Ev Idx": 2534 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116655.812, "dur": 0.661 , "args": { "External id": 2536,"Record function id": 0, "Ev Idx": 2535 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116655.992, "dur": 0.391 , "args": { "External id": 2537,"Record function id": 0, "Ev Idx": 2536 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116656.193, "dur": 0.080 , "args": { "External id": 2538,"Record function id": 0, "Ev Idx": 2537 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116657.535, "dur": 1.263 , "args": { "External id": 2539,"Record function id": 0, "Ev Idx": 2538 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116657.726, "dur": 0.971 , "args": { "External id": 2540,"Record function id": 0, "Ev Idx": 2539 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381116668.345, "dur": 4.659 , "args": { "External id": 2541,"Record function id": 0, "Ev Idx": 2540 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381116692.641, "dur": 23.864 , "args": { "External id": 2542,"Record function id": 0, "Ev Idx": 2541 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381116716.896, "dur": 341.826 , "args": { "External id": 2543,"Record function id": 0, "Ev Idx": 2542 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381116721.865, "dur": 10.480 , "args": { "External id": 2544,"Record function id": 0, "Ev Idx": 2543 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116725.382, "dur": 0.420 , "args": { "External id": 2545,"Record function id": 0, "Ev Idx": 2544 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381116726.253, "dur": 0.060 , "args": { "External id": 2546,"Record function id": 0, "Ev Idx": 2545 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381116726.714, "dur": 0.030 , "args": { "External id": 2547,"Record function id": 0, "Ev Idx": 2546 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116727.926, "dur": 0.241 , "args": { "External id": 2548,"Record function id": 0, "Ev Idx": 2547 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381116728.517, "dur": 0.030 , "args": { "External id": 2549,"Record function id": 0, "Ev Idx": 2548 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381116728.878, "dur": 0.030 , "args": { "External id": 2550,"Record function id": 0, "Ev Idx": 2549 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381116735.230, "dur": 1.092 , "args": { "External id": 2551,"Record function id": 0, "Ev Idx": 2550 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381116738.626, "dur": 317.211 , "args": { "External id": 2552,"Record function id": 0, "Ev Idx": 2551 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116747.423, "dur": 6.251 , "args": { "External id": 2553,"Record function id": 0, "Ev Idx": 2552 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116760.317, "dur": 6.983 , "args": { "External id": 2554,"Record function id": 0, "Ev Idx": 2553 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116773.171, "dur": 5.410 , "args": { "External id": 2555,"Record function id": 0, "Ev Idx": 2554 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116784.141, "dur": 5.039 , "args": { "External id": 2556,"Record function id": 0, "Ev Idx": 2555 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116791.625, "dur": 15.559 , "args": { "External id": 2557,"Record function id": 0, "Ev Idx": 2556 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116792.226, "dur": 0.952 , "args": { "External id": 2558,"Record function id": 0, "Ev Idx": 2557 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116792.446, "dur": 0.622 , "args": { "External id": 2559,"Record function id": 0, "Ev Idx": 2558 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116792.717, "dur": 0.220 , "args": { "External id": 2560,"Record function id": 0, "Ev Idx": 2559 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116796.474, "dur": 0.711 , "args": { "External id": 2561,"Record function id": 0, "Ev Idx": 2560 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116796.674, "dur": 0.421 , "args": { "External id": 2562,"Record function id": 0, "Ev Idx": 2561 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116796.925, "dur": 0.070 , "args": { "External id": 2563,"Record function id": 0, "Ev Idx": 2562 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116797.456, "dur": 0.721 , "args": { "External id": 2564,"Record function id": 0, "Ev Idx": 2563 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116797.636, "dur": 0.451 , "args": { "External id": 2565,"Record function id": 0, "Ev Idx": 2564 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116797.846, "dur": 0.141 , "args": { "External id": 2566,"Record function id": 0, "Ev Idx": 2565 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116799.550, "dur": 1.452 , "args": { "External id": 2567,"Record function id": 0, "Ev Idx": 2566 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116799.790, "dur": 1.092 , "args": { "External id": 2568,"Record function id": 0, "Ev Idx": 2567 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116812.854, "dur": 4.619 , "args": { "External id": 2569,"Record function id": 0, "Ev Idx": 2568 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116819.627, "dur": 0.972 , "args": { "External id": 2570,"Record function id": 0, "Ev Idx": 2569 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116822.472, "dur": 0.561 , "args": { "External id": 2571,"Record function id": 0, "Ev Idx": 2570 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381116826.800, "dur": 0.511 , "args": { "External id": 2572,"Record function id": 0, "Ev Idx": 2571 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381116827.111, "dur": 0.090 , "args": { "External id": 2573,"Record function id": 0, "Ev Idx": 2572 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381116830.267, "dur": 6.562 , "args": { "External id": 2574,"Record function id": 0, "Ev Idx": 2573 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116839.003, "dur": 10.750 , "args": { "External id": 2575,"Record function id": 0, "Ev Idx": 2574 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116839.554, "dur": 0.701 , "args": { "External id": 2576,"Record function id": 0, "Ev Idx": 2575 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116839.714, "dur": 0.451 , "args": { "External id": 2577,"Record function id": 0, "Ev Idx": 2576 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116839.965, "dur": 0.090 , "args": { "External id": 2578,"Record function id": 0, "Ev Idx": 2577 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116840.536, "dur": 0.631 , "args": { "External id": 2579,"Record function id": 0, "Ev Idx": 2578 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116840.696, "dur": 0.381 , "args": { "External id": 2580,"Record function id": 0, "Ev Idx": 2579 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116840.897, "dur": 0.070 , "args": { "External id": 2581,"Record function id": 0, "Ev Idx": 2580 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116841.428, "dur": 0.621 , "args": { "External id": 2582,"Record function id": 0, "Ev Idx": 2581 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116841.588, "dur": 0.371 , "args": { "External id": 2583,"Record function id": 0, "Ev Idx": 2582 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116841.778, "dur": 0.080 , "args": { "External id": 2584,"Record function id": 0, "Ev Idx": 2583 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116843.141, "dur": 1.332 , "args": { "External id": 2585,"Record function id": 0, "Ev Idx": 2584 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116843.351, "dur": 1.022 , "args": { "External id": 2586,"Record function id": 0, "Ev Idx": 2585 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381116853.079, "dur": 3.968 , "args": { "External id": 2587,"Record function id": 0, "Ev Idx": 2586 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116858.129, "dur": 0.641 , "args": { "External id": 2588,"Record function id": 0, "Ev Idx": 2587 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381116860.142, "dur": 0.561 , "args": { "External id": 2589,"Record function id": 0, "Ev Idx": 2588 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381116864.330, "dur": 0.321 , "args": { "External id": 2590,"Record function id": 0, "Ev Idx": 2589 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381116864.511, "dur": 0.040 , "args": { "External id": 2591,"Record function id": 0, "Ev Idx": 2590 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381116867.145, "dur": 5.761 , "args": { "External id": 2592,"Record function id": 0, "Ev Idx": 2591 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116875.110, "dur": 10.299 , "args": { "External id": 2593,"Record function id": 0, "Ev Idx": 2592 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116875.631, "dur": 0.712 , "args": { "External id": 2594,"Record function id": 0, "Ev Idx": 2593 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116875.832, "dur": 0.420 , "args": { "External id": 2595,"Record function id": 0, "Ev Idx": 2594 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116876.072, "dur": 0.080 , "args": { "External id": 2596,"Record function id": 0, "Ev Idx": 2595 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116876.613, "dur": 0.621 , "args": { "External id": 2597,"Record function id": 0, "Ev Idx": 2596 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116876.773, "dur": 0.371 , "args": { "External id": 2598,"Record function id": 0, "Ev Idx": 2597 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116876.964, "dur": 0.080 , "args": { "External id": 2599,"Record function id": 0, "Ev Idx": 2598 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116877.505, "dur": 0.631 , "args": { "External id": 2600,"Record function id": 0, "Ev Idx": 2599 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116877.665, "dur": 0.371 , "args": { "External id": 2601,"Record function id": 0, "Ev Idx": 2600 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116877.855, "dur": 0.081 , "args": { "External id": 2602,"Record function id": 0, "Ev Idx": 2601 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116879.228, "dur": 1.292 , "args": { "External id": 2603,"Record function id": 0, "Ev Idx": 2602 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116879.428, "dur": 0.982 , "args": { "External id": 2604,"Record function id": 0, "Ev Idx": 2603 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381116889.948, "dur": 4.188 , "args": { "External id": 2605,"Record function id": 0, "Ev Idx": 2604 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381116899.426, "dur": 4.788 , "args": { "External id": 2606,"Record function id": 0, "Ev Idx": 2605 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381116909.013, "dur": 4.389 , "args": { "External id": 2607,"Record function id": 0, "Ev Idx": 2606 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381116915.656, "dur": 17.472 , "args": { "External id": 2608,"Record function id": 0, "Ev Idx": 2607 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116916.137, "dur": 0.501 , "args": { "External id": 2609,"Record function id": 0, "Ev Idx": 2608 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116916.387, "dur": 0.140 , "args": { "External id": 2610,"Record function id": 0, "Ev Idx": 2609 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116916.888, "dur": 0.401 , "args": { "External id": 2611,"Record function id": 0, "Ev Idx": 2610 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116917.088, "dur": 0.091 , "args": { "External id": 2612,"Record function id": 0, "Ev Idx": 2611 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116917.539, "dur": 0.381 , "args": { "External id": 2613,"Record function id": 0, "Ev Idx": 2612 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116917.730, "dur": 0.080 , "args": { "External id": 2614,"Record function id": 0, "Ev Idx": 2613 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381116918.261, "dur": 13.395 , "args": { "External id": 2615,"Record function id": 0, "Ev Idx": 2614 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381116919.172, "dur": 1.924 , "args": { "External id": 2616,"Record function id": 0, "Ev Idx": 2615 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116919.523, "dur": 1.443 , "args": { "External id": 2617,"Record function id": 0, "Ev Idx": 2616 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116921.547, "dur": 1.282 , "args": { "External id": 2618,"Record function id": 0, "Ev Idx": 2617 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116923.290, "dur": 1.192 , "args": { "External id": 2619,"Record function id": 0, "Ev Idx": 2618 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116924.773, "dur": 1.292 , "args": { "External id": 2620,"Record function id": 0, "Ev Idx": 2619 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116930.353, "dur": 0.942 , "args": { "External id": 2621,"Record function id": 0, "Ev Idx": 2620 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116931.926, "dur": 0.471 , "args": { "External id": 2622,"Record function id": 0, "Ev Idx": 2621 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116932.156, "dur": 0.121 , "args": { "External id": 2623,"Record function id": 0, "Ev Idx": 2622 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116940.101, "dur": 6.502 , "args": { "External id": 2624,"Record function id": 0, "Ev Idx": 2623 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116949.148, "dur": 10.730 , "args": { "External id": 2625,"Record function id": 0, "Ev Idx": 2624 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116949.789, "dur": 0.682 , "args": { "External id": 2626,"Record function id": 0, "Ev Idx": 2625 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116949.960, "dur": 0.410 , "args": { "External id": 2627,"Record function id": 0, "Ev Idx": 2626 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116950.170, "dur": 0.100 , "args": { "External id": 2628,"Record function id": 0, "Ev Idx": 2627 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116950.761, "dur": 0.641 , "args": { "External id": 2629,"Record function id": 0, "Ev Idx": 2628 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116950.941, "dur": 0.371 , "args": { "External id": 2630,"Record function id": 0, "Ev Idx": 2629 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116951.132, "dur": 0.070 , "args": { "External id": 2631,"Record function id": 0, "Ev Idx": 2630 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116951.653, "dur": 0.631 , "args": { "External id": 2632,"Record function id": 0, "Ev Idx": 2631 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116951.803, "dur": 0.391 , "args": { "External id": 2633,"Record function id": 0, "Ev Idx": 2632 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116952.013, "dur": 0.081 , "args": { "External id": 2634,"Record function id": 0, "Ev Idx": 2633 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116953.446, "dur": 1.272 , "args": { "External id": 2635,"Record function id": 0, "Ev Idx": 2634 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116953.626, "dur": 0.982 , "args": { "External id": 2636,"Record function id": 0, "Ev Idx": 2635 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381116963.876, "dur": 5.249 , "args": { "External id": 2637,"Record function id": 0, "Ev Idx": 2636 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116971.941, "dur": 5.600 , "args": { "External id": 2638,"Record function id": 0, "Ev Idx": 2637 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381116980.917, "dur": 4.919 , "args": { "External id": 2639,"Record function id": 0, "Ev Idx": 2638 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381116989.443, "dur": 10.430 , "args": { "External id": 2640,"Record function id": 0, "Ev Idx": 2639 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116989.984, "dur": 0.701 , "args": { "External id": 2641,"Record function id": 0, "Ev Idx": 2640 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116990.164, "dur": 0.421 , "args": { "External id": 2642,"Record function id": 0, "Ev Idx": 2641 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116990.395, "dur": 0.090 , "args": { "External id": 2643,"Record function id": 0, "Ev Idx": 2642 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116990.976, "dur": 0.641 , "args": { "External id": 2644,"Record function id": 0, "Ev Idx": 2643 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116991.136, "dur": 0.391 , "args": { "External id": 2645,"Record function id": 0, "Ev Idx": 2644 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116991.347, "dur": 0.070 , "args": { "External id": 2646,"Record function id": 0, "Ev Idx": 2645 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381116991.868, "dur": 0.661 , "args": { "External id": 2647,"Record function id": 0, "Ev Idx": 2646 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381116992.028, "dur": 0.411 , "args": { "External id": 2648,"Record function id": 0, "Ev Idx": 2647 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381116992.258, "dur": 0.070 , "args": { "External id": 2649,"Record function id": 0, "Ev Idx": 2648 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116993.591, "dur": 1.272 , "args": { "External id": 2650,"Record function id": 0, "Ev Idx": 2649 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381116993.781, "dur": 0.972 , "args": { "External id": 2651,"Record function id": 0, "Ev Idx": 2650 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117003.630, "dur": 10.068 , "args": { "External id": 2652,"Record function id": 0, "Ev Idx": 2651 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117004.140, "dur": 0.642 , "args": { "External id": 2653,"Record function id": 0, "Ev Idx": 2652 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117004.311, "dur": 0.370 , "args": { "External id": 2654,"Record function id": 0, "Ev Idx": 2653 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117004.511, "dur": 0.070 , "args": { "External id": 2655,"Record function id": 0, "Ev Idx": 2654 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117005.052, "dur": 0.621 , "args": { "External id": 2656,"Record function id": 0, "Ev Idx": 2655 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117005.202, "dur": 0.381 , "args": { "External id": 2657,"Record function id": 0, "Ev Idx": 2656 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117005.403, "dur": 0.070 , "args": { "External id": 2658,"Record function id": 0, "Ev Idx": 2657 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117005.934, "dur": 0.621 , "args": { "External id": 2659,"Record function id": 0, "Ev Idx": 2658 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117006.084, "dur": 0.381 , "args": { "External id": 2660,"Record function id": 0, "Ev Idx": 2659 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117006.284, "dur": 0.081 , "args": { "External id": 2661,"Record function id": 0, "Ev Idx": 2660 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117007.567, "dur": 1.242 , "args": { "External id": 2662,"Record function id": 0, "Ev Idx": 2661 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117007.737, "dur": 0.972 , "args": { "External id": 2663,"Record function id": 0, "Ev Idx": 2662 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381117018.337, "dur": 4.218 , "args": { "External id": 2664,"Record function id": 0, "Ev Idx": 2663 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117025.921, "dur": 5.340 , "args": { "External id": 2665,"Record function id": 0, "Ev Idx": 2664 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117034.517, "dur": 10.419 , "args": { "External id": 2666,"Record function id": 0, "Ev Idx": 2665 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117035.048, "dur": 0.661 , "args": { "External id": 2667,"Record function id": 0, "Ev Idx": 2666 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117035.218, "dur": 0.401 , "args": { "External id": 2668,"Record function id": 0, "Ev Idx": 2667 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117035.439, "dur": 0.080 , "args": { "External id": 2669,"Record function id": 0, "Ev Idx": 2668 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117036.010, "dur": 0.661 , "args": { "External id": 2670,"Record function id": 0, "Ev Idx": 2669 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117036.160, "dur": 0.421 , "args": { "External id": 2671,"Record function id": 0, "Ev Idx": 2670 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117036.390, "dur": 0.081 , "args": { "External id": 2672,"Record function id": 0, "Ev Idx": 2671 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117036.921, "dur": 0.612 , "args": { "External id": 2673,"Record function id": 0, "Ev Idx": 2672 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117037.082, "dur": 0.360 , "args": { "External id": 2674,"Record function id": 0, "Ev Idx": 2673 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117037.272, "dur": 0.070 , "args": { "External id": 2675,"Record function id": 0, "Ev Idx": 2674 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117038.615, "dur": 1.282 , "args": { "External id": 2676,"Record function id": 0, "Ev Idx": 2675 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117038.825, "dur": 0.972 , "args": { "External id": 2677,"Record function id": 0, "Ev Idx": 2676 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381117049.505, "dur": 4.769 , "args": { "External id": 2678,"Record function id": 0, "Ev Idx": 2677 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381117073.209, "dur": 24.145 , "args": { "External id": 2679,"Record function id": 0, "Ev Idx": 2678 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381117097.745, "dur": 337.508 , "args": { "External id": 2680,"Record function id": 0, "Ev Idx": 2679 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381117102.564, "dur": 10.519 , "args": { "External id": 2681,"Record function id": 0, "Ev Idx": 2680 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117106.020, "dur": 0.391 , "args": { "External id": 2682,"Record function id": 0, "Ev Idx": 2681 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381117106.902, "dur": 0.050 , "args": { "External id": 2683,"Record function id": 0, "Ev Idx": 2682 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381117107.343, "dur": 0.040 , "args": { "External id": 2684,"Record function id": 0, "Ev Idx": 2683 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117108.545, "dur": 0.250 , "args": { "External id": 2685,"Record function id": 0, "Ev Idx": 2684 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381117109.146, "dur": 0.040 , "args": { "External id": 2686,"Record function id": 0, "Ev Idx": 2685 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381117109.507, "dur": 0.020 , "args": { "External id": 2687,"Record function id": 0, "Ev Idx": 2686 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381117115.828, "dur": 1.102 , "args": { "External id": 2688,"Record function id": 0, "Ev Idx": 2687 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381117119.135, "dur": 313.042 , "args": { "External id": 2689,"Record function id": 0, "Ev Idx": 2688 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117127.981, "dur": 6.332 , "args": { "External id": 2690,"Record function id": 0, "Ev Idx": 2689 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117140.925, "dur": 7.484 , "args": { "External id": 2691,"Record function id": 0, "Ev Idx": 2690 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117154.390, "dur": 5.360 , "args": { "External id": 2692,"Record function id": 0, "Ev Idx": 2691 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117165.250, "dur": 5.080 , "args": { "External id": 2693,"Record function id": 0, "Ev Idx": 2692 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117172.654, "dur": 12.233 , "args": { "External id": 2694,"Record function id": 0, "Ev Idx": 2693 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117173.265, "dur": 0.972 , "args": { "External id": 2695,"Record function id": 0, "Ev Idx": 2694 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117173.486, "dur": 0.641 , "args": { "External id": 2696,"Record function id": 0, "Ev Idx": 2695 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117173.776, "dur": 0.201 , "args": { "External id": 2697,"Record function id": 0, "Ev Idx": 2696 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117174.588, "dur": 0.641 , "args": { "External id": 2698,"Record function id": 0, "Ev Idx": 2697 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117174.758, "dur": 0.381 , "args": { "External id": 2699,"Record function id": 0, "Ev Idx": 2698 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117174.968, "dur": 0.071 , "args": { "External id": 2700,"Record function id": 0, "Ev Idx": 2699 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117175.499, "dur": 0.642 , "args": { "External id": 2701,"Record function id": 0, "Ev Idx": 2700 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117175.670, "dur": 0.380 , "args": { "External id": 2702,"Record function id": 0, "Ev Idx": 2701 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117175.850, "dur": 0.100 , "args": { "External id": 2703,"Record function id": 0, "Ev Idx": 2702 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117177.403, "dur": 1.383 , "args": { "External id": 2704,"Record function id": 0, "Ev Idx": 2703 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117177.613, "dur": 1.062 , "args": { "External id": 2705,"Record function id": 0, "Ev Idx": 2704 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117190.347, "dur": 4.328 , "args": { "External id": 2706,"Record function id": 0, "Ev Idx": 2705 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117196.759, "dur": 0.882 , "args": { "External id": 2707,"Record function id": 0, "Ev Idx": 2706 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117199.414, "dur": 0.561 , "args": { "External id": 2708,"Record function id": 0, "Ev Idx": 2707 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381117203.812, "dur": 0.701 , "args": { "External id": 2709,"Record function id": 0, "Ev Idx": 2708 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381117204.343, "dur": 0.050 , "args": { "External id": 2710,"Record function id": 0, "Ev Idx": 2709 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381117207.379, "dur": 6.402 , "args": { "External id": 2711,"Record function id": 0, "Ev Idx": 2710 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117215.955, "dur": 10.599 , "args": { "External id": 2712,"Record function id": 0, "Ev Idx": 2711 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117216.486, "dur": 0.701 , "args": { "External id": 2713,"Record function id": 0, "Ev Idx": 2712 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117216.656, "dur": 0.431 , "args": { "External id": 2714,"Record function id": 0, "Ev Idx": 2713 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117216.896, "dur": 0.081 , "args": { "External id": 2715,"Record function id": 0, "Ev Idx": 2714 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117217.448, "dur": 0.631 , "args": { "External id": 2716,"Record function id": 0, "Ev Idx": 2715 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117217.608, "dur": 0.381 , "args": { "External id": 2717,"Record function id": 0, "Ev Idx": 2716 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117217.808, "dur": 0.080 , "args": { "External id": 2718,"Record function id": 0, "Ev Idx": 2717 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117218.339, "dur": 0.631 , "args": { "External id": 2719,"Record function id": 0, "Ev Idx": 2718 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117218.499, "dur": 0.381 , "args": { "External id": 2720,"Record function id": 0, "Ev Idx": 2719 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117218.700, "dur": 0.080 , "args": { "External id": 2721,"Record function id": 0, "Ev Idx": 2720 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117220.102, "dur": 1.283 , "args": { "External id": 2722,"Record function id": 0, "Ev Idx": 2721 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117220.283, "dur": 1.002 , "args": { "External id": 2723,"Record function id": 0, "Ev Idx": 2722 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117229.881, "dur": 3.807 , "args": { "External id": 2724,"Record function id": 0, "Ev Idx": 2723 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117234.810, "dur": 0.661 , "args": { "External id": 2725,"Record function id": 0, "Ev Idx": 2724 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117236.824, "dur": 0.581 , "args": { "External id": 2726,"Record function id": 0, "Ev Idx": 2725 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381117241.031, "dur": 0.331 , "args": { "External id": 2727,"Record function id": 0, "Ev Idx": 2726 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381117241.202, "dur": 0.050 , "args": { "External id": 2728,"Record function id": 0, "Ev Idx": 2727 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381117244.037, "dur": 5.711 , "args": { "External id": 2729,"Record function id": 0, "Ev Idx": 2728 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117251.992, "dur": 10.309 , "args": { "External id": 2730,"Record function id": 0, "Ev Idx": 2729 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117252.563, "dur": 0.691 , "args": { "External id": 2731,"Record function id": 0, "Ev Idx": 2730 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117252.743, "dur": 0.421 , "args": { "External id": 2732,"Record function id": 0, "Ev Idx": 2731 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117252.974, "dur": 0.080 , "args": { "External id": 2733,"Record function id": 0, "Ev Idx": 2732 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117253.535, "dur": 0.621 , "args": { "External id": 2734,"Record function id": 0, "Ev Idx": 2733 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117253.685, "dur": 0.381 , "args": { "External id": 2735,"Record function id": 0, "Ev Idx": 2734 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117253.875, "dur": 0.080 , "args": { "External id": 2736,"Record function id": 0, "Ev Idx": 2735 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117254.416, "dur": 0.611 , "args": { "External id": 2737,"Record function id": 0, "Ev Idx": 2736 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117254.567, "dur": 0.370 , "args": { "External id": 2738,"Record function id": 0, "Ev Idx": 2737 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117254.757, "dur": 0.080 , "args": { "External id": 2739,"Record function id": 0, "Ev Idx": 2738 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117256.079, "dur": 1.283 , "args": { "External id": 2740,"Record function id": 0, "Ev Idx": 2739 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117256.280, "dur": 0.982 , "args": { "External id": 2741,"Record function id": 0, "Ev Idx": 2740 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381117266.880, "dur": 4.237 , "args": { "External id": 2742,"Record function id": 0, "Ev Idx": 2741 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381117276.467, "dur": 4.990 , "args": { "External id": 2743,"Record function id": 0, "Ev Idx": 2742 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381117286.406, "dur": 4.628 , "args": { "External id": 2744,"Record function id": 0, "Ev Idx": 2743 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381117293.209, "dur": 17.031 , "args": { "External id": 2745,"Record function id": 0, "Ev Idx": 2744 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117293.699, "dur": 0.491 , "args": { "External id": 2746,"Record function id": 0, "Ev Idx": 2745 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117293.930, "dur": 0.150 , "args": { "External id": 2747,"Record function id": 0, "Ev Idx": 2746 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117294.431, "dur": 0.401 , "args": { "External id": 2748,"Record function id": 0, "Ev Idx": 2747 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117294.631, "dur": 0.100 , "args": { "External id": 2749,"Record function id": 0, "Ev Idx": 2748 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117295.082, "dur": 0.381 , "args": { "External id": 2750,"Record function id": 0, "Ev Idx": 2749 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117295.282, "dur": 0.081 , "args": { "External id": 2751,"Record function id": 0, "Ev Idx": 2750 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381117295.813, "dur": 12.924 , "args": { "External id": 2752,"Record function id": 0, "Ev Idx": 2751 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381117296.695, "dur": 1.894 , "args": { "External id": 2753,"Record function id": 0, "Ev Idx": 2752 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117296.986, "dur": 1.472 , "args": { "External id": 2754,"Record function id": 0, "Ev Idx": 2753 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117299.029, "dur": 1.223 , "args": { "External id": 2755,"Record function id": 0, "Ev Idx": 2754 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117300.672, "dur": 1.223 , "args": { "External id": 2756,"Record function id": 0, "Ev Idx": 2755 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117302.165, "dur": 1.253 , "args": { "External id": 2757,"Record function id": 0, "Ev Idx": 2756 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117307.465, "dur": 0.932 , "args": { "External id": 2758,"Record function id": 0, "Ev Idx": 2757 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117309.008, "dur": 0.451 , "args": { "External id": 2759,"Record function id": 0, "Ev Idx": 2758 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117309.228, "dur": 0.131 , "args": { "External id": 2760,"Record function id": 0, "Ev Idx": 2759 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117317.093, "dur": 6.682 , "args": { "External id": 2761,"Record function id": 0, "Ev Idx": 2760 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117326.310, "dur": 10.650 , "args": { "External id": 2762,"Record function id": 0, "Ev Idx": 2761 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117326.911, "dur": 0.692 , "args": { "External id": 2763,"Record function id": 0, "Ev Idx": 2762 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117327.092, "dur": 0.420 , "args": { "External id": 2764,"Record function id": 0, "Ev Idx": 2763 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117327.312, "dur": 0.080 , "args": { "External id": 2765,"Record function id": 0, "Ev Idx": 2764 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117327.903, "dur": 0.651 , "args": { "External id": 2766,"Record function id": 0, "Ev Idx": 2765 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117328.073, "dur": 0.391 , "args": { "External id": 2767,"Record function id": 0, "Ev Idx": 2766 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117328.294, "dur": 0.070 , "args": { "External id": 2768,"Record function id": 0, "Ev Idx": 2767 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117328.815, "dur": 0.721 , "args": { "External id": 2769,"Record function id": 0, "Ev Idx": 2768 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117328.995, "dur": 0.451 , "args": { "External id": 2770,"Record function id": 0, "Ev Idx": 2769 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117329.196, "dur": 0.140 , "args": { "External id": 2771,"Record function id": 0, "Ev Idx": 2770 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117330.638, "dur": 1.273 , "args": { "External id": 2772,"Record function id": 0, "Ev Idx": 2771 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117330.829, "dur": 0.971 , "args": { "External id": 2773,"Record function id": 0, "Ev Idx": 2772 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381117340.947, "dur": 5.190 , "args": { "External id": 2774,"Record function id": 0, "Ev Idx": 2773 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117348.942, "dur": 5.310 , "args": { "External id": 2775,"Record function id": 0, "Ev Idx": 2774 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117357.548, "dur": 4.829 , "args": { "External id": 2776,"Record function id": 0, "Ev Idx": 2775 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117365.954, "dur": 10.229 , "args": { "External id": 2777,"Record function id": 0, "Ev Idx": 2776 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117366.465, "dur": 0.681 , "args": { "External id": 2778,"Record function id": 0, "Ev Idx": 2777 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117366.645, "dur": 0.401 , "args": { "External id": 2779,"Record function id": 0, "Ev Idx": 2778 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117366.866, "dur": 0.070 , "args": { "External id": 2780,"Record function id": 0, "Ev Idx": 2779 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117367.427, "dur": 0.631 , "args": { "External id": 2781,"Record function id": 0, "Ev Idx": 2780 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117367.577, "dur": 0.391 , "args": { "External id": 2782,"Record function id": 0, "Ev Idx": 2781 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117367.797, "dur": 0.071 , "args": { "External id": 2783,"Record function id": 0, "Ev Idx": 2782 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117368.318, "dur": 0.632 , "args": { "External id": 2784,"Record function id": 0, "Ev Idx": 2783 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117368.479, "dur": 0.380 , "args": { "External id": 2785,"Record function id": 0, "Ev Idx": 2784 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117368.679, "dur": 0.080 , "args": { "External id": 2786,"Record function id": 0, "Ev Idx": 2785 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117370.001, "dur": 1.253 , "args": { "External id": 2787,"Record function id": 0, "Ev Idx": 2786 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117370.182, "dur": 0.962 , "args": { "External id": 2788,"Record function id": 0, "Ev Idx": 2787 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117379.970, "dur": 9.949 , "args": { "External id": 2789,"Record function id": 0, "Ev Idx": 2788 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117380.451, "dur": 0.641 , "args": { "External id": 2790,"Record function id": 0, "Ev Idx": 2789 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117380.621, "dur": 0.381 , "args": { "External id": 2791,"Record function id": 0, "Ev Idx": 2790 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117380.822, "dur": 0.080 , "args": { "External id": 2792,"Record function id": 0, "Ev Idx": 2791 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117381.353, "dur": 0.631 , "args": { "External id": 2793,"Record function id": 0, "Ev Idx": 2792 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117381.513, "dur": 0.381 , "args": { "External id": 2794,"Record function id": 0, "Ev Idx": 2793 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117381.713, "dur": 0.070 , "args": { "External id": 2795,"Record function id": 0, "Ev Idx": 2794 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117382.234, "dur": 0.712 , "args": { "External id": 2796,"Record function id": 0, "Ev Idx": 2795 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117382.405, "dur": 0.450 , "args": { "External id": 2797,"Record function id": 0, "Ev Idx": 2796 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117382.615, "dur": 0.130 , "args": { "External id": 2798,"Record function id": 0, "Ev Idx": 2797 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117383.937, "dur": 1.243 , "args": { "External id": 2799,"Record function id": 0, "Ev Idx": 2798 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117384.108, "dur": 0.962 , "args": { "External id": 2800,"Record function id": 0, "Ev Idx": 2799 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381117394.798, "dur": 4.127 , "args": { "External id": 2801,"Record function id": 0, "Ev Idx": 2800 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117402.322, "dur": 5.630 , "args": { "External id": 2802,"Record function id": 0, "Ev Idx": 2801 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117411.188, "dur": 10.089 , "args": { "External id": 2803,"Record function id": 0, "Ev Idx": 2802 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117411.679, "dur": 0.651 , "args": { "External id": 2804,"Record function id": 0, "Ev Idx": 2803 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117411.859, "dur": 0.381 , "args": { "External id": 2805,"Record function id": 0, "Ev Idx": 2804 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117412.060, "dur": 0.080 , "args": { "External id": 2806,"Record function id": 0, "Ev Idx": 2805 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117412.601, "dur": 0.661 , "args": { "External id": 2807,"Record function id": 0, "Ev Idx": 2806 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117412.751, "dur": 0.421 , "args": { "External id": 2808,"Record function id": 0, "Ev Idx": 2807 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117412.992, "dur": 0.070 , "args": { "External id": 2809,"Record function id": 0, "Ev Idx": 2808 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117413.512, "dur": 0.632 , "args": { "External id": 2810,"Record function id": 0, "Ev Idx": 2809 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117413.673, "dur": 0.380 , "args": { "External id": 2811,"Record function id": 0, "Ev Idx": 2810 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117413.873, "dur": 0.080 , "args": { "External id": 2812,"Record function id": 0, "Ev Idx": 2811 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117415.166, "dur": 1.242 , "args": { "External id": 2813,"Record function id": 0, "Ev Idx": 2812 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117415.346, "dur": 0.952 , "args": { "External id": 2814,"Record function id": 0, "Ev Idx": 2813 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381117425.966, "dur": 4.638 , "args": { "External id": 2815,"Record function id": 0, "Ev Idx": 2814 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381117449.499, "dur": 23.073 , "args": { "External id": 2816,"Record function id": 0, "Ev Idx": 2815 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381117472.913, "dur": 337.829 , "args": { "External id": 2817,"Record function id": 0, "Ev Idx": 2816 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381117477.712, "dur": 10.279 , "args": { "External id": 2818,"Record function id": 0, "Ev Idx": 2817 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117481.078, "dur": 0.401 , "args": { "External id": 2819,"Record function id": 0, "Ev Idx": 2818 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381117481.950, "dur": 0.030 , "args": { "External id": 2820,"Record function id": 0, "Ev Idx": 2819 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381117482.361, "dur": 0.040 , "args": { "External id": 2821,"Record function id": 0, "Ev Idx": 2820 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117483.553, "dur": 0.260 , "args": { "External id": 2822,"Record function id": 0, "Ev Idx": 2821 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381117484.154, "dur": 0.030 , "args": { "External id": 2823,"Record function id": 0, "Ev Idx": 2822 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381117484.505, "dur": 0.030 , "args": { "External id": 2824,"Record function id": 0, "Ev Idx": 2823 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381117490.696, "dur": 1.453 , "args": { "External id": 2825,"Record function id": 0, "Ev Idx": 2824 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381117494.153, "dur": 313.593 , "args": { "External id": 2826,"Record function id": 0, "Ev Idx": 2825 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117502.849, "dur": 6.181 , "args": { "External id": 2827,"Record function id": 0, "Ev Idx": 2826 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117515.562, "dur": 7.013 , "args": { "External id": 2828,"Record function id": 0, "Ev Idx": 2827 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117532.384, "dur": 5.600 , "args": { "External id": 2829,"Record function id": 0, "Ev Idx": 2828 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117543.354, "dur": 5.090 , "args": { "External id": 2830,"Record function id": 0, "Ev Idx": 2829 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117550.928, "dur": 12.203 , "args": { "External id": 2831,"Record function id": 0, "Ev Idx": 2830 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117551.549, "dur": 0.952 , "args": { "External id": 2832,"Record function id": 0, "Ev Idx": 2831 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117551.770, "dur": 0.631 , "args": { "External id": 2833,"Record function id": 0, "Ev Idx": 2832 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117552.060, "dur": 0.201 , "args": { "External id": 2834,"Record function id": 0, "Ev Idx": 2833 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117552.822, "dur": 0.631 , "args": { "External id": 2835,"Record function id": 0, "Ev Idx": 2834 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117552.982, "dur": 0.381 , "args": { "External id": 2836,"Record function id": 0, "Ev Idx": 2835 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117553.192, "dur": 0.071 , "args": { "External id": 2837,"Record function id": 0, "Ev Idx": 2836 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117553.723, "dur": 0.632 , "args": { "External id": 2838,"Record function id": 0, "Ev Idx": 2837 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117553.884, "dur": 0.380 , "args": { "External id": 2839,"Record function id": 0, "Ev Idx": 2838 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117554.074, "dur": 0.080 , "args": { "External id": 2840,"Record function id": 0, "Ev Idx": 2839 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117555.657, "dur": 1.453 , "args": { "External id": 2841,"Record function id": 0, "Ev Idx": 2840 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117555.887, "dur": 1.103 , "args": { "External id": 2842,"Record function id": 0, "Ev Idx": 2841 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117568.521, "dur": 4.579 , "args": { "External id": 2843,"Record function id": 0, "Ev Idx": 2842 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117575.324, "dur": 0.921 , "args": { "External id": 2844,"Record function id": 0, "Ev Idx": 2843 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117577.999, "dur": 0.561 , "args": { "External id": 2845,"Record function id": 0, "Ev Idx": 2844 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381117582.277, "dur": 0.440 , "args": { "External id": 2846,"Record function id": 0, "Ev Idx": 2845 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381117582.537, "dur": 0.070 , "args": { "External id": 2847,"Record function id": 0, "Ev Idx": 2846 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381117585.603, "dur": 6.091 , "args": { "External id": 2848,"Record function id": 0, "Ev Idx": 2847 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117593.868, "dur": 10.710 , "args": { "External id": 2849,"Record function id": 0, "Ev Idx": 2848 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117594.399, "dur": 0.742 , "args": { "External id": 2850,"Record function id": 0, "Ev Idx": 2849 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117594.570, "dur": 0.470 , "args": { "External id": 2851,"Record function id": 0, "Ev Idx": 2850 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117594.830, "dur": 0.100 , "args": { "External id": 2852,"Record function id": 0, "Ev Idx": 2851 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117595.411, "dur": 0.631 , "args": { "External id": 2853,"Record function id": 0, "Ev Idx": 2852 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117595.581, "dur": 0.381 , "args": { "External id": 2854,"Record function id": 0, "Ev Idx": 2853 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117595.782, "dur": 0.070 , "args": { "External id": 2855,"Record function id": 0, "Ev Idx": 2854 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117596.303, "dur": 0.641 , "args": { "External id": 2856,"Record function id": 0, "Ev Idx": 2855 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117596.463, "dur": 0.381 , "args": { "External id": 2857,"Record function id": 0, "Ev Idx": 2856 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117596.653, "dur": 0.091 , "args": { "External id": 2858,"Record function id": 0, "Ev Idx": 2857 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117598.036, "dur": 1.302 , "args": { "External id": 2859,"Record function id": 0, "Ev Idx": 2858 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117598.246, "dur": 0.982 , "args": { "External id": 2860,"Record function id": 0, "Ev Idx": 2859 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117607.834, "dur": 3.958 , "args": { "External id": 2861,"Record function id": 0, "Ev Idx": 2860 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117612.844, "dur": 0.621 , "args": { "External id": 2862,"Record function id": 0, "Ev Idx": 2861 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117614.797, "dur": 0.591 , "args": { "External id": 2863,"Record function id": 0, "Ev Idx": 2862 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381117618.865, "dur": 0.340 , "args": { "External id": 2864,"Record function id": 0, "Ev Idx": 2863 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381117619.055, "dur": 0.050 , "args": { "External id": 2865,"Record function id": 0, "Ev Idx": 2864 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381117621.720, "dur": 5.530 , "args": { "External id": 2866,"Record function id": 0, "Ev Idx": 2865 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117629.464, "dur": 10.250 , "args": { "External id": 2867,"Record function id": 0, "Ev Idx": 2866 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117629.995, "dur": 0.722 , "args": { "External id": 2868,"Record function id": 0, "Ev Idx": 2867 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117630.176, "dur": 0.451 , "args": { "External id": 2869,"Record function id": 0, "Ev Idx": 2868 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117630.436, "dur": 0.090 , "args": { "External id": 2870,"Record function id": 0, "Ev Idx": 2869 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117630.997, "dur": 0.631 , "args": { "External id": 2871,"Record function id": 0, "Ev Idx": 2870 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117631.168, "dur": 0.360 , "args": { "External id": 2872,"Record function id": 0, "Ev Idx": 2871 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117631.358, "dur": 0.070 , "args": { "External id": 2873,"Record function id": 0, "Ev Idx": 2872 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117631.879, "dur": 0.621 , "args": { "External id": 2874,"Record function id": 0, "Ev Idx": 2873 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117632.029, "dur": 0.381 , "args": { "External id": 2875,"Record function id": 0, "Ev Idx": 2874 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117632.220, "dur": 0.080 , "args": { "External id": 2876,"Record function id": 0, "Ev Idx": 2875 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117633.552, "dur": 1.272 , "args": { "External id": 2877,"Record function id": 0, "Ev Idx": 2876 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117633.722, "dur": 0.992 , "args": { "External id": 2878,"Record function id": 0, "Ev Idx": 2877 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381117644.102, "dur": 4.288 , "args": { "External id": 2879,"Record function id": 0, "Ev Idx": 2878 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381117653.740, "dur": 4.658 , "args": { "External id": 2880,"Record function id": 0, "Ev Idx": 2879 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381117663.327, "dur": 4.449 , "args": { "External id": 2881,"Record function id": 0, "Ev Idx": 2880 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381117670.080, "dur": 17.032 , "args": { "External id": 2882,"Record function id": 0, "Ev Idx": 2881 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117670.581, "dur": 0.541 , "args": { "External id": 2883,"Record function id": 0, "Ev Idx": 2882 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117670.851, "dur": 0.151 , "args": { "External id": 2884,"Record function id": 0, "Ev Idx": 2883 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117671.362, "dur": 0.391 , "args": { "External id": 2885,"Record function id": 0, "Ev Idx": 2884 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117671.563, "dur": 0.080 , "args": { "External id": 2886,"Record function id": 0, "Ev Idx": 2885 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117671.994, "dur": 0.390 , "args": { "External id": 2887,"Record function id": 0, "Ev Idx": 2886 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117672.194, "dur": 0.080 , "args": { "External id": 2888,"Record function id": 0, "Ev Idx": 2887 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381117672.735, "dur": 12.914 , "args": { "External id": 2889,"Record function id": 0, "Ev Idx": 2888 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381117673.557, "dur": 1.823 , "args": { "External id": 2890,"Record function id": 0, "Ev Idx": 2889 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117673.857, "dur": 1.403 , "args": { "External id": 2891,"Record function id": 0, "Ev Idx": 2890 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117675.821, "dur": 1.212 , "args": { "External id": 2892,"Record function id": 0, "Ev Idx": 2891 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117677.484, "dur": 1.182 , "args": { "External id": 2893,"Record function id": 0, "Ev Idx": 2892 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117678.957, "dur": 1.242 , "args": { "External id": 2894,"Record function id": 0, "Ev Idx": 2893 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117684.337, "dur": 0.961 , "args": { "External id": 2895,"Record function id": 0, "Ev Idx": 2894 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117685.930, "dur": 0.460 , "args": { "External id": 2896,"Record function id": 0, "Ev Idx": 2895 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117686.140, "dur": 0.140 , "args": { "External id": 2897,"Record function id": 0, "Ev Idx": 2896 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117693.644, "dur": 6.382 , "args": { "External id": 2898,"Record function id": 0, "Ev Idx": 2897 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117702.380, "dur": 10.600 , "args": { "External id": 2899,"Record function id": 0, "Ev Idx": 2898 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117702.991, "dur": 0.732 , "args": { "External id": 2900,"Record function id": 0, "Ev Idx": 2899 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117703.182, "dur": 0.451 , "args": { "External id": 2901,"Record function id": 0, "Ev Idx": 2900 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117703.412, "dur": 0.110 , "args": { "External id": 2902,"Record function id": 0, "Ev Idx": 2901 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117703.993, "dur": 0.641 , "args": { "External id": 2903,"Record function id": 0, "Ev Idx": 2902 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117704.153, "dur": 0.381 , "args": { "External id": 2904,"Record function id": 0, "Ev Idx": 2903 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117704.364, "dur": 0.070 , "args": { "External id": 2905,"Record function id": 0, "Ev Idx": 2904 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117704.905, "dur": 0.641 , "args": { "External id": 2906,"Record function id": 0, "Ev Idx": 2905 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117705.055, "dur": 0.401 , "args": { "External id": 2907,"Record function id": 0, "Ev Idx": 2906 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117705.266, "dur": 0.080 , "args": { "External id": 2908,"Record function id": 0, "Ev Idx": 2907 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117706.648, "dur": 1.293 , "args": { "External id": 2909,"Record function id": 0, "Ev Idx": 2908 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117706.849, "dur": 0.981 , "args": { "External id": 2910,"Record function id": 0, "Ev Idx": 2909 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381117717.078, "dur": 5.219 , "args": { "External id": 2911,"Record function id": 0, "Ev Idx": 2910 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117725.253, "dur": 5.209 , "args": { "External id": 2912,"Record function id": 0, "Ev Idx": 2911 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117733.759, "dur": 4.969 , "args": { "External id": 2913,"Record function id": 0, "Ev Idx": 2912 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117742.244, "dur": 10.390 , "args": { "External id": 2914,"Record function id": 0, "Ev Idx": 2913 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117742.755, "dur": 0.672 , "args": { "External id": 2915,"Record function id": 0, "Ev Idx": 2914 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117742.916, "dur": 0.420 , "args": { "External id": 2916,"Record function id": 0, "Ev Idx": 2915 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117743.146, "dur": 0.080 , "args": { "External id": 2917,"Record function id": 0, "Ev Idx": 2916 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117743.697, "dur": 0.661 , "args": { "External id": 2918,"Record function id": 0, "Ev Idx": 2917 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117743.877, "dur": 0.391 , "args": { "External id": 2919,"Record function id": 0, "Ev Idx": 2918 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117744.088, "dur": 0.070 , "args": { "External id": 2920,"Record function id": 0, "Ev Idx": 2919 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117744.609, "dur": 0.631 , "args": { "External id": 2921,"Record function id": 0, "Ev Idx": 2920 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117744.769, "dur": 0.381 , "args": { "External id": 2922,"Record function id": 0, "Ev Idx": 2921 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117744.969, "dur": 0.081 , "args": { "External id": 2923,"Record function id": 0, "Ev Idx": 2922 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117746.282, "dur": 1.262 , "args": { "External id": 2924,"Record function id": 0, "Ev Idx": 2923 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117746.482, "dur": 0.962 , "args": { "External id": 2925,"Record function id": 0, "Ev Idx": 2924 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117756.351, "dur": 10.028 , "args": { "External id": 2926,"Record function id": 0, "Ev Idx": 2925 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117756.842, "dur": 0.641 , "args": { "External id": 2927,"Record function id": 0, "Ev Idx": 2926 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117757.002, "dur": 0.391 , "args": { "External id": 2928,"Record function id": 0, "Ev Idx": 2927 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117757.202, "dur": 0.080 , "args": { "External id": 2929,"Record function id": 0, "Ev Idx": 2928 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117757.753, "dur": 0.621 , "args": { "External id": 2930,"Record function id": 0, "Ev Idx": 2929 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117757.904, "dur": 0.380 , "args": { "External id": 2931,"Record function id": 0, "Ev Idx": 2930 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117758.104, "dur": 0.070 , "args": { "External id": 2932,"Record function id": 0, "Ev Idx": 2931 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117758.625, "dur": 0.641 , "args": { "External id": 2933,"Record function id": 0, "Ev Idx": 2932 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117758.795, "dur": 0.381 , "args": { "External id": 2934,"Record function id": 0, "Ev Idx": 2933 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117758.996, "dur": 0.080 , "args": { "External id": 2935,"Record function id": 0, "Ev Idx": 2934 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117760.328, "dur": 1.232 , "args": { "External id": 2936,"Record function id": 0, "Ev Idx": 2935 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117760.508, "dur": 0.952 , "args": { "External id": 2937,"Record function id": 0, "Ev Idx": 2936 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381117771.128, "dur": 4.158 , "args": { "External id": 2938,"Record function id": 0, "Ev Idx": 2937 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117778.452, "dur": 5.240 , "args": { "External id": 2939,"Record function id": 0, "Ev Idx": 2938 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117786.827, "dur": 10.209 , "args": { "External id": 2940,"Record function id": 0, "Ev Idx": 2939 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117787.318, "dur": 0.742 , "args": { "External id": 2941,"Record function id": 0, "Ev Idx": 2940 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117787.499, "dur": 0.471 , "args": { "External id": 2942,"Record function id": 0, "Ev Idx": 2941 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117787.709, "dur": 0.150 , "args": { "External id": 2943,"Record function id": 0, "Ev Idx": 2942 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117788.330, "dur": 0.651 , "args": { "External id": 2944,"Record function id": 0, "Ev Idx": 2943 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117788.490, "dur": 0.401 , "args": { "External id": 2945,"Record function id": 0, "Ev Idx": 2944 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117788.721, "dur": 0.070 , "args": { "External id": 2946,"Record function id": 0, "Ev Idx": 2945 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117789.242, "dur": 0.651 , "args": { "External id": 2947,"Record function id": 0, "Ev Idx": 2946 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117789.412, "dur": 0.391 , "args": { "External id": 2948,"Record function id": 0, "Ev Idx": 2947 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117789.613, "dur": 0.070 , "args": { "External id": 2949,"Record function id": 0, "Ev Idx": 2948 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117790.955, "dur": 1.262 , "args": { "External id": 2950,"Record function id": 0, "Ev Idx": 2949 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117791.135, "dur": 0.982 , "args": { "External id": 2951,"Record function id": 0, "Ev Idx": 2950 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381117801.535, "dur": 4.598 , "args": { "External id": 2952,"Record function id": 0, "Ev Idx": 2951 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381117825.559, "dur": 22.122 , "args": { "External id": 2953,"Record function id": 0, "Ev Idx": 2952 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381117848.021, "dur": 339.673 , "args": { "External id": 2954,"Record function id": 0, "Ev Idx": 2953 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381117852.860, "dur": 10.309 , "args": { "External id": 2955,"Record function id": 0, "Ev Idx": 2954 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117856.207, "dur": 0.370 , "args": { "External id": 2956,"Record function id": 0, "Ev Idx": 2955 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381117857.058, "dur": 0.030 , "args": { "External id": 2957,"Record function id": 0, "Ev Idx": 2956 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381117857.509, "dur": 0.040 , "args": { "External id": 2958,"Record function id": 0, "Ev Idx": 2957 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117858.701, "dur": 0.211 , "args": { "External id": 2959,"Record function id": 0, "Ev Idx": 2958 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381117859.262, "dur": 0.040 , "args": { "External id": 2960,"Record function id": 0, "Ev Idx": 2959 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381117859.633, "dur": 0.030 , "args": { "External id": 2961,"Record function id": 0, "Ev Idx": 2960 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381117865.895, "dur": 1.021 , "args": { "External id": 2962,"Record function id": 0, "Ev Idx": 2961 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381117869.050, "dur": 315.648 , "args": { "External id": 2963,"Record function id": 0, "Ev Idx": 2962 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117877.687, "dur": 6.411 , "args": { "External id": 2964,"Record function id": 0, "Ev Idx": 2963 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117890.711, "dur": 6.893 , "args": { "External id": 2965,"Record function id": 0, "Ev Idx": 2964 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117903.444, "dur": 5.441 , "args": { "External id": 2966,"Record function id": 0, "Ev Idx": 2965 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381117914.245, "dur": 5.249 , "args": { "External id": 2967,"Record function id": 0, "Ev Idx": 2966 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117921.899, "dur": 12.012 , "args": { "External id": 2968,"Record function id": 0, "Ev Idx": 2967 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117922.490, "dur": 0.932 , "args": { "External id": 2969,"Record function id": 0, "Ev Idx": 2968 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117922.720, "dur": 0.591 , "args": { "External id": 2970,"Record function id": 0, "Ev Idx": 2969 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117923.031, "dur": 0.150 , "args": { "External id": 2971,"Record function id": 0, "Ev Idx": 2970 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117923.732, "dur": 0.621 , "args": { "External id": 2972,"Record function id": 0, "Ev Idx": 2971 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117923.903, "dur": 0.360 , "args": { "External id": 2973,"Record function id": 0, "Ev Idx": 2972 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117924.093, "dur": 0.070 , "args": { "External id": 2974,"Record function id": 0, "Ev Idx": 2973 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117924.624, "dur": 0.661 , "args": { "External id": 2975,"Record function id": 0, "Ev Idx": 2974 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117924.804, "dur": 0.381 , "args": { "External id": 2976,"Record function id": 0, "Ev Idx": 2975 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117925.005, "dur": 0.080 , "args": { "External id": 2977,"Record function id": 0, "Ev Idx": 2976 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117926.517, "dur": 1.433 , "args": { "External id": 2978,"Record function id": 0, "Ev Idx": 2977 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117926.738, "dur": 1.102 , "args": { "External id": 2979,"Record function id": 0, "Ev Idx": 2978 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117939.201, "dur": 4.408 , "args": { "External id": 2980,"Record function id": 0, "Ev Idx": 2979 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117945.823, "dur": 0.962 , "args": { "External id": 2981,"Record function id": 0, "Ev Idx": 2980 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117948.629, "dur": 0.561 , "args": { "External id": 2982,"Record function id": 0, "Ev Idx": 2981 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381117952.997, "dur": 0.471 , "args": { "External id": 2983,"Record function id": 0, "Ev Idx": 2982 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381117953.247, "dur": 0.090 , "args": { "External id": 2984,"Record function id": 0, "Ev Idx": 2983 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381117956.323, "dur": 6.532 , "args": { "External id": 2985,"Record function id": 0, "Ev Idx": 2984 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381117965.019, "dur": 10.750 , "args": { "External id": 2986,"Record function id": 0, "Ev Idx": 2985 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117965.560, "dur": 0.731 , "args": { "External id": 2987,"Record function id": 0, "Ev Idx": 2986 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117965.760, "dur": 0.441 , "args": { "External id": 2988,"Record function id": 0, "Ev Idx": 2987 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117966.011, "dur": 0.080 , "args": { "External id": 2989,"Record function id": 0, "Ev Idx": 2988 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117966.572, "dur": 0.631 , "args": { "External id": 2990,"Record function id": 0, "Ev Idx": 2989 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117966.732, "dur": 0.371 , "args": { "External id": 2991,"Record function id": 0, "Ev Idx": 2990 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117966.923, "dur": 0.080 , "args": { "External id": 2992,"Record function id": 0, "Ev Idx": 2991 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381117967.454, "dur": 0.701 , "args": { "External id": 2993,"Record function id": 0, "Ev Idx": 2992 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381117967.614, "dur": 0.451 , "args": { "External id": 2994,"Record function id": 0, "Ev Idx": 2993 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381117967.794, "dur": 0.171 , "args": { "External id": 2995,"Record function id": 0, "Ev Idx": 2994 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117969.287, "dur": 1.282 , "args": { "External id": 2996,"Record function id": 0, "Ev Idx": 2995 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381117969.497, "dur": 0.972 , "args": { "External id": 2997,"Record function id": 0, "Ev Idx": 2996 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381117978.925, "dur": 3.837 , "args": { "External id": 2998,"Record function id": 0, "Ev Idx": 2997 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117983.794, "dur": 0.621 , "args": { "External id": 2999,"Record function id": 0, "Ev Idx": 2998 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381117985.718, "dur": 0.561 , "args": { "External id": 3000,"Record function id": 0, "Ev Idx": 2999 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381117990.016, "dur": 0.330 , "args": { "External id": 3001,"Record function id": 0, "Ev Idx": 3000 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381117990.206, "dur": 0.040 , "args": { "External id": 3002,"Record function id": 0, "Ev Idx": 3001 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381117992.891, "dur": 5.420 , "args": { "External id": 3003,"Record function id": 0, "Ev Idx": 3002 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118000.465, "dur": 10.490 , "args": { "External id": 3004,"Record function id": 0, "Ev Idx": 3003 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118001.016, "dur": 0.761 , "args": { "External id": 3005,"Record function id": 0, "Ev Idx": 3004 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118001.216, "dur": 0.471 , "args": { "External id": 3006,"Record function id": 0, "Ev Idx": 3005 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118001.457, "dur": 0.120 , "args": { "External id": 3007,"Record function id": 0, "Ev Idx": 3006 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118002.058, "dur": 0.671 , "args": { "External id": 3008,"Record function id": 0, "Ev Idx": 3007 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118002.208, "dur": 0.431 , "args": { "External id": 3009,"Record function id": 0, "Ev Idx": 3008 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118002.459, "dur": 0.070 , "args": { "External id": 3010,"Record function id": 0, "Ev Idx": 3009 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118002.990, "dur": 0.631 , "args": { "External id": 3011,"Record function id": 0, "Ev Idx": 3010 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118003.140, "dur": 0.391 , "args": { "External id": 3012,"Record function id": 0, "Ev Idx": 3011 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118003.340, "dur": 0.091 , "args": { "External id": 3013,"Record function id": 0, "Ev Idx": 3012 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118004.683, "dur": 1.282 , "args": { "External id": 3014,"Record function id": 0, "Ev Idx": 3013 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118004.873, "dur": 0.982 , "args": { "External id": 3015,"Record function id": 0, "Ev Idx": 3014 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381118015.583, "dur": 4.258 , "args": { "External id": 3016,"Record function id": 0, "Ev Idx": 3015 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381118025.041, "dur": 4.899 , "args": { "External id": 3017,"Record function id": 0, "Ev Idx": 3016 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381118035.029, "dur": 4.328 , "args": { "External id": 3018,"Record function id": 0, "Ev Idx": 3017 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381118041.391, "dur": 17.142 , "args": { "External id": 3019,"Record function id": 0, "Ev Idx": 3018 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118041.872, "dur": 0.511 , "args": { "External id": 3020,"Record function id": 0, "Ev Idx": 3019 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118042.113, "dur": 0.160 , "args": { "External id": 3021,"Record function id": 0, "Ev Idx": 3020 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118042.624, "dur": 0.410 , "args": { "External id": 3022,"Record function id": 0, "Ev Idx": 3021 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118042.834, "dur": 0.090 , "args": { "External id": 3023,"Record function id": 0, "Ev Idx": 3022 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118043.275, "dur": 0.390 , "args": { "External id": 3024,"Record function id": 0, "Ev Idx": 3023 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118043.465, "dur": 0.090 , "args": { "External id": 3025,"Record function id": 0, "Ev Idx": 3024 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381118043.986, "dur": 13.084 , "args": { "External id": 3026,"Record function id": 0, "Ev Idx": 3025 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381118044.908, "dur": 1.903 , "args": { "External id": 3027,"Record function id": 0, "Ev Idx": 3026 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118045.198, "dur": 1.483 , "args": { "External id": 3028,"Record function id": 0, "Ev Idx": 3027 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118047.252, "dur": 1.212 , "args": { "External id": 3029,"Record function id": 0, "Ev Idx": 3028 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118048.885, "dur": 1.202 , "args": { "External id": 3030,"Record function id": 0, "Ev Idx": 3029 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118050.368, "dur": 1.252 , "args": { "External id": 3031,"Record function id": 0, "Ev Idx": 3030 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118055.788, "dur": 0.942 , "args": { "External id": 3032,"Record function id": 0, "Ev Idx": 3031 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118057.341, "dur": 0.461 , "args": { "External id": 3033,"Record function id": 0, "Ev Idx": 3032 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118057.551, "dur": 0.141 , "args": { "External id": 3034,"Record function id": 0, "Ev Idx": 3033 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118065.586, "dur": 6.552 , "args": { "External id": 3035,"Record function id": 0, "Ev Idx": 3034 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118074.543, "dur": 10.740 , "args": { "External id": 3036,"Record function id": 0, "Ev Idx": 3035 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118075.134, "dur": 0.701 , "args": { "External id": 3037,"Record function id": 0, "Ev Idx": 3036 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118075.324, "dur": 0.411 , "args": { "External id": 3038,"Record function id": 0, "Ev Idx": 3037 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118075.545, "dur": 0.090 , "args": { "External id": 3039,"Record function id": 0, "Ev Idx": 3038 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118076.136, "dur": 0.661 , "args": { "External id": 3040,"Record function id": 0, "Ev Idx": 3039 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118076.306, "dur": 0.401 , "args": { "External id": 3041,"Record function id": 0, "Ev Idx": 3040 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118076.527, "dur": 0.080 , "args": { "External id": 3042,"Record function id": 0, "Ev Idx": 3041 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118077.058, "dur": 0.651 , "args": { "External id": 3043,"Record function id": 0, "Ev Idx": 3042 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118077.218, "dur": 0.401 , "args": { "External id": 3044,"Record function id": 0, "Ev Idx": 3043 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118077.438, "dur": 0.080 , "args": { "External id": 3045,"Record function id": 0, "Ev Idx": 3044 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118078.831, "dur": 1.252 , "args": { "External id": 3046,"Record function id": 0, "Ev Idx": 3045 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118079.011, "dur": 0.972 , "args": { "External id": 3047,"Record function id": 0, "Ev Idx": 3046 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381118089.391, "dur": 5.179 , "args": { "External id": 3048,"Record function id": 0, "Ev Idx": 3047 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118097.586, "dur": 5.330 , "args": { "External id": 3049,"Record function id": 0, "Ev Idx": 3048 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118106.132, "dur": 4.979 , "args": { "External id": 3050,"Record function id": 0, "Ev Idx": 3049 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118114.738, "dur": 10.299 , "args": { "External id": 3051,"Record function id": 0, "Ev Idx": 3050 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118115.279, "dur": 0.671 , "args": { "External id": 3052,"Record function id": 0, "Ev Idx": 3051 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118115.459, "dur": 0.401 , "args": { "External id": 3053,"Record function id": 0, "Ev Idx": 3052 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118115.669, "dur": 0.081 , "args": { "External id": 3054,"Record function id": 0, "Ev Idx": 3053 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118116.241, "dur": 0.631 , "args": { "External id": 3055,"Record function id": 0, "Ev Idx": 3054 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118116.401, "dur": 0.381 , "args": { "External id": 3056,"Record function id": 0, "Ev Idx": 3055 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118116.611, "dur": 0.070 , "args": { "External id": 3057,"Record function id": 0, "Ev Idx": 3056 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118117.152, "dur": 0.661 , "args": { "External id": 3058,"Record function id": 0, "Ev Idx": 3057 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118117.333, "dur": 0.380 , "args": { "External id": 3059,"Record function id": 0, "Ev Idx": 3058 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118117.543, "dur": 0.070 , "args": { "External id": 3060,"Record function id": 0, "Ev Idx": 3059 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118118.855, "dur": 1.263 , "args": { "External id": 3061,"Record function id": 0, "Ev Idx": 3060 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118119.046, "dur": 0.962 , "args": { "External id": 3062,"Record function id": 0, "Ev Idx": 3061 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118128.744, "dur": 13.505 , "args": { "External id": 3063,"Record function id": 0, "Ev Idx": 3062 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118129.245, "dur": 0.621 , "args": { "External id": 3064,"Record function id": 0, "Ev Idx": 3063 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118129.395, "dur": 0.381 , "args": { "External id": 3065,"Record function id": 0, "Ev Idx": 3064 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118129.595, "dur": 0.081 , "args": { "External id": 3066,"Record function id": 0, "Ev Idx": 3065 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118130.126, "dur": 0.632 , "args": { "External id": 3067,"Record function id": 0, "Ev Idx": 3066 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118130.297, "dur": 0.370 , "args": { "External id": 3068,"Record function id": 0, "Ev Idx": 3067 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118130.487, "dur": 0.070 , "args": { "External id": 3069,"Record function id": 0, "Ev Idx": 3068 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118131.018, "dur": 0.651 , "args": { "External id": 3070,"Record function id": 0, "Ev Idx": 3069 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118131.188, "dur": 0.381 , "args": { "External id": 3071,"Record function id": 0, "Ev Idx": 3070 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118131.389, "dur": 0.080 , "args": { "External id": 3072,"Record function id": 0, "Ev Idx": 3071 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118135.857, "dur": 1.252 , "args": { "External id": 3073,"Record function id": 0, "Ev Idx": 3072 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118136.047, "dur": 0.962 , "args": { "External id": 3074,"Record function id": 0, "Ev Idx": 3073 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381118146.968, "dur": 4.418 , "args": { "External id": 3075,"Record function id": 0, "Ev Idx": 3074 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118154.742, "dur": 5.170 , "args": { "External id": 3076,"Record function id": 0, "Ev Idx": 3075 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118163.318, "dur": 10.209 , "args": { "External id": 3077,"Record function id": 0, "Ev Idx": 3076 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118163.859, "dur": 0.661 , "args": { "External id": 3078,"Record function id": 0, "Ev Idx": 3077 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118164.019, "dur": 0.411 , "args": { "External id": 3079,"Record function id": 0, "Ev Idx": 3078 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118164.240, "dur": 0.080 , "args": { "External id": 3080,"Record function id": 0, "Ev Idx": 3079 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118164.791, "dur": 0.611 , "args": { "External id": 3081,"Record function id": 0, "Ev Idx": 3080 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118164.941, "dur": 0.371 , "args": { "External id": 3082,"Record function id": 0, "Ev Idx": 3081 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118165.142, "dur": 0.060 , "args": { "External id": 3083,"Record function id": 0, "Ev Idx": 3082 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118165.653, "dur": 0.641 , "args": { "External id": 3084,"Record function id": 0, "Ev Idx": 3083 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118165.813, "dur": 0.391 , "args": { "External id": 3085,"Record function id": 0, "Ev Idx": 3084 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118166.023, "dur": 0.080 , "args": { "External id": 3086,"Record function id": 0, "Ev Idx": 3085 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118167.376, "dur": 1.232 , "args": { "External id": 3087,"Record function id": 0, "Ev Idx": 3086 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118167.546, "dur": 0.952 , "args": { "External id": 3088,"Record function id": 0, "Ev Idx": 3087 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381118178.346, "dur": 4.839 , "args": { "External id": 3089,"Record function id": 0, "Ev Idx": 3088 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381118202.491, "dur": 23.804 , "args": { "External id": 3090,"Record function id": 0, "Ev Idx": 3089 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381118226.666, "dur": 339.091 , "args": { "External id": 3091,"Record function id": 0, "Ev Idx": 3090 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381118231.425, "dur": 10.359 , "args": { "External id": 3092,"Record function id": 0, "Ev Idx": 3091 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118234.711, "dur": 0.381 , "args": { "External id": 3093,"Record function id": 0, "Ev Idx": 3092 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381118235.583, "dur": 0.040 , "args": { "External id": 3094,"Record function id": 0, "Ev Idx": 3093 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381118235.993, "dur": 0.041 , "args": { "External id": 3095,"Record function id": 0, "Ev Idx": 3094 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118237.226, "dur": 0.250 , "args": { "External id": 3096,"Record function id": 0, "Ev Idx": 3095 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381118237.837, "dur": 0.040 , "args": { "External id": 3097,"Record function id": 0, "Ev Idx": 3096 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381118238.228, "dur": 0.030 , "args": { "External id": 3098,"Record function id": 0, "Ev Idx": 3097 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381118244.519, "dur": 1.032 , "args": { "External id": 3099,"Record function id": 0, "Ev Idx": 3098 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381118247.595, "dur": 315.107 , "args": { "External id": 3100,"Record function id": 0, "Ev Idx": 3099 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118256.131, "dur": 6.121 , "args": { "External id": 3101,"Record function id": 0, "Ev Idx": 3100 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118268.674, "dur": 7.053 , "args": { "External id": 3102,"Record function id": 0, "Ev Idx": 3101 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118281.689, "dur": 5.219 , "args": { "External id": 3103,"Record function id": 0, "Ev Idx": 3102 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118292.138, "dur": 5.350 , "args": { "External id": 3104,"Record function id": 0, "Ev Idx": 3103 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118300.163, "dur": 12.183 , "args": { "External id": 3105,"Record function id": 0, "Ev Idx": 3104 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118300.764, "dur": 0.992 , "args": { "External id": 3106,"Record function id": 0, "Ev Idx": 3105 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118300.984, "dur": 0.662 , "args": { "External id": 3107,"Record function id": 0, "Ev Idx": 3106 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118301.325, "dur": 0.180 , "args": { "External id": 3108,"Record function id": 0, "Ev Idx": 3107 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118302.056, "dur": 0.622 , "args": { "External id": 3109,"Record function id": 0, "Ev Idx": 3108 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118302.217, "dur": 0.370 , "args": { "External id": 3110,"Record function id": 0, "Ev Idx": 3109 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118302.407, "dur": 0.080 , "args": { "External id": 3111,"Record function id": 0, "Ev Idx": 3110 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118302.958, "dur": 0.671 , "args": { "External id": 3112,"Record function id": 0, "Ev Idx": 3111 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118303.108, "dur": 0.431 , "args": { "External id": 3113,"Record function id": 0, "Ev Idx": 3112 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118303.299, "dur": 0.130 , "args": { "External id": 3114,"Record function id": 0, "Ev Idx": 3113 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118304.862, "dur": 1.402 , "args": { "External id": 3115,"Record function id": 0, "Ev Idx": 3114 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118305.072, "dur": 1.092 , "args": { "External id": 3116,"Record function id": 0, "Ev Idx": 3115 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118317.746, "dur": 4.488 , "args": { "External id": 3117,"Record function id": 0, "Ev Idx": 3116 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118324.368, "dur": 0.922 , "args": { "External id": 3118,"Record function id": 0, "Ev Idx": 3117 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118327.033, "dur": 0.551 , "args": { "External id": 3119,"Record function id": 0, "Ev Idx": 3118 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381118331.371, "dur": 0.451 , "args": { "External id": 3120,"Record function id": 0, "Ev Idx": 3119 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381118331.631, "dur": 0.061 , "args": { "External id": 3121,"Record function id": 0, "Ev Idx": 3120 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381118335.008, "dur": 6.522 , "args": { "External id": 3122,"Record function id": 0, "Ev Idx": 3121 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118343.554, "dur": 10.619 , "args": { "External id": 3123,"Record function id": 0, "Ev Idx": 3122 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118344.115, "dur": 0.721 , "args": { "External id": 3124,"Record function id": 0, "Ev Idx": 3123 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118344.285, "dur": 0.451 , "args": { "External id": 3125,"Record function id": 0, "Ev Idx": 3124 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118344.525, "dur": 0.111 , "args": { "External id": 3126,"Record function id": 0, "Ev Idx": 3125 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118345.117, "dur": 0.611 , "args": { "External id": 3127,"Record function id": 0, "Ev Idx": 3126 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118345.277, "dur": 0.361 , "args": { "External id": 3128,"Record function id": 0, "Ev Idx": 3127 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118345.467, "dur": 0.070 , "args": { "External id": 3129,"Record function id": 0, "Ev Idx": 3128 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118345.978, "dur": 0.631 , "args": { "External id": 3130,"Record function id": 0, "Ev Idx": 3129 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118346.138, "dur": 0.371 , "args": { "External id": 3131,"Record function id": 0, "Ev Idx": 3130 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118346.319, "dur": 0.090 , "args": { "External id": 3132,"Record function id": 0, "Ev Idx": 3131 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118347.731, "dur": 1.283 , "args": { "External id": 3133,"Record function id": 0, "Ev Idx": 3132 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118347.922, "dur": 0.992 , "args": { "External id": 3134,"Record function id": 0, "Ev Idx": 3133 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118357.850, "dur": 3.727 , "args": { "External id": 3135,"Record function id": 0, "Ev Idx": 3134 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118362.569, "dur": 0.601 , "args": { "External id": 3136,"Record function id": 0, "Ev Idx": 3135 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118364.553, "dur": 0.551 , "args": { "External id": 3137,"Record function id": 0, "Ev Idx": 3136 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381118368.721, "dur": 0.350 , "args": { "External id": 3138,"Record function id": 0, "Ev Idx": 3137 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381118368.931, "dur": 0.040 , "args": { "External id": 3139,"Record function id": 0, "Ev Idx": 3138 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381118371.556, "dur": 5.370 , "args": { "External id": 3140,"Record function id": 0, "Ev Idx": 3139 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118379.070, "dur": 10.239 , "args": { "External id": 3141,"Record function id": 0, "Ev Idx": 3140 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118379.611, "dur": 0.701 , "args": { "External id": 3142,"Record function id": 0, "Ev Idx": 3141 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118379.791, "dur": 0.421 , "args": { "External id": 3143,"Record function id": 0, "Ev Idx": 3142 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118380.022, "dur": 0.090 , "args": { "External id": 3144,"Record function id": 0, "Ev Idx": 3143 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118380.573, "dur": 0.621 , "args": { "External id": 3145,"Record function id": 0, "Ev Idx": 3144 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118380.733, "dur": 0.361 , "args": { "External id": 3146,"Record function id": 0, "Ev Idx": 3145 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118380.923, "dur": 0.070 , "args": { "External id": 3147,"Record function id": 0, "Ev Idx": 3146 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118381.434, "dur": 0.621 , "args": { "External id": 3148,"Record function id": 0, "Ev Idx": 3147 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118381.594, "dur": 0.371 , "args": { "External id": 3149,"Record function id": 0, "Ev Idx": 3148 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118381.775, "dur": 0.090 , "args": { "External id": 3150,"Record function id": 0, "Ev Idx": 3149 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118383.087, "dur": 1.273 , "args": { "External id": 3151,"Record function id": 0, "Ev Idx": 3150 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118383.268, "dur": 0.991 , "args": { "External id": 3152,"Record function id": 0, "Ev Idx": 3151 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381118393.988, "dur": 4.097 , "args": { "External id": 3153,"Record function id": 0, "Ev Idx": 3152 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381118403.525, "dur": 4.609 , "args": { "External id": 3154,"Record function id": 0, "Ev Idx": 3153 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381118413.193, "dur": 4.348 , "args": { "External id": 3155,"Record function id": 0, "Ev Idx": 3154 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381118419.746, "dur": 17.031 , "args": { "External id": 3156,"Record function id": 0, "Ev Idx": 3155 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118420.226, "dur": 0.501 , "args": { "External id": 3157,"Record function id": 0, "Ev Idx": 3156 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118420.467, "dur": 0.150 , "args": { "External id": 3158,"Record function id": 0, "Ev Idx": 3157 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118420.968, "dur": 0.411 , "args": { "External id": 3159,"Record function id": 0, "Ev Idx": 3158 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118421.178, "dur": 0.090 , "args": { "External id": 3160,"Record function id": 0, "Ev Idx": 3159 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118421.619, "dur": 0.391 , "args": { "External id": 3161,"Record function id": 0, "Ev Idx": 3160 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118421.809, "dur": 0.091 , "args": { "External id": 3162,"Record function id": 0, "Ev Idx": 3161 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381118422.340, "dur": 12.944 , "args": { "External id": 3163,"Record function id": 0, "Ev Idx": 3162 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381118423.102, "dur": 1.783 , "args": { "External id": 3164,"Record function id": 0, "Ev Idx": 3163 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118423.392, "dur": 1.393 , "args": { "External id": 3165,"Record function id": 0, "Ev Idx": 3164 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118425.366, "dur": 1.252 , "args": { "External id": 3166,"Record function id": 0, "Ev Idx": 3165 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118427.039, "dur": 1.292 , "args": { "External id": 3167,"Record function id": 0, "Ev Idx": 3166 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118428.612, "dur": 1.363 , "args": { "External id": 3168,"Record function id": 0, "Ev Idx": 3167 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118433.992, "dur": 0.932 , "args": { "External id": 3169,"Record function id": 0, "Ev Idx": 3168 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118435.565, "dur": 0.461 , "args": { "External id": 3170,"Record function id": 0, "Ev Idx": 3169 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118435.785, "dur": 0.131 , "args": { "External id": 3171,"Record function id": 0, "Ev Idx": 3170 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118443.790, "dur": 6.733 , "args": { "External id": 3172,"Record function id": 0, "Ev Idx": 3171 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118452.947, "dur": 10.550 , "args": { "External id": 3173,"Record function id": 0, "Ev Idx": 3172 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118453.558, "dur": 0.712 , "args": { "External id": 3174,"Record function id": 0, "Ev Idx": 3173 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118453.739, "dur": 0.441 , "args": { "External id": 3175,"Record function id": 0, "Ev Idx": 3174 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118453.969, "dur": 0.100 , "args": { "External id": 3176,"Record function id": 0, "Ev Idx": 3175 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118454.550, "dur": 0.641 , "args": { "External id": 3177,"Record function id": 0, "Ev Idx": 3176 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118454.711, "dur": 0.380 , "args": { "External id": 3178,"Record function id": 0, "Ev Idx": 3177 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118454.921, "dur": 0.070 , "args": { "External id": 3179,"Record function id": 0, "Ev Idx": 3178 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118455.442, "dur": 0.641 , "args": { "External id": 3180,"Record function id": 0, "Ev Idx": 3179 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118455.592, "dur": 0.391 , "args": { "External id": 3181,"Record function id": 0, "Ev Idx": 3180 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118455.803, "dur": 0.080 , "args": { "External id": 3182,"Record function id": 0, "Ev Idx": 3181 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118457.175, "dur": 1.253 , "args": { "External id": 3183,"Record function id": 0, "Ev Idx": 3182 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118457.356, "dur": 0.961 , "args": { "External id": 3184,"Record function id": 0, "Ev Idx": 3183 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381118467.514, "dur": 5.250 , "args": { "External id": 3185,"Record function id": 0, "Ev Idx": 3184 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118476.191, "dur": 5.229 , "args": { "External id": 3186,"Record function id": 0, "Ev Idx": 3185 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118484.797, "dur": 5.179 , "args": { "External id": 3187,"Record function id": 0, "Ev Idx": 3186 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118493.583, "dur": 10.259 , "args": { "External id": 3188,"Record function id": 0, "Ev Idx": 3187 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118494.114, "dur": 0.691 , "args": { "External id": 3189,"Record function id": 0, "Ev Idx": 3188 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118494.294, "dur": 0.421 , "args": { "External id": 3190,"Record function id": 0, "Ev Idx": 3189 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118494.525, "dur": 0.080 , "args": { "External id": 3191,"Record function id": 0, "Ev Idx": 3190 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118495.076, "dur": 0.631 , "args": { "External id": 3192,"Record function id": 0, "Ev Idx": 3191 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118495.236, "dur": 0.381 , "args": { "External id": 3193,"Record function id": 0, "Ev Idx": 3192 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118495.436, "dur": 0.081 , "args": { "External id": 3194,"Record function id": 0, "Ev Idx": 3193 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118495.967, "dur": 0.642 , "args": { "External id": 3195,"Record function id": 0, "Ev Idx": 3194 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118496.138, "dur": 0.380 , "args": { "External id": 3196,"Record function id": 0, "Ev Idx": 3195 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118496.328, "dur": 0.080 , "args": { "External id": 3197,"Record function id": 0, "Ev Idx": 3196 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118497.620, "dur": 1.263 , "args": { "External id": 3198,"Record function id": 0, "Ev Idx": 3197 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118497.811, "dur": 0.962 , "args": { "External id": 3199,"Record function id": 0, "Ev Idx": 3198 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118507.529, "dur": 10.069 , "args": { "External id": 3200,"Record function id": 0, "Ev Idx": 3199 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118508.030, "dur": 0.651 , "args": { "External id": 3201,"Record function id": 0, "Ev Idx": 3200 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118508.210, "dur": 0.381 , "args": { "External id": 3202,"Record function id": 0, "Ev Idx": 3201 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118508.411, "dur": 0.070 , "args": { "External id": 3203,"Record function id": 0, "Ev Idx": 3202 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118508.942, "dur": 0.641 , "args": { "External id": 3204,"Record function id": 0, "Ev Idx": 3203 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118509.112, "dur": 0.381 , "args": { "External id": 3205,"Record function id": 0, "Ev Idx": 3204 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118509.312, "dur": 0.070 , "args": { "External id": 3206,"Record function id": 0, "Ev Idx": 3205 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118509.833, "dur": 0.631 , "args": { "External id": 3207,"Record function id": 0, "Ev Idx": 3206 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118509.983, "dur": 0.391 , "args": { "External id": 3208,"Record function id": 0, "Ev Idx": 3207 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118510.194, "dur": 0.070 , "args": { "External id": 3209,"Record function id": 0, "Ev Idx": 3208 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118511.476, "dur": 1.233 , "args": { "External id": 3210,"Record function id": 0, "Ev Idx": 3209 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118511.657, "dur": 0.951 , "args": { "External id": 3211,"Record function id": 0, "Ev Idx": 3210 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381118522.367, "dur": 4.097 , "args": { "External id": 3212,"Record function id": 0, "Ev Idx": 3211 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118532.686, "dur": 5.239 , "args": { "External id": 3213,"Record function id": 0, "Ev Idx": 3212 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118541.272, "dur": 10.299 , "args": { "External id": 3214,"Record function id": 0, "Ev Idx": 3213 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118541.813, "dur": 0.701 , "args": { "External id": 3215,"Record function id": 0, "Ev Idx": 3214 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118541.993, "dur": 0.431 , "args": { "External id": 3216,"Record function id": 0, "Ev Idx": 3215 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118542.213, "dur": 0.111 , "args": { "External id": 3217,"Record function id": 0, "Ev Idx": 3216 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118542.795, "dur": 0.661 , "args": { "External id": 3218,"Record function id": 0, "Ev Idx": 3217 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118542.955, "dur": 0.411 , "args": { "External id": 3219,"Record function id": 0, "Ev Idx": 3218 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118543.185, "dur": 0.070 , "args": { "External id": 3220,"Record function id": 0, "Ev Idx": 3219 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118543.706, "dur": 0.641 , "args": { "External id": 3221,"Record function id": 0, "Ev Idx": 3220 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118543.887, "dur": 0.370 , "args": { "External id": 3222,"Record function id": 0, "Ev Idx": 3221 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118544.077, "dur": 0.080 , "args": { "External id": 3223,"Record function id": 0, "Ev Idx": 3222 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118545.399, "dur": 1.223 , "args": { "External id": 3224,"Record function id": 0, "Ev Idx": 3223 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118545.580, "dur": 0.931 , "args": { "External id": 3225,"Record function id": 0, "Ev Idx": 3224 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381118556.350, "dur": 4.799 , "args": { "External id": 3226,"Record function id": 0, "Ev Idx": 3225 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381118581.226, "dur": 23.384 , "args": { "External id": 3227,"Record function id": 0, "Ev Idx": 3226 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381118605.000, "dur": 335.515 , "args": { "External id": 3228,"Record function id": 0, "Ev Idx": 3227 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381118609.739, "dur": 10.710 , "args": { "External id": 3229,"Record function id": 0, "Ev Idx": 3228 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118613.196, "dur": 0.400 , "args": { "External id": 3230,"Record function id": 0, "Ev Idx": 3229 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381118614.087, "dur": 0.040 , "args": { "External id": 3231,"Record function id": 0, "Ev Idx": 3230 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381118614.488, "dur": 0.040 , "args": { "External id": 3232,"Record function id": 0, "Ev Idx": 3231 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118615.891, "dur": 0.310 , "args": { "External id": 3233,"Record function id": 0, "Ev Idx": 3232 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381118616.562, "dur": 0.030 , "args": { "External id": 3234,"Record function id": 0, "Ev Idx": 3233 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381118616.903, "dur": 0.030 , "args": { "External id": 3235,"Record function id": 0, "Ev Idx": 3234 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381118623.395, "dur": 1.102 , "args": { "External id": 3236,"Record function id": 0, "Ev Idx": 3235 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381118626.671, "dur": 310.918 , "args": { "External id": 3237,"Record function id": 0, "Ev Idx": 3236 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118635.297, "dur": 6.221 , "args": { "External id": 3238,"Record function id": 0, "Ev Idx": 3237 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118647.990, "dur": 6.993 , "args": { "External id": 3239,"Record function id": 0, "Ev Idx": 3238 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118660.814, "dur": 5.521 , "args": { "External id": 3240,"Record function id": 0, "Ev Idx": 3239 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118671.735, "dur": 4.869 , "args": { "External id": 3241,"Record function id": 0, "Ev Idx": 3240 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118679.178, "dur": 12.203 , "args": { "External id": 3242,"Record function id": 0, "Ev Idx": 3241 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118679.810, "dur": 0.931 , "args": { "External id": 3243,"Record function id": 0, "Ev Idx": 3242 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118680.040, "dur": 0.601 , "args": { "External id": 3244,"Record function id": 0, "Ev Idx": 3243 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118680.351, "dur": 0.140 , "args": { "External id": 3245,"Record function id": 0, "Ev Idx": 3244 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118681.042, "dur": 0.631 , "args": { "External id": 3246,"Record function id": 0, "Ev Idx": 3245 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118681.212, "dur": 0.371 , "args": { "External id": 3247,"Record function id": 0, "Ev Idx": 3246 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118681.403, "dur": 0.070 , "args": { "External id": 3248,"Record function id": 0, "Ev Idx": 3247 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118681.944, "dur": 0.721 , "args": { "External id": 3249,"Record function id": 0, "Ev Idx": 3248 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118682.104, "dur": 0.471 , "args": { "External id": 3250,"Record function id": 0, "Ev Idx": 3249 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118682.304, "dur": 0.161 , "args": { "External id": 3251,"Record function id": 0, "Ev Idx": 3250 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118683.937, "dur": 1.413 , "args": { "External id": 3252,"Record function id": 0, "Ev Idx": 3251 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118684.148, "dur": 1.082 , "args": { "External id": 3253,"Record function id": 0, "Ev Idx": 3252 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118696.671, "dur": 4.398 , "args": { "External id": 3254,"Record function id": 0, "Ev Idx": 3253 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118703.394, "dur": 0.921 , "args": { "External id": 3255,"Record function id": 0, "Ev Idx": 3254 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118706.169, "dur": 0.561 , "args": { "External id": 3256,"Record function id": 0, "Ev Idx": 3255 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381118710.346, "dur": 0.401 , "args": { "External id": 3257,"Record function id": 0, "Ev Idx": 3256 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381118710.587, "dur": 0.050 , "args": { "External id": 3258,"Record function id": 0, "Ev Idx": 3257 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381118713.823, "dur": 6.021 , "args": { "External id": 3259,"Record function id": 0, "Ev Idx": 3258 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118722.088, "dur": 10.710 , "args": { "External id": 3260,"Record function id": 0, "Ev Idx": 3259 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118722.619, "dur": 0.752 , "args": { "External id": 3261,"Record function id": 0, "Ev Idx": 3260 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118722.810, "dur": 0.471 , "args": { "External id": 3262,"Record function id": 0, "Ev Idx": 3261 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118723.060, "dur": 0.110 , "args": { "External id": 3263,"Record function id": 0, "Ev Idx": 3262 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118723.671, "dur": 0.621 , "args": { "External id": 3264,"Record function id": 0, "Ev Idx": 3263 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118723.832, "dur": 0.370 , "args": { "External id": 3265,"Record function id": 0, "Ev Idx": 3264 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118724.022, "dur": 0.080 , "args": { "External id": 3266,"Record function id": 0, "Ev Idx": 3265 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118724.553, "dur": 0.611 , "args": { "External id": 3267,"Record function id": 0, "Ev Idx": 3266 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118724.713, "dur": 0.361 , "args": { "External id": 3268,"Record function id": 0, "Ev Idx": 3267 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118724.894, "dur": 0.080 , "args": { "External id": 3269,"Record function id": 0, "Ev Idx": 3268 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118726.306, "dur": 1.363 , "args": { "External id": 3270,"Record function id": 0, "Ev Idx": 3269 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118726.527, "dur": 1.031 , "args": { "External id": 3271,"Record function id": 0, "Ev Idx": 3270 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118736.054, "dur": 4.128 , "args": { "External id": 3272,"Record function id": 0, "Ev Idx": 3271 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118741.274, "dur": 0.621 , "args": { "External id": 3273,"Record function id": 0, "Ev Idx": 3272 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381118743.218, "dur": 0.591 , "args": { "External id": 3274,"Record function id": 0, "Ev Idx": 3273 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381118747.536, "dur": 0.340 , "args": { "External id": 3275,"Record function id": 0, "Ev Idx": 3274 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381118747.726, "dur": 0.040 , "args": { "External id": 3276,"Record function id": 0, "Ev Idx": 3275 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381118750.481, "dur": 5.901 , "args": { "External id": 3277,"Record function id": 0, "Ev Idx": 3276 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118758.536, "dur": 10.470 , "args": { "External id": 3278,"Record function id": 0, "Ev Idx": 3277 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118759.067, "dur": 0.792 , "args": { "External id": 3279,"Record function id": 0, "Ev Idx": 3278 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118759.298, "dur": 0.460 , "args": { "External id": 3280,"Record function id": 0, "Ev Idx": 3279 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118759.538, "dur": 0.100 , "args": { "External id": 3281,"Record function id": 0, "Ev Idx": 3280 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118760.129, "dur": 0.621 , "args": { "External id": 3282,"Record function id": 0, "Ev Idx": 3281 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118760.289, "dur": 0.371 , "args": { "External id": 3283,"Record function id": 0, "Ev Idx": 3282 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118760.480, "dur": 0.070 , "args": { "External id": 3284,"Record function id": 0, "Ev Idx": 3283 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118761.001, "dur": 0.711 , "args": { "External id": 3285,"Record function id": 0, "Ev Idx": 3284 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118761.171, "dur": 0.451 , "args": { "External id": 3286,"Record function id": 0, "Ev Idx": 3285 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118761.361, "dur": 0.151 , "args": { "External id": 3287,"Record function id": 0, "Ev Idx": 3286 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118762.764, "dur": 1.312 , "args": { "External id": 3288,"Record function id": 0, "Ev Idx": 3287 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118762.954, "dur": 1.022 , "args": { "External id": 3289,"Record function id": 0, "Ev Idx": 3288 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381118773.484, "dur": 4.238 , "args": { "External id": 3290,"Record function id": 0, "Ev Idx": 3289 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381118782.972, "dur": 4.769 , "args": { "External id": 3291,"Record function id": 0, "Ev Idx": 3290 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381118792.690, "dur": 4.398 , "args": { "External id": 3292,"Record function id": 0, "Ev Idx": 3291 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381118799.292, "dur": 17.222 , "args": { "External id": 3293,"Record function id": 0, "Ev Idx": 3292 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118799.773, "dur": 0.491 , "args": { "External id": 3294,"Record function id": 0, "Ev Idx": 3293 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118800.003, "dur": 0.151 , "args": { "External id": 3295,"Record function id": 0, "Ev Idx": 3294 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118800.504, "dur": 0.391 , "args": { "External id": 3296,"Record function id": 0, "Ev Idx": 3295 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118800.705, "dur": 0.080 , "args": { "External id": 3297,"Record function id": 0, "Ev Idx": 3296 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118801.145, "dur": 0.381 , "args": { "External id": 3298,"Record function id": 0, "Ev Idx": 3297 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118801.336, "dur": 0.080 , "args": { "External id": 3299,"Record function id": 0, "Ev Idx": 3298 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381118801.877, "dur": 13.124 , "args": { "External id": 3300,"Record function id": 0, "Ev Idx": 3299 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381118802.748, "dur": 1.994 , "args": { "External id": 3301,"Record function id": 0, "Ev Idx": 3300 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118803.139, "dur": 1.483 , "args": { "External id": 3302,"Record function id": 0, "Ev Idx": 3301 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118805.173, "dur": 1.232 , "args": { "External id": 3303,"Record function id": 0, "Ev Idx": 3302 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118806.826, "dur": 1.192 , "args": { "External id": 3304,"Record function id": 0, "Ev Idx": 3303 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118808.289, "dur": 1.282 , "args": { "External id": 3305,"Record function id": 0, "Ev Idx": 3304 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118813.709, "dur": 0.922 , "args": { "External id": 3306,"Record function id": 0, "Ev Idx": 3305 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118815.312, "dur": 0.471 , "args": { "External id": 3307,"Record function id": 0, "Ev Idx": 3306 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118815.522, "dur": 0.151 , "args": { "External id": 3308,"Record function id": 0, "Ev Idx": 3307 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118823.267, "dur": 6.532 , "args": { "External id": 3309,"Record function id": 0, "Ev Idx": 3308 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118832.213, "dur": 10.881 , "args": { "External id": 3310,"Record function id": 0, "Ev Idx": 3309 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118832.804, "dur": 0.742 , "args": { "External id": 3311,"Record function id": 0, "Ev Idx": 3310 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118832.985, "dur": 0.461 , "args": { "External id": 3312,"Record function id": 0, "Ev Idx": 3311 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118833.245, "dur": 0.080 , "args": { "External id": 3313,"Record function id": 0, "Ev Idx": 3312 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118833.826, "dur": 0.672 , "args": { "External id": 3314,"Record function id": 0, "Ev Idx": 3313 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118834.007, "dur": 0.390 , "args": { "External id": 3315,"Record function id": 0, "Ev Idx": 3314 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118834.227, "dur": 0.070 , "args": { "External id": 3316,"Record function id": 0, "Ev Idx": 3315 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118834.758, "dur": 0.661 , "args": { "External id": 3317,"Record function id": 0, "Ev Idx": 3316 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118834.918, "dur": 0.411 , "args": { "External id": 3318,"Record function id": 0, "Ev Idx": 3317 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118835.149, "dur": 0.080 , "args": { "External id": 3319,"Record function id": 0, "Ev Idx": 3318 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118836.621, "dur": 1.273 , "args": { "External id": 3320,"Record function id": 0, "Ev Idx": 3319 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118836.832, "dur": 0.962 , "args": { "External id": 3321,"Record function id": 0, "Ev Idx": 3320 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381118847.161, "dur": 5.330 , "args": { "External id": 3322,"Record function id": 0, "Ev Idx": 3321 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118855.246, "dur": 5.160 , "args": { "External id": 3323,"Record function id": 0, "Ev Idx": 3322 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381118863.622, "dur": 4.829 , "args": { "External id": 3324,"Record function id": 0, "Ev Idx": 3323 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118871.957, "dur": 10.239 , "args": { "External id": 3325,"Record function id": 0, "Ev Idx": 3324 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118872.488, "dur": 0.672 , "args": { "External id": 3326,"Record function id": 0, "Ev Idx": 3325 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118872.669, "dur": 0.400 , "args": { "External id": 3327,"Record function id": 0, "Ev Idx": 3326 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118872.879, "dur": 0.080 , "args": { "External id": 3328,"Record function id": 0, "Ev Idx": 3327 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118873.460, "dur": 0.651 , "args": { "External id": 3329,"Record function id": 0, "Ev Idx": 3328 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118873.620, "dur": 0.391 , "args": { "External id": 3330,"Record function id": 0, "Ev Idx": 3329 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118873.831, "dur": 0.080 , "args": { "External id": 3331,"Record function id": 0, "Ev Idx": 3330 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118874.362, "dur": 0.641 , "args": { "External id": 3332,"Record function id": 0, "Ev Idx": 3331 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118874.522, "dur": 0.391 , "args": { "External id": 3333,"Record function id": 0, "Ev Idx": 3332 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118874.722, "dur": 0.091 , "args": { "External id": 3334,"Record function id": 0, "Ev Idx": 3333 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118876.055, "dur": 1.272 , "args": { "External id": 3335,"Record function id": 0, "Ev Idx": 3334 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118876.235, "dur": 0.992 , "args": { "External id": 3336,"Record function id": 0, "Ev Idx": 3335 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118885.763, "dur": 9.948 , "args": { "External id": 3337,"Record function id": 0, "Ev Idx": 3336 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118886.254, "dur": 0.661 , "args": { "External id": 3338,"Record function id": 0, "Ev Idx": 3337 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118886.434, "dur": 0.381 , "args": { "External id": 3339,"Record function id": 0, "Ev Idx": 3338 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118886.635, "dur": 0.080 , "args": { "External id": 3340,"Record function id": 0, "Ev Idx": 3339 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118887.166, "dur": 0.621 , "args": { "External id": 3341,"Record function id": 0, "Ev Idx": 3340 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118887.326, "dur": 0.371 , "args": { "External id": 3342,"Record function id": 0, "Ev Idx": 3341 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118887.526, "dur": 0.070 , "args": { "External id": 3343,"Record function id": 0, "Ev Idx": 3342 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118888.047, "dur": 0.641 , "args": { "External id": 3344,"Record function id": 0, "Ev Idx": 3343 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118888.208, "dur": 0.390 , "args": { "External id": 3345,"Record function id": 0, "Ev Idx": 3344 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118888.408, "dur": 0.080 , "args": { "External id": 3346,"Record function id": 0, "Ev Idx": 3345 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118889.690, "dur": 1.223 , "args": { "External id": 3347,"Record function id": 0, "Ev Idx": 3346 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118889.861, "dur": 0.951 , "args": { "External id": 3348,"Record function id": 0, "Ev Idx": 3347 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381118900.440, "dur": 4.118 , "args": { "External id": 3349,"Record function id": 0, "Ev Idx": 3348 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381118907.884, "dur": 5.270 , "args": { "External id": 3350,"Record function id": 0, "Ev Idx": 3349 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381118916.380, "dur": 10.179 , "args": { "External id": 3351,"Record function id": 0, "Ev Idx": 3350 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118916.901, "dur": 0.681 , "args": { "External id": 3352,"Record function id": 0, "Ev Idx": 3351 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118917.081, "dur": 0.411 , "args": { "External id": 3353,"Record function id": 0, "Ev Idx": 3352 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118917.302, "dur": 0.080 , "args": { "External id": 3354,"Record function id": 0, "Ev Idx": 3353 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118917.883, "dur": 0.661 , "args": { "External id": 3355,"Record function id": 0, "Ev Idx": 3354 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118918.043, "dur": 0.411 , "args": { "External id": 3356,"Record function id": 0, "Ev Idx": 3355 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118918.273, "dur": 0.081 , "args": { "External id": 3357,"Record function id": 0, "Ev Idx": 3356 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381118918.794, "dur": 0.632 , "args": { "External id": 3358,"Record function id": 0, "Ev Idx": 3357 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381118918.955, "dur": 0.380 , "args": { "External id": 3359,"Record function id": 0, "Ev Idx": 3358 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381118919.155, "dur": 0.070 , "args": { "External id": 3360,"Record function id": 0, "Ev Idx": 3359 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118920.478, "dur": 1.262 , "args": { "External id": 3361,"Record function id": 0, "Ev Idx": 3360 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118920.658, "dur": 0.982 , "args": { "External id": 3362,"Record function id": 0, "Ev Idx": 3361 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381118931.238, "dur": 4.809 , "args": { "External id": 3363,"Record function id": 0, "Ev Idx": 3362 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381118955.092, "dur": 22.492 , "args": { "External id": 3364,"Record function id": 0, "Ev Idx": 3363 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381118977.985, "dur": 334.713 , "args": { "External id": 3365,"Record function id": 0, "Ev Idx": 3364 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381118982.814, "dur": 10.159 , "args": { "External id": 3366,"Record function id": 0, "Ev Idx": 3365 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118986.030, "dur": 0.340 , "args": { "External id": 3367,"Record function id": 0, "Ev Idx": 3366 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381118986.831, "dur": 0.030 , "args": { "External id": 3368,"Record function id": 0, "Ev Idx": 3367 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381118987.232, "dur": 0.040 , "args": { "External id": 3369,"Record function id": 0, "Ev Idx": 3368 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381118988.474, "dur": 0.231 , "args": { "External id": 3370,"Record function id": 0, "Ev Idx": 3369 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381118989.045, "dur": 0.020 , "args": { "External id": 3371,"Record function id": 0, "Ev Idx": 3370 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381118989.386, "dur": 0.020 , "args": { "External id": 3372,"Record function id": 0, "Ev Idx": 3371 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381118995.748, "dur": 1.112 , "args": { "External id": 3373,"Record function id": 0, "Ev Idx": 3372 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381118998.924, "dur": 310.768 , "args": { "External id": 3374,"Record function id": 0, "Ev Idx": 3373 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119007.249, "dur": 6.332 , "args": { "External id": 3375,"Record function id": 0, "Ev Idx": 3374 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119019.993, "dur": 7.023 , "args": { "External id": 3376,"Record function id": 0, "Ev Idx": 3375 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119032.827, "dur": 5.209 , "args": { "External id": 3377,"Record function id": 0, "Ev Idx": 3376 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119043.537, "dur": 5.129 , "args": { "External id": 3378,"Record function id": 0, "Ev Idx": 3377 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119051.121, "dur": 11.992 , "args": { "External id": 3379,"Record function id": 0, "Ev Idx": 3378 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119051.772, "dur": 0.962 , "args": { "External id": 3380,"Record function id": 0, "Ev Idx": 3379 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119051.972, "dur": 0.652 , "args": { "External id": 3381,"Record function id": 0, "Ev Idx": 3380 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119052.293, "dur": 0.200 , "args": { "External id": 3382,"Record function id": 0, "Ev Idx": 3381 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119053.054, "dur": 0.682 , "args": { "External id": 3383,"Record function id": 0, "Ev Idx": 3382 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119053.265, "dur": 0.380 , "args": { "External id": 3384,"Record function id": 0, "Ev Idx": 3383 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119053.465, "dur": 0.080 , "args": { "External id": 3385,"Record function id": 0, "Ev Idx": 3384 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119053.996, "dur": 0.651 , "args": { "External id": 3386,"Record function id": 0, "Ev Idx": 3385 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119054.146, "dur": 0.411 , "args": { "External id": 3387,"Record function id": 0, "Ev Idx": 3386 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119054.347, "dur": 0.110 , "args": { "External id": 3388,"Record function id": 0, "Ev Idx": 3387 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119055.940, "dur": 1.372 , "args": { "External id": 3389,"Record function id": 0, "Ev Idx": 3388 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119056.150, "dur": 1.042 , "args": { "External id": 3390,"Record function id": 0, "Ev Idx": 3389 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119068.563, "dur": 4.368 , "args": { "External id": 3391,"Record function id": 0, "Ev Idx": 3390 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119075.145, "dur": 0.872 , "args": { "External id": 3392,"Record function id": 0, "Ev Idx": 3391 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119077.790, "dur": 0.571 , "args": { "External id": 3393,"Record function id": 0, "Ev Idx": 3392 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381119082.028, "dur": 0.431 , "args": { "External id": 3394,"Record function id": 0, "Ev Idx": 3393 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381119082.279, "dur": 0.080 , "args": { "External id": 3395,"Record function id": 0, "Ev Idx": 3394 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381119085.334, "dur": 6.362 , "args": { "External id": 3396,"Record function id": 0, "Ev Idx": 3395 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119093.830, "dur": 10.640 , "args": { "External id": 3397,"Record function id": 0, "Ev Idx": 3396 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119094.371, "dur": 0.802 , "args": { "External id": 3398,"Record function id": 0, "Ev Idx": 3397 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119094.542, "dur": 0.541 , "args": { "External id": 3399,"Record function id": 0, "Ev Idx": 3398 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119094.832, "dur": 0.130 , "args": { "External id": 3400,"Record function id": 0, "Ev Idx": 3399 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119095.443, "dur": 0.621 , "args": { "External id": 3401,"Record function id": 0, "Ev Idx": 3400 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119095.604, "dur": 0.370 , "args": { "External id": 3402,"Record function id": 0, "Ev Idx": 3401 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119095.794, "dur": 0.070 , "args": { "External id": 3403,"Record function id": 0, "Ev Idx": 3402 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119096.315, "dur": 0.701 , "args": { "External id": 3404,"Record function id": 0, "Ev Idx": 3403 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119096.475, "dur": 0.451 , "args": { "External id": 3405,"Record function id": 0, "Ev Idx": 3404 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119096.666, "dur": 0.150 , "args": { "External id": 3406,"Record function id": 0, "Ev Idx": 3405 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119098.108, "dur": 1.323 , "args": { "External id": 3407,"Record function id": 0, "Ev Idx": 3406 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119098.309, "dur": 1.011 , "args": { "External id": 3408,"Record function id": 0, "Ev Idx": 3407 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119107.776, "dur": 3.968 , "args": { "External id": 3409,"Record function id": 0, "Ev Idx": 3408 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119112.745, "dur": 0.592 , "args": { "External id": 3410,"Record function id": 0, "Ev Idx": 3409 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119114.809, "dur": 0.551 , "args": { "External id": 3411,"Record function id": 0, "Ev Idx": 3410 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381119119.328, "dur": 0.320 , "args": { "External id": 3412,"Record function id": 0, "Ev Idx": 3411 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381119119.508, "dur": 0.040 , "args": { "External id": 3413,"Record function id": 0, "Ev Idx": 3412 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381119122.173, "dur": 5.570 , "args": { "External id": 3414,"Record function id": 0, "Ev Idx": 3413 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119129.897, "dur": 10.500 , "args": { "External id": 3415,"Record function id": 0, "Ev Idx": 3414 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119130.438, "dur": 0.762 , "args": { "External id": 3416,"Record function id": 0, "Ev Idx": 3415 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119130.629, "dur": 0.481 , "args": { "External id": 3417,"Record function id": 0, "Ev Idx": 3416 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119130.859, "dur": 0.130 , "args": { "External id": 3418,"Record function id": 0, "Ev Idx": 3417 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119131.480, "dur": 0.642 , "args": { "External id": 3419,"Record function id": 0, "Ev Idx": 3418 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119131.651, "dur": 0.380 , "args": { "External id": 3420,"Record function id": 0, "Ev Idx": 3419 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119131.841, "dur": 0.090 , "args": { "External id": 3421,"Record function id": 0, "Ev Idx": 3420 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119132.402, "dur": 0.621 , "args": { "External id": 3422,"Record function id": 0, "Ev Idx": 3421 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119132.562, "dur": 0.371 , "args": { "External id": 3423,"Record function id": 0, "Ev Idx": 3422 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119132.753, "dur": 0.080 , "args": { "External id": 3424,"Record function id": 0, "Ev Idx": 3423 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119134.095, "dur": 1.243 , "args": { "External id": 3425,"Record function id": 0, "Ev Idx": 3424 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119134.266, "dur": 0.961 , "args": { "External id": 3426,"Record function id": 0, "Ev Idx": 3425 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381119144.865, "dur": 4.308 , "args": { "External id": 3427,"Record function id": 0, "Ev Idx": 3426 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381119154.333, "dur": 4.849 , "args": { "External id": 3428,"Record function id": 0, "Ev Idx": 3427 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381119164.001, "dur": 4.238 , "args": { "External id": 3429,"Record function id": 0, "Ev Idx": 3428 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381119170.373, "dur": 17.312 , "args": { "External id": 3430,"Record function id": 0, "Ev Idx": 3429 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119170.824, "dur": 0.621 , "args": { "External id": 3431,"Record function id": 0, "Ev Idx": 3430 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119171.134, "dur": 0.191 , "args": { "External id": 3432,"Record function id": 0, "Ev Idx": 3431 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119171.695, "dur": 0.401 , "args": { "External id": 3433,"Record function id": 0, "Ev Idx": 3432 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119171.906, "dur": 0.090 , "args": { "External id": 3434,"Record function id": 0, "Ev Idx": 3433 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119172.346, "dur": 0.461 , "args": { "External id": 3435,"Record function id": 0, "Ev Idx": 3434 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119172.617, "dur": 0.090 , "args": { "External id": 3436,"Record function id": 0, "Ev Idx": 3435 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381119173.168, "dur": 13.004 , "args": { "External id": 3437,"Record function id": 0, "Ev Idx": 3436 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381119174.000, "dur": 1.843 , "args": { "External id": 3438,"Record function id": 0, "Ev Idx": 3437 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119174.350, "dur": 1.393 , "args": { "External id": 3439,"Record function id": 0, "Ev Idx": 3438 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119176.304, "dur": 1.262 , "args": { "External id": 3440,"Record function id": 0, "Ev Idx": 3439 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119178.007, "dur": 1.182 , "args": { "External id": 3441,"Record function id": 0, "Ev Idx": 3440 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119179.470, "dur": 1.232 , "args": { "External id": 3442,"Record function id": 0, "Ev Idx": 3441 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119184.880, "dur": 0.931 , "args": { "External id": 3443,"Record function id": 0, "Ev Idx": 3442 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119186.473, "dur": 0.491 , "args": { "External id": 3444,"Record function id": 0, "Ev Idx": 3443 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119186.683, "dur": 0.170 , "args": { "External id": 3445,"Record function id": 0, "Ev Idx": 3444 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119194.658, "dur": 6.562 , "args": { "External id": 3446,"Record function id": 0, "Ev Idx": 3445 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119203.595, "dur": 10.599 , "args": { "External id": 3447,"Record function id": 0, "Ev Idx": 3446 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119204.176, "dur": 0.681 , "args": { "External id": 3448,"Record function id": 0, "Ev Idx": 3447 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119204.346, "dur": 0.411 , "args": { "External id": 3449,"Record function id": 0, "Ev Idx": 3448 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119204.576, "dur": 0.071 , "args": { "External id": 3450,"Record function id": 0, "Ev Idx": 3449 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119205.137, "dur": 0.682 , "args": { "External id": 3451,"Record function id": 0, "Ev Idx": 3450 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119205.298, "dur": 0.431 , "args": { "External id": 3452,"Record function id": 0, "Ev Idx": 3451 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119205.548, "dur": 0.070 , "args": { "External id": 3453,"Record function id": 0, "Ev Idx": 3452 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119206.069, "dur": 0.641 , "args": { "External id": 3454,"Record function id": 0, "Ev Idx": 3453 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119206.230, "dur": 0.390 , "args": { "External id": 3455,"Record function id": 0, "Ev Idx": 3454 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119206.430, "dur": 0.080 , "args": { "External id": 3456,"Record function id": 0, "Ev Idx": 3455 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119207.792, "dur": 1.263 , "args": { "External id": 3457,"Record function id": 0, "Ev Idx": 3456 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119207.973, "dur": 0.982 , "args": { "External id": 3458,"Record function id": 0, "Ev Idx": 3457 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381119218.212, "dur": 5.240 , "args": { "External id": 3459,"Record function id": 0, "Ev Idx": 3458 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119226.337, "dur": 5.590 , "args": { "External id": 3460,"Record function id": 0, "Ev Idx": 3459 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119235.314, "dur": 5.079 , "args": { "External id": 3461,"Record function id": 0, "Ev Idx": 3460 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119243.739, "dur": 10.540 , "args": { "External id": 3462,"Record function id": 0, "Ev Idx": 3461 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119244.260, "dur": 0.772 , "args": { "External id": 3463,"Record function id": 0, "Ev Idx": 3462 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119244.441, "dur": 0.491 , "args": { "External id": 3464,"Record function id": 0, "Ev Idx": 3463 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119244.681, "dur": 0.140 , "args": { "External id": 3465,"Record function id": 0, "Ev Idx": 3464 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119245.312, "dur": 0.682 , "args": { "External id": 3466,"Record function id": 0, "Ev Idx": 3465 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119245.483, "dur": 0.420 , "args": { "External id": 3467,"Record function id": 0, "Ev Idx": 3466 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119245.723, "dur": 0.070 , "args": { "External id": 3468,"Record function id": 0, "Ev Idx": 3467 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119246.254, "dur": 0.661 , "args": { "External id": 3469,"Record function id": 0, "Ev Idx": 3468 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119246.414, "dur": 0.411 , "args": { "External id": 3470,"Record function id": 0, "Ev Idx": 3469 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119246.655, "dur": 0.070 , "args": { "External id": 3471,"Record function id": 0, "Ev Idx": 3470 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119247.957, "dur": 1.283 , "args": { "External id": 3472,"Record function id": 0, "Ev Idx": 3471 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119248.148, "dur": 0.991 , "args": { "External id": 3473,"Record function id": 0, "Ev Idx": 3472 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119258.006, "dur": 9.928 , "args": { "External id": 3474,"Record function id": 0, "Ev Idx": 3473 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119258.497, "dur": 0.641 , "args": { "External id": 3475,"Record function id": 0, "Ev Idx": 3474 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119258.657, "dur": 0.391 , "args": { "External id": 3476,"Record function id": 0, "Ev Idx": 3475 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119258.857, "dur": 0.081 , "args": { "External id": 3477,"Record function id": 0, "Ev Idx": 3476 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119259.398, "dur": 0.632 , "args": { "External id": 3478,"Record function id": 0, "Ev Idx": 3477 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119259.569, "dur": 0.370 , "args": { "External id": 3479,"Record function id": 0, "Ev Idx": 3478 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119259.759, "dur": 0.070 , "args": { "External id": 3480,"Record function id": 0, "Ev Idx": 3479 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119260.280, "dur": 0.661 , "args": { "External id": 3481,"Record function id": 0, "Ev Idx": 3480 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119260.440, "dur": 0.411 , "args": { "External id": 3482,"Record function id": 0, "Ev Idx": 3481 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119260.661, "dur": 0.090 , "args": { "External id": 3483,"Record function id": 0, "Ev Idx": 3482 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119261.933, "dur": 1.253 , "args": { "External id": 3484,"Record function id": 0, "Ev Idx": 3483 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119262.114, "dur": 0.971 , "args": { "External id": 3485,"Record function id": 0, "Ev Idx": 3484 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381119272.473, "dur": 4.328 , "args": { "External id": 3486,"Record function id": 0, "Ev Idx": 3485 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119280.057, "dur": 5.320 , "args": { "External id": 3487,"Record function id": 0, "Ev Idx": 3486 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119288.573, "dur": 10.159 , "args": { "External id": 3488,"Record function id": 0, "Ev Idx": 3487 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119289.104, "dur": 0.661 , "args": { "External id": 3489,"Record function id": 0, "Ev Idx": 3488 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119289.264, "dur": 0.411 , "args": { "External id": 3490,"Record function id": 0, "Ev Idx": 3489 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119289.484, "dur": 0.081 , "args": { "External id": 3491,"Record function id": 0, "Ev Idx": 3490 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119290.056, "dur": 0.651 , "args": { "External id": 3492,"Record function id": 0, "Ev Idx": 3491 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119290.216, "dur": 0.401 , "args": { "External id": 3493,"Record function id": 0, "Ev Idx": 3492 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119290.436, "dur": 0.070 , "args": { "External id": 3494,"Record function id": 0, "Ev Idx": 3493 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119290.957, "dur": 0.631 , "args": { "External id": 3495,"Record function id": 0, "Ev Idx": 3494 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119291.118, "dur": 0.380 , "args": { "External id": 3496,"Record function id": 0, "Ev Idx": 3495 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119291.318, "dur": 0.080 , "args": { "External id": 3497,"Record function id": 0, "Ev Idx": 3496 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119292.610, "dur": 1.243 , "args": { "External id": 3498,"Record function id": 0, "Ev Idx": 3497 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119292.791, "dur": 0.951 , "args": { "External id": 3499,"Record function id": 0, "Ev Idx": 3498 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381119303.440, "dur": 4.729 , "args": { "External id": 3500,"Record function id": 0, "Ev Idx": 3499 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381119327.014, "dur": 24.396 , "args": { "External id": 3501,"Record function id": 0, "Ev Idx": 3500 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381119351.800, "dur": 346.926 , "args": { "External id": 3502,"Record function id": 0, "Ev Idx": 3501 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381119356.579, "dur": 10.139 , "args": { "External id": 3503,"Record function id": 0, "Ev Idx": 3502 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119359.745, "dur": 0.411 , "args": { "External id": 3504,"Record function id": 0, "Ev Idx": 3503 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381119360.597, "dur": 0.040 , "args": { "External id": 3505,"Record function id": 0, "Ev Idx": 3504 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381119361.018, "dur": 0.050 , "args": { "External id": 3506,"Record function id": 0, "Ev Idx": 3505 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119362.200, "dur": 0.210 , "args": { "External id": 3507,"Record function id": 0, "Ev Idx": 3506 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381119362.751, "dur": 0.050 , "args": { "External id": 3508,"Record function id": 0, "Ev Idx": 3507 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381119363.152, "dur": 0.030 , "args": { "External id": 3509,"Record function id": 0, "Ev Idx": 3508 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381119369.523, "dur": 1.183 , "args": { "External id": 3510,"Record function id": 0, "Ev Idx": 3509 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381119372.780, "dur": 322.911 , "args": { "External id": 3511,"Record function id": 0, "Ev Idx": 3510 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119381.325, "dur": 6.102 , "args": { "External id": 3512,"Record function id": 0, "Ev Idx": 3511 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119393.688, "dur": 6.803 , "args": { "External id": 3513,"Record function id": 0, "Ev Idx": 3512 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119406.382, "dur": 5.250 , "args": { "External id": 3514,"Record function id": 0, "Ev Idx": 3513 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119417.342, "dur": 5.040 , "args": { "External id": 3515,"Record function id": 0, "Ev Idx": 3514 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119424.806, "dur": 12.313 , "args": { "External id": 3516,"Record function id": 0, "Ev Idx": 3515 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119425.407, "dur": 0.922 , "args": { "External id": 3517,"Record function id": 0, "Ev Idx": 3516 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119425.608, "dur": 0.621 , "args": { "External id": 3518,"Record function id": 0, "Ev Idx": 3517 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119425.898, "dur": 0.191 , "args": { "External id": 3519,"Record function id": 0, "Ev Idx": 3518 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119426.650, "dur": 0.611 , "args": { "External id": 3520,"Record function id": 0, "Ev Idx": 3519 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119426.800, "dur": 0.371 , "args": { "External id": 3521,"Record function id": 0, "Ev Idx": 3520 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119427.000, "dur": 0.071 , "args": { "External id": 3522,"Record function id": 0, "Ev Idx": 3521 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119427.541, "dur": 0.622 , "args": { "External id": 3523,"Record function id": 0, "Ev Idx": 3522 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119427.692, "dur": 0.380 , "args": { "External id": 3524,"Record function id": 0, "Ev Idx": 3523 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119427.882, "dur": 0.080 , "args": { "External id": 3525,"Record function id": 0, "Ev Idx": 3524 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119429.405, "dur": 1.403 , "args": { "External id": 3526,"Record function id": 0, "Ev Idx": 3525 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119429.625, "dur": 1.082 , "args": { "External id": 3527,"Record function id": 0, "Ev Idx": 3526 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119442.650, "dur": 4.548 , "args": { "External id": 3528,"Record function id": 0, "Ev Idx": 3527 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119449.372, "dur": 0.872 , "args": { "External id": 3529,"Record function id": 0, "Ev Idx": 3528 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119452.027, "dur": 0.581 , "args": { "External id": 3530,"Record function id": 0, "Ev Idx": 3529 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381119456.345, "dur": 0.451 , "args": { "External id": 3531,"Record function id": 0, "Ev Idx": 3530 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381119456.646, "dur": 0.050 , "args": { "External id": 3532,"Record function id": 0, "Ev Idx": 3531 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381119459.731, "dur": 6.162 , "args": { "External id": 3533,"Record function id": 0, "Ev Idx": 3532 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119467.967, "dur": 10.750 , "args": { "External id": 3534,"Record function id": 0, "Ev Idx": 3533 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119468.508, "dur": 0.751 , "args": { "External id": 3535,"Record function id": 0, "Ev Idx": 3534 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119468.698, "dur": 0.461 , "args": { "External id": 3536,"Record function id": 0, "Ev Idx": 3535 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119468.939, "dur": 0.110 , "args": { "External id": 3537,"Record function id": 0, "Ev Idx": 3536 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119469.550, "dur": 0.641 , "args": { "External id": 3538,"Record function id": 0, "Ev Idx": 3537 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119469.720, "dur": 0.381 , "args": { "External id": 3539,"Record function id": 0, "Ev Idx": 3538 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119469.920, "dur": 0.080 , "args": { "External id": 3540,"Record function id": 0, "Ev Idx": 3539 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119470.451, "dur": 0.692 , "args": { "External id": 3541,"Record function id": 0, "Ev Idx": 3540 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119470.622, "dur": 0.430 , "args": { "External id": 3542,"Record function id": 0, "Ev Idx": 3541 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119470.802, "dur": 0.090 , "args": { "External id": 3543,"Record function id": 0, "Ev Idx": 3542 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119472.245, "dur": 1.352 , "args": { "External id": 3544,"Record function id": 0, "Ev Idx": 3543 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119472.475, "dur": 1.012 , "args": { "External id": 3545,"Record function id": 0, "Ev Idx": 3544 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119481.953, "dur": 3.827 , "args": { "External id": 3546,"Record function id": 0, "Ev Idx": 3545 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119486.852, "dur": 0.601 , "args": { "External id": 3547,"Record function id": 0, "Ev Idx": 3546 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119488.725, "dur": 0.561 , "args": { "External id": 3548,"Record function id": 0, "Ev Idx": 3547 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381119493.304, "dur": 0.320 , "args": { "External id": 3549,"Record function id": 0, "Ev Idx": 3548 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381119493.484, "dur": 0.030 , "args": { "External id": 3550,"Record function id": 0, "Ev Idx": 3549 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381119496.169, "dur": 6.382 , "args": { "External id": 3551,"Record function id": 0, "Ev Idx": 3550 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119504.745, "dur": 10.480 , "args": { "External id": 3552,"Record function id": 0, "Ev Idx": 3551 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119505.296, "dur": 0.792 , "args": { "External id": 3553,"Record function id": 0, "Ev Idx": 3552 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119505.477, "dur": 0.520 , "args": { "External id": 3554,"Record function id": 0, "Ev Idx": 3553 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119505.797, "dur": 0.090 , "args": { "External id": 3555,"Record function id": 0, "Ev Idx": 3554 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119506.378, "dur": 0.621 , "args": { "External id": 3556,"Record function id": 0, "Ev Idx": 3555 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119506.539, "dur": 0.370 , "args": { "External id": 3557,"Record function id": 0, "Ev Idx": 3556 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119506.729, "dur": 0.070 , "args": { "External id": 3558,"Record function id": 0, "Ev Idx": 3557 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119507.250, "dur": 0.651 , "args": { "External id": 3559,"Record function id": 0, "Ev Idx": 3558 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119507.410, "dur": 0.401 , "args": { "External id": 3560,"Record function id": 0, "Ev Idx": 3559 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119507.611, "dur": 0.090 , "args": { "External id": 3561,"Record function id": 0, "Ev Idx": 3560 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119508.943, "dur": 1.302 , "args": { "External id": 3562,"Record function id": 0, "Ev Idx": 3561 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119509.123, "dur": 1.022 , "args": { "External id": 3563,"Record function id": 0, "Ev Idx": 3562 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381119519.853, "dur": 4.318 , "args": { "External id": 3564,"Record function id": 0, "Ev Idx": 3563 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381119530.122, "dur": 4.719 , "args": { "External id": 3565,"Record function id": 0, "Ev Idx": 3564 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381119543.527, "dur": 4.599 , "args": { "External id": 3566,"Record function id": 0, "Ev Idx": 3565 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381119550.520, "dur": 17.573 , "args": { "External id": 3567,"Record function id": 0, "Ev Idx": 3566 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119550.971, "dur": 0.511 , "args": { "External id": 3568,"Record function id": 0, "Ev Idx": 3567 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119551.212, "dur": 0.160 , "args": { "External id": 3569,"Record function id": 0, "Ev Idx": 3568 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119551.723, "dur": 0.400 , "args": { "External id": 3570,"Record function id": 0, "Ev Idx": 3569 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119551.923, "dur": 0.090 , "args": { "External id": 3571,"Record function id": 0, "Ev Idx": 3570 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119552.364, "dur": 0.381 , "args": { "External id": 3572,"Record function id": 0, "Ev Idx": 3571 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119552.554, "dur": 0.090 , "args": { "External id": 3573,"Record function id": 0, "Ev Idx": 3572 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381119553.085, "dur": 13.465 , "args": { "External id": 3574,"Record function id": 0, "Ev Idx": 3573 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381119553.997, "dur": 1.863 , "args": { "External id": 3575,"Record function id": 0, "Ev Idx": 3574 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119554.317, "dur": 1.443 , "args": { "External id": 3576,"Record function id": 0, "Ev Idx": 3575 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119556.301, "dur": 1.252 , "args": { "External id": 3577,"Record function id": 0, "Ev Idx": 3576 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119557.984, "dur": 1.192 , "args": { "External id": 3578,"Record function id": 0, "Ev Idx": 3577 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119559.457, "dur": 1.373 , "args": { "External id": 3579,"Record function id": 0, "Ev Idx": 3578 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119565.158, "dur": 1.022 , "args": { "External id": 3580,"Record function id": 0, "Ev Idx": 3579 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119566.851, "dur": 0.491 , "args": { "External id": 3581,"Record function id": 0, "Ev Idx": 3580 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119567.071, "dur": 0.160 , "args": { "External id": 3582,"Record function id": 0, "Ev Idx": 3581 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119575.627, "dur": 6.883 , "args": { "External id": 3583,"Record function id": 0, "Ev Idx": 3582 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119585.175, "dur": 14.377 , "args": { "External id": 3584,"Record function id": 0, "Ev Idx": 3583 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119589.062, "dur": 0.711 , "args": { "External id": 3585,"Record function id": 0, "Ev Idx": 3584 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119589.252, "dur": 0.431 , "args": { "External id": 3586,"Record function id": 0, "Ev Idx": 3585 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119589.473, "dur": 0.090 , "args": { "External id": 3587,"Record function id": 0, "Ev Idx": 3586 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119590.104, "dur": 0.711 , "args": { "External id": 3588,"Record function id": 0, "Ev Idx": 3587 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119590.284, "dur": 0.441 , "args": { "External id": 3589,"Record function id": 0, "Ev Idx": 3588 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119590.535, "dur": 0.080 , "args": { "External id": 3590,"Record function id": 0, "Ev Idx": 3589 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119591.086, "dur": 0.631 , "args": { "External id": 3591,"Record function id": 0, "Ev Idx": 3590 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119591.246, "dur": 0.381 , "args": { "External id": 3592,"Record function id": 0, "Ev Idx": 3591 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119591.447, "dur": 0.080 , "args": { "External id": 3593,"Record function id": 0, "Ev Idx": 3592 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119592.849, "dur": 1.323 , "args": { "External id": 3594,"Record function id": 0, "Ev Idx": 3593 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119593.040, "dur": 1.021 , "args": { "External id": 3595,"Record function id": 0, "Ev Idx": 3594 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381119603.800, "dur": 5.259 , "args": { "External id": 3596,"Record function id": 0, "Ev Idx": 3595 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119612.185, "dur": 5.520 , "args": { "External id": 3597,"Record function id": 0, "Ev Idx": 3596 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119621.092, "dur": 4.939 , "args": { "External id": 3598,"Record function id": 0, "Ev Idx": 3597 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119629.537, "dur": 10.440 , "args": { "External id": 3599,"Record function id": 0, "Ev Idx": 3598 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119630.078, "dur": 0.672 , "args": { "External id": 3600,"Record function id": 0, "Ev Idx": 3599 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119630.239, "dur": 0.421 , "args": { "External id": 3601,"Record function id": 0, "Ev Idx": 3600 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119630.459, "dur": 0.090 , "args": { "External id": 3602,"Record function id": 0, "Ev Idx": 3601 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119631.020, "dur": 0.661 , "args": { "External id": 3603,"Record function id": 0, "Ev Idx": 3602 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119631.201, "dur": 0.390 , "args": { "External id": 3604,"Record function id": 0, "Ev Idx": 3603 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119631.421, "dur": 0.070 , "args": { "External id": 3605,"Record function id": 0, "Ev Idx": 3604 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119631.932, "dur": 0.641 , "args": { "External id": 3606,"Record function id": 0, "Ev Idx": 3605 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119632.092, "dur": 0.391 , "args": { "External id": 3607,"Record function id": 0, "Ev Idx": 3606 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119632.303, "dur": 0.070 , "args": { "External id": 3608,"Record function id": 0, "Ev Idx": 3607 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119633.615, "dur": 1.282 , "args": { "External id": 3609,"Record function id": 0, "Ev Idx": 3608 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119633.815, "dur": 0.972 , "args": { "External id": 3610,"Record function id": 0, "Ev Idx": 3609 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119643.704, "dur": 9.928 , "args": { "External id": 3611,"Record function id": 0, "Ev Idx": 3610 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119644.215, "dur": 0.651 , "args": { "External id": 3612,"Record function id": 0, "Ev Idx": 3611 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119644.385, "dur": 0.381 , "args": { "External id": 3613,"Record function id": 0, "Ev Idx": 3612 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119644.585, "dur": 0.081 , "args": { "External id": 3614,"Record function id": 0, "Ev Idx": 3613 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119645.126, "dur": 0.622 , "args": { "External id": 3615,"Record function id": 0, "Ev Idx": 3614 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119645.287, "dur": 0.370 , "args": { "External id": 3616,"Record function id": 0, "Ev Idx": 3615 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119645.487, "dur": 0.070 , "args": { "External id": 3617,"Record function id": 0, "Ev Idx": 3616 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119646.008, "dur": 0.631 , "args": { "External id": 3618,"Record function id": 0, "Ev Idx": 3617 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119646.158, "dur": 0.391 , "args": { "External id": 3619,"Record function id": 0, "Ev Idx": 3618 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119646.359, "dur": 0.080 , "args": { "External id": 3620,"Record function id": 0, "Ev Idx": 3619 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119647.631, "dur": 1.252 , "args": { "External id": 3621,"Record function id": 0, "Ev Idx": 3620 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119647.821, "dur": 0.952 , "args": { "External id": 3622,"Record function id": 0, "Ev Idx": 3621 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381119658.401, "dur": 4.238 , "args": { "External id": 3623,"Record function id": 0, "Ev Idx": 3622 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119666.055, "dur": 5.320 , "args": { "External id": 3624,"Record function id": 0, "Ev Idx": 3623 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119674.621, "dur": 10.229 , "args": { "External id": 3625,"Record function id": 0, "Ev Idx": 3624 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119675.122, "dur": 0.672 , "args": { "External id": 3626,"Record function id": 0, "Ev Idx": 3625 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119675.303, "dur": 0.400 , "args": { "External id": 3627,"Record function id": 0, "Ev Idx": 3626 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119675.513, "dur": 0.080 , "args": { "External id": 3628,"Record function id": 0, "Ev Idx": 3627 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119676.064, "dur": 0.671 , "args": { "External id": 3629,"Record function id": 0, "Ev Idx": 3628 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119676.234, "dur": 0.411 , "args": { "External id": 3630,"Record function id": 0, "Ev Idx": 3629 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119676.455, "dur": 0.090 , "args": { "External id": 3631,"Record function id": 0, "Ev Idx": 3630 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119676.996, "dur": 0.611 , "args": { "External id": 3632,"Record function id": 0, "Ev Idx": 3631 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119677.156, "dur": 0.361 , "args": { "External id": 3633,"Record function id": 0, "Ev Idx": 3632 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119677.346, "dur": 0.071 , "args": { "External id": 3634,"Record function id": 0, "Ev Idx": 3633 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119678.659, "dur": 1.262 , "args": { "External id": 3635,"Record function id": 0, "Ev Idx": 3634 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119678.839, "dur": 0.982 , "args": { "External id": 3636,"Record function id": 0, "Ev Idx": 3635 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381119689.499, "dur": 4.699 , "args": { "External id": 3637,"Record function id": 0, "Ev Idx": 3636 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381119713.494, "dur": 24.405 , "args": { "External id": 3638,"Record function id": 0, "Ev Idx": 3637 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381119738.310, "dur": 339.131 , "args": { "External id": 3639,"Record function id": 0, "Ev Idx": 3638 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381119743.279, "dur": 10.620 , "args": { "External id": 3640,"Record function id": 0, "Ev Idx": 3639 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119746.746, "dur": 0.370 , "args": { "External id": 3641,"Record function id": 0, "Ev Idx": 3640 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381119747.567, "dur": 0.040 , "args": { "External id": 3642,"Record function id": 0, "Ev Idx": 3641 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381119747.988, "dur": 0.040 , "args": { "External id": 3643,"Record function id": 0, "Ev Idx": 3642 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119749.250, "dur": 0.291 , "args": { "External id": 3644,"Record function id": 0, "Ev Idx": 3643 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381119749.881, "dur": 0.041 , "args": { "External id": 3645,"Record function id": 0, "Ev Idx": 3644 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381119750.292, "dur": 0.020 , "args": { "External id": 3646,"Record function id": 0, "Ev Idx": 3645 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381119756.844, "dur": 1.173 , "args": { "External id": 3647,"Record function id": 0, "Ev Idx": 3646 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381119760.161, "dur": 314.275 , "args": { "External id": 3648,"Record function id": 0, "Ev Idx": 3647 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119768.616, "dur": 6.322 , "args": { "External id": 3649,"Record function id": 0, "Ev Idx": 3648 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119781.500, "dur": 7.785 , "args": { "External id": 3650,"Record function id": 0, "Ev Idx": 3649 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119795.316, "dur": 5.530 , "args": { "External id": 3651,"Record function id": 0, "Ev Idx": 3650 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119806.427, "dur": 5.730 , "args": { "External id": 3652,"Record function id": 0, "Ev Idx": 3651 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119814.712, "dur": 12.103 , "args": { "External id": 3653,"Record function id": 0, "Ev Idx": 3652 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119815.293, "dur": 1.072 , "args": { "External id": 3654,"Record function id": 0, "Ev Idx": 3653 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119815.524, "dur": 0.721 , "args": { "External id": 3655,"Record function id": 0, "Ev Idx": 3654 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119815.834, "dur": 0.241 , "args": { "External id": 3656,"Record function id": 0, "Ev Idx": 3655 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119816.666, "dur": 0.671 , "args": { "External id": 3657,"Record function id": 0, "Ev Idx": 3656 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119816.846, "dur": 0.401 , "args": { "External id": 3658,"Record function id": 0, "Ev Idx": 3657 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119817.067, "dur": 0.070 , "args": { "External id": 3659,"Record function id": 0, "Ev Idx": 3658 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119817.608, "dur": 0.641 , "args": { "External id": 3660,"Record function id": 0, "Ev Idx": 3659 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119817.758, "dur": 0.401 , "args": { "External id": 3661,"Record function id": 0, "Ev Idx": 3660 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119817.968, "dur": 0.080 , "args": { "External id": 3662,"Record function id": 0, "Ev Idx": 3661 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119819.481, "dur": 1.433 , "args": { "External id": 3663,"Record function id": 0, "Ev Idx": 3662 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119819.701, "dur": 1.102 , "args": { "External id": 3664,"Record function id": 0, "Ev Idx": 3663 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119832.215, "dur": 4.719 , "args": { "External id": 3665,"Record function id": 0, "Ev Idx": 3664 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119839.057, "dur": 0.932 , "args": { "External id": 3666,"Record function id": 0, "Ev Idx": 3665 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119841.752, "dur": 0.562 , "args": { "External id": 3667,"Record function id": 0, "Ev Idx": 3666 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381119846.030, "dur": 0.491 , "args": { "External id": 3668,"Record function id": 0, "Ev Idx": 3667 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381119846.351, "dur": 0.050 , "args": { "External id": 3669,"Record function id": 0, "Ev Idx": 3668 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381119849.527, "dur": 6.272 , "args": { "External id": 3670,"Record function id": 0, "Ev Idx": 3669 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119857.852, "dur": 10.771 , "args": { "External id": 3671,"Record function id": 0, "Ev Idx": 3670 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119858.393, "dur": 0.762 , "args": { "External id": 3672,"Record function id": 0, "Ev Idx": 3671 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119858.584, "dur": 0.471 , "args": { "External id": 3673,"Record function id": 0, "Ev Idx": 3672 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119858.844, "dur": 0.100 , "args": { "External id": 3674,"Record function id": 0, "Ev Idx": 3673 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119859.425, "dur": 0.662 , "args": { "External id": 3675,"Record function id": 0, "Ev Idx": 3674 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119859.606, "dur": 0.390 , "args": { "External id": 3676,"Record function id": 0, "Ev Idx": 3675 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119859.816, "dur": 0.070 , "args": { "External id": 3677,"Record function id": 0, "Ev Idx": 3676 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119860.347, "dur": 0.721 , "args": { "External id": 3678,"Record function id": 0, "Ev Idx": 3677 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119860.507, "dur": 0.471 , "args": { "External id": 3679,"Record function id": 0, "Ev Idx": 3678 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119860.708, "dur": 0.170 , "args": { "External id": 3680,"Record function id": 0, "Ev Idx": 3679 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119862.170, "dur": 1.293 , "args": { "External id": 3681,"Record function id": 0, "Ev Idx": 3680 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119862.381, "dur": 0.972 , "args": { "External id": 3682,"Record function id": 0, "Ev Idx": 3681 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381119871.869, "dur": 3.827 , "args": { "External id": 3683,"Record function id": 0, "Ev Idx": 3682 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119876.748, "dur": 0.601 , "args": { "External id": 3684,"Record function id": 0, "Ev Idx": 3683 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381119878.681, "dur": 0.561 , "args": { "External id": 3685,"Record function id": 0, "Ev Idx": 3684 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381119883.089, "dur": 0.321 , "args": { "External id": 3686,"Record function id": 0, "Ev Idx": 3685 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381119883.260, "dur": 0.040 , "args": { "External id": 3687,"Record function id": 0, "Ev Idx": 3686 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381119885.955, "dur": 5.470 , "args": { "External id": 3688,"Record function id": 0, "Ev Idx": 3687 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119893.729, "dur": 10.370 , "args": { "External id": 3689,"Record function id": 0, "Ev Idx": 3688 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119894.260, "dur": 0.732 , "args": { "External id": 3690,"Record function id": 0, "Ev Idx": 3689 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119894.461, "dur": 0.440 , "args": { "External id": 3691,"Record function id": 0, "Ev Idx": 3690 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119894.701, "dur": 0.090 , "args": { "External id": 3692,"Record function id": 0, "Ev Idx": 3691 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119895.262, "dur": 0.641 , "args": { "External id": 3693,"Record function id": 0, "Ev Idx": 3692 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119895.432, "dur": 0.381 , "args": { "External id": 3694,"Record function id": 0, "Ev Idx": 3693 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119895.633, "dur": 0.080 , "args": { "External id": 3695,"Record function id": 0, "Ev Idx": 3694 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119896.154, "dur": 0.621 , "args": { "External id": 3696,"Record function id": 0, "Ev Idx": 3695 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119896.314, "dur": 0.371 , "args": { "External id": 3697,"Record function id": 0, "Ev Idx": 3696 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119896.504, "dur": 0.081 , "args": { "External id": 3698,"Record function id": 0, "Ev Idx": 3697 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119897.817, "dur": 1.262 , "args": { "External id": 3699,"Record function id": 0, "Ev Idx": 3698 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119898.017, "dur": 0.962 , "args": { "External id": 3700,"Record function id": 0, "Ev Idx": 3699 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381119908.517, "dur": 4.268 , "args": { "External id": 3701,"Record function id": 0, "Ev Idx": 3700 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381119918.165, "dur": 4.949 , "args": { "External id": 3702,"Record function id": 0, "Ev Idx": 3701 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381119928.033, "dur": 4.699 , "args": { "External id": 3703,"Record function id": 0, "Ev Idx": 3702 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381119934.986, "dur": 17.102 , "args": { "External id": 3704,"Record function id": 0, "Ev Idx": 3703 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119935.467, "dur": 0.511 , "args": { "External id": 3705,"Record function id": 0, "Ev Idx": 3704 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119935.707, "dur": 0.151 , "args": { "External id": 3706,"Record function id": 0, "Ev Idx": 3705 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119936.218, "dur": 0.401 , "args": { "External id": 3707,"Record function id": 0, "Ev Idx": 3706 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119936.419, "dur": 0.090 , "args": { "External id": 3708,"Record function id": 0, "Ev Idx": 3707 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119936.870, "dur": 0.380 , "args": { "External id": 3709,"Record function id": 0, "Ev Idx": 3708 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119937.060, "dur": 0.080 , "args": { "External id": 3710,"Record function id": 0, "Ev Idx": 3709 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381119937.601, "dur": 12.994 , "args": { "External id": 3711,"Record function id": 0, "Ev Idx": 3710 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381119938.533, "dur": 1.873 , "args": { "External id": 3712,"Record function id": 0, "Ev Idx": 3711 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119938.853, "dur": 1.433 , "args": { "External id": 3713,"Record function id": 0, "Ev Idx": 3712 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119940.867, "dur": 1.232 , "args": { "External id": 3714,"Record function id": 0, "Ev Idx": 3713 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119942.500, "dur": 1.182 , "args": { "External id": 3715,"Record function id": 0, "Ev Idx": 3714 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119943.963, "dur": 1.282 , "args": { "External id": 3716,"Record function id": 0, "Ev Idx": 3715 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119949.283, "dur": 0.971 , "args": { "External id": 3717,"Record function id": 0, "Ev Idx": 3716 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119950.896, "dur": 0.450 , "args": { "External id": 3718,"Record function id": 0, "Ev Idx": 3717 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119951.116, "dur": 0.130 , "args": { "External id": 3719,"Record function id": 0, "Ev Idx": 3718 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119959.051, "dur": 6.863 , "args": { "External id": 3720,"Record function id": 0, "Ev Idx": 3719 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381119968.468, "dur": 10.780 , "args": { "External id": 3721,"Record function id": 0, "Ev Idx": 3720 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119969.079, "dur": 0.672 , "args": { "External id": 3722,"Record function id": 0, "Ev Idx": 3721 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119969.250, "dur": 0.411 , "args": { "External id": 3723,"Record function id": 0, "Ev Idx": 3722 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119969.470, "dur": 0.080 , "args": { "External id": 3724,"Record function id": 0, "Ev Idx": 3723 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119970.041, "dur": 0.682 , "args": { "External id": 3725,"Record function id": 0, "Ev Idx": 3724 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119970.212, "dur": 0.420 , "args": { "External id": 3726,"Record function id": 0, "Ev Idx": 3725 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119970.452, "dur": 0.070 , "args": { "External id": 3727,"Record function id": 0, "Ev Idx": 3726 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381119970.983, "dur": 0.621 , "args": { "External id": 3728,"Record function id": 0, "Ev Idx": 3727 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381119971.133, "dur": 0.381 , "args": { "External id": 3729,"Record function id": 0, "Ev Idx": 3728 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381119971.334, "dur": 0.070 , "args": { "External id": 3730,"Record function id": 0, "Ev Idx": 3729 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119972.676, "dur": 1.273 , "args": { "External id": 3731,"Record function id": 0, "Ev Idx": 3730 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381119972.877, "dur": 0.961 , "args": { "External id": 3732,"Record function id": 0, "Ev Idx": 3731 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381119983.346, "dur": 5.300 , "args": { "External id": 3733,"Record function id": 0, "Ev Idx": 3732 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381119991.672, "dur": 5.229 , "args": { "External id": 3734,"Record function id": 0, "Ev Idx": 3733 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120000.157, "dur": 4.990 , "args": { "External id": 3735,"Record function id": 0, "Ev Idx": 3734 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120008.603, "dur": 10.259 , "args": { "External id": 3736,"Record function id": 0, "Ev Idx": 3735 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120009.114, "dur": 0.681 , "args": { "External id": 3737,"Record function id": 0, "Ev Idx": 3736 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120009.304, "dur": 0.401 , "args": { "External id": 3738,"Record function id": 0, "Ev Idx": 3737 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120009.505, "dur": 0.090 , "args": { "External id": 3739,"Record function id": 0, "Ev Idx": 3738 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120010.076, "dur": 0.651 , "args": { "External id": 3740,"Record function id": 0, "Ev Idx": 3739 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120010.226, "dur": 0.411 , "args": { "External id": 3741,"Record function id": 0, "Ev Idx": 3740 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120010.467, "dur": 0.070 , "args": { "External id": 3742,"Record function id": 0, "Ev Idx": 3741 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120010.977, "dur": 0.672 , "args": { "External id": 3743,"Record function id": 0, "Ev Idx": 3742 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120011.148, "dur": 0.411 , "args": { "External id": 3744,"Record function id": 0, "Ev Idx": 3743 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120011.378, "dur": 0.070 , "args": { "External id": 3745,"Record function id": 0, "Ev Idx": 3744 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120012.681, "dur": 1.272 , "args": { "External id": 3746,"Record function id": 0, "Ev Idx": 3745 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120012.871, "dur": 0.972 , "args": { "External id": 3747,"Record function id": 0, "Ev Idx": 3746 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120022.659, "dur": 10.079 , "args": { "External id": 3748,"Record function id": 0, "Ev Idx": 3747 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120023.160, "dur": 0.641 , "args": { "External id": 3749,"Record function id": 0, "Ev Idx": 3748 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120023.330, "dur": 0.381 , "args": { "External id": 3750,"Record function id": 0, "Ev Idx": 3749 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120023.531, "dur": 0.070 , "args": { "External id": 3751,"Record function id": 0, "Ev Idx": 3750 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120024.062, "dur": 0.621 , "args": { "External id": 3752,"Record function id": 0, "Ev Idx": 3751 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120024.212, "dur": 0.381 , "args": { "External id": 3753,"Record function id": 0, "Ev Idx": 3752 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120024.423, "dur": 0.070 , "args": { "External id": 3754,"Record function id": 0, "Ev Idx": 3753 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120024.943, "dur": 0.642 , "args": { "External id": 3755,"Record function id": 0, "Ev Idx": 3754 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120025.094, "dur": 0.401 , "args": { "External id": 3756,"Record function id": 0, "Ev Idx": 3755 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120025.314, "dur": 0.080 , "args": { "External id": 3757,"Record function id": 0, "Ev Idx": 3756 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120026.577, "dur": 1.302 , "args": { "External id": 3758,"Record function id": 0, "Ev Idx": 3757 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120026.757, "dur": 1.012 , "args": { "External id": 3759,"Record function id": 0, "Ev Idx": 3758 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381120037.327, "dur": 4.207 , "args": { "External id": 3760,"Record function id": 0, "Ev Idx": 3759 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120044.750, "dur": 5.180 , "args": { "External id": 3761,"Record function id": 0, "Ev Idx": 3760 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120053.206, "dur": 10.189 , "args": { "External id": 3762,"Record function id": 0, "Ev Idx": 3761 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120053.737, "dur": 0.661 , "args": { "External id": 3763,"Record function id": 0, "Ev Idx": 3762 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120053.897, "dur": 0.411 , "args": { "External id": 3764,"Record function id": 0, "Ev Idx": 3763 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120054.118, "dur": 0.090 , "args": { "External id": 3765,"Record function id": 0, "Ev Idx": 3764 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120054.679, "dur": 0.661 , "args": { "External id": 3766,"Record function id": 0, "Ev Idx": 3765 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120054.849, "dur": 0.401 , "args": { "External id": 3767,"Record function id": 0, "Ev Idx": 3766 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120055.070, "dur": 0.070 , "args": { "External id": 3768,"Record function id": 0, "Ev Idx": 3767 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120055.591, "dur": 0.631 , "args": { "External id": 3769,"Record function id": 0, "Ev Idx": 3768 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120055.751, "dur": 0.381 , "args": { "External id": 3770,"Record function id": 0, "Ev Idx": 3769 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120055.961, "dur": 0.070 , "args": { "External id": 3771,"Record function id": 0, "Ev Idx": 3770 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120057.234, "dur": 1.272 , "args": { "External id": 3772,"Record function id": 0, "Ev Idx": 3771 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120057.414, "dur": 0.982 , "args": { "External id": 3773,"Record function id": 0, "Ev Idx": 3772 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381120068.064, "dur": 4.769 , "args": { "External id": 3774,"Record function id": 0, "Ev Idx": 3773 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381120092.219, "dur": 23.764 , "args": { "External id": 3775,"Record function id": 0, "Ev Idx": 3774 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381120116.334, "dur": 337.949 , "args": { "External id": 3776,"Record function id": 0, "Ev Idx": 3775 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381120120.972, "dur": 10.470 , "args": { "External id": 3777,"Record function id": 0, "Ev Idx": 3776 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120124.298, "dur": 0.351 , "args": { "External id": 3778,"Record function id": 0, "Ev Idx": 3777 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381120125.100, "dur": 0.040 , "args": { "External id": 3779,"Record function id": 0, "Ev Idx": 3778 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381120125.491, "dur": 0.050 , "args": { "External id": 3780,"Record function id": 0, "Ev Idx": 3779 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120126.913, "dur": 0.281 , "args": { "External id": 3781,"Record function id": 0, "Ev Idx": 3780 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381120127.524, "dur": 0.030 , "args": { "External id": 3782,"Record function id": 0, "Ev Idx": 3781 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381120127.885, "dur": 0.020 , "args": { "External id": 3783,"Record function id": 0, "Ev Idx": 3782 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381120134.157, "dur": 1.132 , "args": { "External id": 3784,"Record function id": 0, "Ev Idx": 3783 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381120137.333, "dur": 313.904 , "args": { "External id": 3785,"Record function id": 0, "Ev Idx": 3784 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120145.548, "dur": 6.522 , "args": { "External id": 3786,"Record function id": 0, "Ev Idx": 3785 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120158.712, "dur": 7.074 , "args": { "External id": 3787,"Record function id": 0, "Ev Idx": 3786 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120171.657, "dur": 5.340 , "args": { "External id": 3788,"Record function id": 0, "Ev Idx": 3787 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120182.356, "dur": 5.130 , "args": { "External id": 3789,"Record function id": 0, "Ev Idx": 3788 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120189.961, "dur": 12.272 , "args": { "External id": 3790,"Record function id": 0, "Ev Idx": 3789 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120190.602, "dur": 1.052 , "args": { "External id": 3791,"Record function id": 0, "Ev Idx": 3790 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120190.812, "dur": 0.742 , "args": { "External id": 3792,"Record function id": 0, "Ev Idx": 3791 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120191.123, "dur": 0.290 , "args": { "External id": 3793,"Record function id": 0, "Ev Idx": 3792 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120191.974, "dur": 0.622 , "args": { "External id": 3794,"Record function id": 0, "Ev Idx": 3793 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120192.135, "dur": 0.370 , "args": { "External id": 3795,"Record function id": 0, "Ev Idx": 3794 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120192.335, "dur": 0.070 , "args": { "External id": 3796,"Record function id": 0, "Ev Idx": 3795 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120192.856, "dur": 0.701 , "args": { "External id": 3797,"Record function id": 0, "Ev Idx": 3796 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120193.016, "dur": 0.441 , "args": { "External id": 3798,"Record function id": 0, "Ev Idx": 3797 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120193.207, "dur": 0.150 , "args": { "External id": 3799,"Record function id": 0, "Ev Idx": 3798 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120194.780, "dur": 1.432 , "args": { "External id": 3800,"Record function id": 0, "Ev Idx": 3799 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120194.990, "dur": 1.102 , "args": { "External id": 3801,"Record function id": 0, "Ev Idx": 3800 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120207.664, "dur": 4.428 , "args": { "External id": 3802,"Record function id": 0, "Ev Idx": 3801 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120214.326, "dur": 0.902 , "args": { "External id": 3803,"Record function id": 0, "Ev Idx": 3802 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120217.121, "dur": 0.541 , "args": { "External id": 3804,"Record function id": 0, "Ev Idx": 3803 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381120221.409, "dur": 0.421 , "args": { "External id": 3805,"Record function id": 0, "Ev Idx": 3804 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381120221.670, "dur": 0.050 , "args": { "External id": 3806,"Record function id": 0, "Ev Idx": 3805 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381120224.795, "dur": 6.312 , "args": { "External id": 3807,"Record function id": 0, "Ev Idx": 3806 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120233.231, "dur": 11.021 , "args": { "External id": 3808,"Record function id": 0, "Ev Idx": 3807 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120233.762, "dur": 0.732 , "args": { "External id": 3809,"Record function id": 0, "Ev Idx": 3808 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120233.943, "dur": 0.460 , "args": { "External id": 3810,"Record function id": 0, "Ev Idx": 3809 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120234.183, "dur": 0.110 , "args": { "External id": 3811,"Record function id": 0, "Ev Idx": 3810 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120234.774, "dur": 0.641 , "args": { "External id": 3812,"Record function id": 0, "Ev Idx": 3811 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120234.944, "dur": 0.371 , "args": { "External id": 3813,"Record function id": 0, "Ev Idx": 3812 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120235.145, "dur": 0.070 , "args": { "External id": 3814,"Record function id": 0, "Ev Idx": 3813 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120235.666, "dur": 0.721 , "args": { "External id": 3815,"Record function id": 0, "Ev Idx": 3814 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120235.836, "dur": 0.461 , "args": { "External id": 3816,"Record function id": 0, "Ev Idx": 3815 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120236.036, "dur": 0.161 , "args": { "External id": 3817,"Record function id": 0, "Ev Idx": 3816 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120237.479, "dur": 1.303 , "args": { "External id": 3818,"Record function id": 0, "Ev Idx": 3817 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120237.679, "dur": 1.002 , "args": { "External id": 3819,"Record function id": 0, "Ev Idx": 3818 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120247.548, "dur": 3.887 , "args": { "External id": 3820,"Record function id": 0, "Ev Idx": 3819 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120252.537, "dur": 0.631 , "args": { "External id": 3821,"Record function id": 0, "Ev Idx": 3820 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120254.561, "dur": 0.591 , "args": { "External id": 3822,"Record function id": 0, "Ev Idx": 3821 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381120258.929, "dur": 0.321 , "args": { "External id": 3823,"Record function id": 0, "Ev Idx": 3822 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381120259.099, "dur": 0.040 , "args": { "External id": 3824,"Record function id": 0, "Ev Idx": 3823 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381120261.885, "dur": 5.570 , "args": { "External id": 3825,"Record function id": 0, "Ev Idx": 3824 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120269.589, "dur": 10.449 , "args": { "External id": 3826,"Record function id": 0, "Ev Idx": 3825 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120270.140, "dur": 0.771 , "args": { "External id": 3827,"Record function id": 0, "Ev Idx": 3826 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120270.330, "dur": 0.491 , "args": { "External id": 3828,"Record function id": 0, "Ev Idx": 3827 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120270.561, "dur": 0.150 , "args": { "External id": 3829,"Record function id": 0, "Ev Idx": 3828 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120271.172, "dur": 0.631 , "args": { "External id": 3830,"Record function id": 0, "Ev Idx": 3829 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120271.342, "dur": 0.371 , "args": { "External id": 3831,"Record function id": 0, "Ev Idx": 3830 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120271.543, "dur": 0.070 , "args": { "External id": 3832,"Record function id": 0, "Ev Idx": 3831 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120272.063, "dur": 0.762 , "args": { "External id": 3833,"Record function id": 0, "Ev Idx": 3832 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120272.234, "dur": 0.501 , "args": { "External id": 3834,"Record function id": 0, "Ev Idx": 3833 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120272.434, "dur": 0.201 , "args": { "External id": 3835,"Record function id": 0, "Ev Idx": 3834 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120273.867, "dur": 1.252 , "args": { "External id": 3836,"Record function id": 0, "Ev Idx": 3835 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120274.037, "dur": 0.982 , "args": { "External id": 3837,"Record function id": 0, "Ev Idx": 3836 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381120285.529, "dur": 4.288 , "args": { "External id": 3838,"Record function id": 0, "Ev Idx": 3837 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381120295.116, "dur": 4.739 , "args": { "External id": 3839,"Record function id": 0, "Ev Idx": 3838 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381120304.885, "dur": 4.428 , "args": { "External id": 3840,"Record function id": 0, "Ev Idx": 3839 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381120311.557, "dur": 17.392 , "args": { "External id": 3841,"Record function id": 0, "Ev Idx": 3840 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120312.018, "dur": 0.491 , "args": { "External id": 3842,"Record function id": 0, "Ev Idx": 3841 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120312.258, "dur": 0.131 , "args": { "External id": 3843,"Record function id": 0, "Ev Idx": 3842 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120312.749, "dur": 0.391 , "args": { "External id": 3844,"Record function id": 0, "Ev Idx": 3843 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120312.940, "dur": 0.090 , "args": { "External id": 3845,"Record function id": 0, "Ev Idx": 3844 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120313.390, "dur": 0.401 , "args": { "External id": 3846,"Record function id": 0, "Ev Idx": 3845 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120313.601, "dur": 0.090 , "args": { "External id": 3847,"Record function id": 0, "Ev Idx": 3846 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381120314.122, "dur": 13.355 , "args": { "External id": 3848,"Record function id": 0, "Ev Idx": 3847 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381120315.043, "dur": 1.934 , "args": { "External id": 3849,"Record function id": 0, "Ev Idx": 3848 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120315.344, "dur": 1.523 , "args": { "External id": 3850,"Record function id": 0, "Ev Idx": 3849 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120317.408, "dur": 1.232 , "args": { "External id": 3851,"Record function id": 0, "Ev Idx": 3850 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120319.071, "dur": 1.332 , "args": { "External id": 3852,"Record function id": 0, "Ev Idx": 3851 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120320.684, "dur": 1.403 , "args": { "External id": 3853,"Record function id": 0, "Ev Idx": 3852 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120326.124, "dur": 1.012 , "args": { "External id": 3854,"Record function id": 0, "Ev Idx": 3853 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120327.747, "dur": 0.471 , "args": { "External id": 3855,"Record function id": 0, "Ev Idx": 3854 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120327.968, "dur": 0.140 , "args": { "External id": 3856,"Record function id": 0, "Ev Idx": 3855 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120335.882, "dur": 6.542 , "args": { "External id": 3857,"Record function id": 0, "Ev Idx": 3856 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120344.899, "dur": 10.560 , "args": { "External id": 3858,"Record function id": 0, "Ev Idx": 3857 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120345.500, "dur": 0.701 , "args": { "External id": 3859,"Record function id": 0, "Ev Idx": 3858 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120345.701, "dur": 0.410 , "args": { "External id": 3860,"Record function id": 0, "Ev Idx": 3859 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120345.921, "dur": 0.080 , "args": { "External id": 3861,"Record function id": 0, "Ev Idx": 3860 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120346.482, "dur": 0.651 , "args": { "External id": 3862,"Record function id": 0, "Ev Idx": 3861 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120346.642, "dur": 0.391 , "args": { "External id": 3863,"Record function id": 0, "Ev Idx": 3862 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120346.853, "dur": 0.080 , "args": { "External id": 3864,"Record function id": 0, "Ev Idx": 3863 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120347.384, "dur": 0.641 , "args": { "External id": 3865,"Record function id": 0, "Ev Idx": 3864 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120347.554, "dur": 0.381 , "args": { "External id": 3866,"Record function id": 0, "Ev Idx": 3865 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120347.754, "dur": 0.071 , "args": { "External id": 3867,"Record function id": 0, "Ev Idx": 3866 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120349.147, "dur": 1.252 , "args": { "External id": 3868,"Record function id": 0, "Ev Idx": 3867 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120349.337, "dur": 0.952 , "args": { "External id": 3869,"Record function id": 0, "Ev Idx": 3868 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381120359.516, "dur": 5.170 , "args": { "External id": 3870,"Record function id": 0, "Ev Idx": 3869 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120367.561, "dur": 5.320 , "args": { "External id": 3871,"Record function id": 0, "Ev Idx": 3870 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120376.187, "dur": 5.110 , "args": { "External id": 3872,"Record function id": 0, "Ev Idx": 3871 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120384.924, "dur": 10.579 , "args": { "External id": 3873,"Record function id": 0, "Ev Idx": 3872 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120385.445, "dur": 0.691 , "args": { "External id": 3874,"Record function id": 0, "Ev Idx": 3873 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120385.625, "dur": 0.421 , "args": { "External id": 3875,"Record function id": 0, "Ev Idx": 3874 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120385.855, "dur": 0.080 , "args": { "External id": 3876,"Record function id": 0, "Ev Idx": 3875 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120386.416, "dur": 0.632 , "args": { "External id": 3877,"Record function id": 0, "Ev Idx": 3876 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120386.577, "dur": 0.380 , "args": { "External id": 3878,"Record function id": 0, "Ev Idx": 3877 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120386.787, "dur": 0.070 , "args": { "External id": 3879,"Record function id": 0, "Ev Idx": 3878 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120387.318, "dur": 0.631 , "args": { "External id": 3880,"Record function id": 0, "Ev Idx": 3879 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120387.468, "dur": 0.381 , "args": { "External id": 3881,"Record function id": 0, "Ev Idx": 3880 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120387.679, "dur": 0.070 , "args": { "External id": 3882,"Record function id": 0, "Ev Idx": 3881 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120389.011, "dur": 1.253 , "args": { "External id": 3883,"Record function id": 0, "Ev Idx": 3882 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120389.181, "dur": 0.982 , "args": { "External id": 3884,"Record function id": 0, "Ev Idx": 3883 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120399.270, "dur": 9.999 , "args": { "External id": 3885,"Record function id": 0, "Ev Idx": 3884 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120399.771, "dur": 0.651 , "args": { "External id": 3886,"Record function id": 0, "Ev Idx": 3885 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120399.931, "dur": 0.391 , "args": { "External id": 3887,"Record function id": 0, "Ev Idx": 3886 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120400.142, "dur": 0.080 , "args": { "External id": 3888,"Record function id": 0, "Ev Idx": 3887 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120400.673, "dur": 0.641 , "args": { "External id": 3889,"Record function id": 0, "Ev Idx": 3888 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120400.843, "dur": 0.381 , "args": { "External id": 3890,"Record function id": 0, "Ev Idx": 3889 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120401.054, "dur": 0.070 , "args": { "External id": 3891,"Record function id": 0, "Ev Idx": 3890 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120401.585, "dur": 0.631 , "args": { "External id": 3892,"Record function id": 0, "Ev Idx": 3891 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120401.745, "dur": 0.371 , "args": { "External id": 3893,"Record function id": 0, "Ev Idx": 3892 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120401.945, "dur": 0.070 , "args": { "External id": 3894,"Record function id": 0, "Ev Idx": 3893 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120403.208, "dur": 1.242 , "args": { "External id": 3895,"Record function id": 0, "Ev Idx": 3894 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120403.388, "dur": 0.962 , "args": { "External id": 3896,"Record function id": 0, "Ev Idx": 3895 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381120414.028, "dur": 4.248 , "args": { "External id": 3897,"Record function id": 0, "Ev Idx": 3896 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120421.472, "dur": 5.340 , "args": { "External id": 3898,"Record function id": 0, "Ev Idx": 3897 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120430.068, "dur": 10.279 , "args": { "External id": 3899,"Record function id": 0, "Ev Idx": 3898 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120430.589, "dur": 0.681 , "args": { "External id": 3900,"Record function id": 0, "Ev Idx": 3899 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120430.759, "dur": 0.421 , "args": { "External id": 3901,"Record function id": 0, "Ev Idx": 3900 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120430.989, "dur": 0.080 , "args": { "External id": 3902,"Record function id": 0, "Ev Idx": 3901 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120431.540, "dur": 0.662 , "args": { "External id": 3903,"Record function id": 0, "Ev Idx": 3902 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120431.701, "dur": 0.410 , "args": { "External id": 3904,"Record function id": 0, "Ev Idx": 3903 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120431.941, "dur": 0.070 , "args": { "External id": 3905,"Record function id": 0, "Ev Idx": 3904 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120432.462, "dur": 0.621 , "args": { "External id": 3906,"Record function id": 0, "Ev Idx": 3905 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120432.612, "dur": 0.381 , "args": { "External id": 3907,"Record function id": 0, "Ev Idx": 3906 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120432.813, "dur": 0.070 , "args": { "External id": 3908,"Record function id": 0, "Ev Idx": 3907 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120434.135, "dur": 1.242 , "args": { "External id": 3909,"Record function id": 0, "Ev Idx": 3908 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120434.316, "dur": 0.961 , "args": { "External id": 3910,"Record function id": 0, "Ev Idx": 3909 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381120444.915, "dur": 4.729 , "args": { "External id": 3911,"Record function id": 0, "Ev Idx": 3910 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381120469.621, "dur": 22.963 , "args": { "External id": 3912,"Record function id": 0, "Ev Idx": 3911 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381120492.955, "dur": 343.018 , "args": { "External id": 3913,"Record function id": 0, "Ev Idx": 3912 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381120497.683, "dur": 10.189 , "args": { "External id": 3914,"Record function id": 0, "Ev Idx": 3913 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120501.020, "dur": 0.380 , "args": { "External id": 3915,"Record function id": 0, "Ev Idx": 3914 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381120501.851, "dur": 0.050 , "args": { "External id": 3916,"Record function id": 0, "Ev Idx": 3915 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381120502.272, "dur": 0.050 , "args": { "External id": 3917,"Record function id": 0, "Ev Idx": 3916 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120503.464, "dur": 0.211 , "args": { "External id": 3918,"Record function id": 0, "Ev Idx": 3917 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381120504.015, "dur": 0.020 , "args": { "External id": 3919,"Record function id": 0, "Ev Idx": 3918 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381120504.366, "dur": 0.020 , "args": { "External id": 3920,"Record function id": 0, "Ev Idx": 3919 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381120510.918, "dur": 1.152 , "args": { "External id": 3921,"Record function id": 0, "Ev Idx": 3920 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381120514.314, "dur": 318.543 , "args": { "External id": 3922,"Record function id": 0, "Ev Idx": 3921 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120522.820, "dur": 6.983 , "args": { "External id": 3923,"Record function id": 0, "Ev Idx": 3922 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120536.596, "dur": 10.660 , "args": { "External id": 3924,"Record function id": 0, "Ev Idx": 3923 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120553.297, "dur": 5.450 , "args": { "External id": 3925,"Record function id": 0, "Ev Idx": 3924 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120564.277, "dur": 5.290 , "args": { "External id": 3926,"Record function id": 0, "Ev Idx": 3925 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120572.202, "dur": 12.463 , "args": { "External id": 3927,"Record function id": 0, "Ev Idx": 3926 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120572.823, "dur": 1.062 , "args": { "External id": 3928,"Record function id": 0, "Ev Idx": 3927 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120573.074, "dur": 0.701 , "args": { "External id": 3929,"Record function id": 0, "Ev Idx": 3928 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120573.404, "dur": 0.231 , "args": { "External id": 3930,"Record function id": 0, "Ev Idx": 3929 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120574.196, "dur": 0.641 , "args": { "External id": 3931,"Record function id": 0, "Ev Idx": 3930 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120574.376, "dur": 0.371 , "args": { "External id": 3932,"Record function id": 0, "Ev Idx": 3931 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120574.577, "dur": 0.070 , "args": { "External id": 3933,"Record function id": 0, "Ev Idx": 3932 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120575.108, "dur": 0.651 , "args": { "External id": 3934,"Record function id": 0, "Ev Idx": 3933 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120575.278, "dur": 0.381 , "args": { "External id": 3935,"Record function id": 0, "Ev Idx": 3934 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120575.478, "dur": 0.080 , "args": { "External id": 3936,"Record function id": 0, "Ev Idx": 3935 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120577.071, "dur": 1.473 , "args": { "External id": 3937,"Record function id": 0, "Ev Idx": 3936 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120577.302, "dur": 1.122 , "args": { "External id": 3938,"Record function id": 0, "Ev Idx": 3937 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120590.156, "dur": 4.518 , "args": { "External id": 3939,"Record function id": 0, "Ev Idx": 3938 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120596.988, "dur": 0.822 , "args": { "External id": 3940,"Record function id": 0, "Ev Idx": 3939 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120599.643, "dur": 0.551 , "args": { "External id": 3941,"Record function id": 0, "Ev Idx": 3940 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381120603.931, "dur": 0.451 , "args": { "External id": 3942,"Record function id": 0, "Ev Idx": 3941 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381120604.212, "dur": 0.060 , "args": { "External id": 3943,"Record function id": 0, "Ev Idx": 3942 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381120607.287, "dur": 5.992 , "args": { "External id": 3944,"Record function id": 0, "Ev Idx": 3943 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120615.352, "dur": 10.901 , "args": { "External id": 3945,"Record function id": 0, "Ev Idx": 3944 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120615.883, "dur": 0.762 , "args": { "External id": 3946,"Record function id": 0, "Ev Idx": 3945 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120616.064, "dur": 0.491 , "args": { "External id": 3947,"Record function id": 0, "Ev Idx": 3946 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120616.344, "dur": 0.101 , "args": { "External id": 3948,"Record function id": 0, "Ev Idx": 3947 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120616.925, "dur": 0.632 , "args": { "External id": 3949,"Record function id": 0, "Ev Idx": 3948 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120617.096, "dur": 0.370 , "args": { "External id": 3950,"Record function id": 0, "Ev Idx": 3949 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120617.286, "dur": 0.070 , "args": { "External id": 3951,"Record function id": 0, "Ev Idx": 3950 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120617.817, "dur": 0.711 , "args": { "External id": 3952,"Record function id": 0, "Ev Idx": 3951 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120617.997, "dur": 0.451 , "args": { "External id": 3953,"Record function id": 0, "Ev Idx": 3952 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120618.178, "dur": 0.160 , "args": { "External id": 3954,"Record function id": 0, "Ev Idx": 3953 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120619.620, "dur": 1.293 , "args": { "External id": 3955,"Record function id": 0, "Ev Idx": 3954 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120619.811, "dur": 0.992 , "args": { "External id": 3956,"Record function id": 0, "Ev Idx": 3955 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120629.329, "dur": 3.726 , "args": { "External id": 3957,"Record function id": 0, "Ev Idx": 3956 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120634.037, "dur": 0.641 , "args": { "External id": 3958,"Record function id": 0, "Ev Idx": 3957 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120636.041, "dur": 0.591 , "args": { "External id": 3959,"Record function id": 0, "Ev Idx": 3958 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381120640.259, "dur": 0.320 , "args": { "External id": 3960,"Record function id": 0, "Ev Idx": 3959 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381120640.439, "dur": 0.040 , "args": { "External id": 3961,"Record function id": 0, "Ev Idx": 3960 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381120643.054, "dur": 5.711 , "args": { "External id": 3962,"Record function id": 0, "Ev Idx": 3961 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120650.999, "dur": 10.379 , "args": { "External id": 3963,"Record function id": 0, "Ev Idx": 3962 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120651.570, "dur": 0.691 , "args": { "External id": 3964,"Record function id": 0, "Ev Idx": 3963 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120651.760, "dur": 0.411 , "args": { "External id": 3965,"Record function id": 0, "Ev Idx": 3964 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120651.991, "dur": 0.070 , "args": { "External id": 3966,"Record function id": 0, "Ev Idx": 3965 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120652.552, "dur": 0.641 , "args": { "External id": 3967,"Record function id": 0, "Ev Idx": 3966 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120652.722, "dur": 0.381 , "args": { "External id": 3968,"Record function id": 0, "Ev Idx": 3967 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120652.922, "dur": 0.071 , "args": { "External id": 3969,"Record function id": 0, "Ev Idx": 3968 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120653.443, "dur": 0.642 , "args": { "External id": 3970,"Record function id": 0, "Ev Idx": 3969 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120653.624, "dur": 0.370 , "args": { "External id": 3971,"Record function id": 0, "Ev Idx": 3970 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120653.814, "dur": 0.080 , "args": { "External id": 3972,"Record function id": 0, "Ev Idx": 3971 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120655.127, "dur": 1.272 , "args": { "External id": 3973,"Record function id": 0, "Ev Idx": 3972 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120655.317, "dur": 0.972 , "args": { "External id": 3974,"Record function id": 0, "Ev Idx": 3973 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381120665.706, "dur": 4.168 , "args": { "External id": 3975,"Record function id": 0, "Ev Idx": 3974 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381120675.524, "dur": 4.920 , "args": { "External id": 3976,"Record function id": 0, "Ev Idx": 3975 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381120685.513, "dur": 4.358 , "args": { "External id": 3977,"Record function id": 0, "Ev Idx": 3976 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381120692.196, "dur": 17.993 , "args": { "External id": 3978,"Record function id": 0, "Ev Idx": 3977 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120692.747, "dur": 0.571 , "args": { "External id": 3979,"Record function id": 0, "Ev Idx": 3978 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120692.987, "dur": 0.210 , "args": { "External id": 3980,"Record function id": 0, "Ev Idx": 3979 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120693.558, "dur": 0.401 , "args": { "External id": 3981,"Record function id": 0, "Ev Idx": 3980 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120693.758, "dur": 0.091 , "args": { "External id": 3982,"Record function id": 0, "Ev Idx": 3981 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120694.209, "dur": 0.381 , "args": { "External id": 3983,"Record function id": 0, "Ev Idx": 3982 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120694.400, "dur": 0.080 , "args": { "External id": 3984,"Record function id": 0, "Ev Idx": 3983 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381120694.931, "dur": 13.685 , "args": { "External id": 3985,"Record function id": 0, "Ev Idx": 3984 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381120695.902, "dur": 1.964 , "args": { "External id": 3986,"Record function id": 0, "Ev Idx": 3985 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120696.213, "dur": 1.533 , "args": { "External id": 3987,"Record function id": 0, "Ev Idx": 3986 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120698.307, "dur": 1.292 , "args": { "External id": 3988,"Record function id": 0, "Ev Idx": 3987 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120700.040, "dur": 1.343 , "args": { "External id": 3989,"Record function id": 0, "Ev Idx": 3988 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120701.683, "dur": 1.403 , "args": { "External id": 3990,"Record function id": 0, "Ev Idx": 3989 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120707.264, "dur": 1.001 , "args": { "External id": 3991,"Record function id": 0, "Ev Idx": 3990 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120708.907, "dur": 0.531 , "args": { "External id": 3992,"Record function id": 0, "Ev Idx": 3991 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120709.127, "dur": 0.200 , "args": { "External id": 3993,"Record function id": 0, "Ev Idx": 3992 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120716.841, "dur": 6.623 , "args": { "External id": 3994,"Record function id": 0, "Ev Idx": 3993 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120726.049, "dur": 10.679 , "args": { "External id": 3995,"Record function id": 0, "Ev Idx": 3994 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120726.770, "dur": 0.691 , "args": { "External id": 3996,"Record function id": 0, "Ev Idx": 3995 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120726.960, "dur": 0.411 , "args": { "External id": 3997,"Record function id": 0, "Ev Idx": 3996 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120727.181, "dur": 0.080 , "args": { "External id": 3998,"Record function id": 0, "Ev Idx": 3997 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120727.772, "dur": 0.651 , "args": { "External id": 3999,"Record function id": 0, "Ev Idx": 3998 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120727.922, "dur": 0.411 , "args": { "External id": 4000,"Record function id": 0, "Ev Idx": 3999 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120728.162, "dur": 0.071 , "args": { "External id": 4001,"Record function id": 0, "Ev Idx": 4000 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120728.673, "dur": 0.642 , "args": { "External id": 4002,"Record function id": 0, "Ev Idx": 4001 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120728.844, "dur": 0.380 , "args": { "External id": 4003,"Record function id": 0, "Ev Idx": 4002 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120729.044, "dur": 0.080 , "args": { "External id": 4004,"Record function id": 0, "Ev Idx": 4003 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120730.427, "dur": 1.252 , "args": { "External id": 4005,"Record function id": 0, "Ev Idx": 4004 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120730.627, "dur": 0.952 , "args": { "External id": 4006,"Record function id": 0, "Ev Idx": 4005 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381120740.736, "dur": 5.290 , "args": { "External id": 4007,"Record function id": 0, "Ev Idx": 4006 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120748.891, "dur": 5.320 , "args": { "External id": 4008,"Record function id": 0, "Ev Idx": 4007 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120757.597, "dur": 5.260 , "args": { "External id": 4009,"Record function id": 0, "Ev Idx": 4008 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120766.374, "dur": 10.690 , "args": { "External id": 4010,"Record function id": 0, "Ev Idx": 4009 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120766.895, "dur": 0.761 , "args": { "External id": 4011,"Record function id": 0, "Ev Idx": 4010 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120767.075, "dur": 0.491 , "args": { "External id": 4012,"Record function id": 0, "Ev Idx": 4011 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120767.305, "dur": 0.151 , "args": { "External id": 4013,"Record function id": 0, "Ev Idx": 4012 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120767.926, "dur": 0.682 , "args": { "External id": 4014,"Record function id": 0, "Ev Idx": 4013 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120768.087, "dur": 0.431 , "args": { "External id": 4015,"Record function id": 0, "Ev Idx": 4014 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120768.327, "dur": 0.090 , "args": { "External id": 4016,"Record function id": 0, "Ev Idx": 4015 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120768.868, "dur": 0.651 , "args": { "External id": 4017,"Record function id": 0, "Ev Idx": 4016 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120769.039, "dur": 0.390 , "args": { "External id": 4018,"Record function id": 0, "Ev Idx": 4017 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120769.249, "dur": 0.080 , "args": { "External id": 4019,"Record function id": 0, "Ev Idx": 4018 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120770.581, "dur": 1.263 , "args": { "External id": 4020,"Record function id": 0, "Ev Idx": 4019 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120770.772, "dur": 0.972 , "args": { "External id": 4021,"Record function id": 0, "Ev Idx": 4020 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120780.820, "dur": 9.949 , "args": { "External id": 4022,"Record function id": 0, "Ev Idx": 4021 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120781.341, "dur": 0.652 , "args": { "External id": 4023,"Record function id": 0, "Ev Idx": 4022 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120781.522, "dur": 0.381 , "args": { "External id": 4024,"Record function id": 0, "Ev Idx": 4023 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120781.722, "dur": 0.070 , "args": { "External id": 4025,"Record function id": 0, "Ev Idx": 4024 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120782.263, "dur": 0.621 , "args": { "External id": 4026,"Record function id": 0, "Ev Idx": 4025 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120782.413, "dur": 0.381 , "args": { "External id": 4027,"Record function id": 0, "Ev Idx": 4026 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120782.624, "dur": 0.070 , "args": { "External id": 4028,"Record function id": 0, "Ev Idx": 4027 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120783.135, "dur": 0.641 , "args": { "External id": 4029,"Record function id": 0, "Ev Idx": 4028 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120783.295, "dur": 0.391 , "args": { "External id": 4030,"Record function id": 0, "Ev Idx": 4029 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120783.505, "dur": 0.081 , "args": { "External id": 4031,"Record function id": 0, "Ev Idx": 4030 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120784.798, "dur": 1.232 , "args": { "External id": 4032,"Record function id": 0, "Ev Idx": 4031 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120784.958, "dur": 0.972 , "args": { "External id": 4033,"Record function id": 0, "Ev Idx": 4032 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381120795.438, "dur": 4.087 , "args": { "External id": 4034,"Record function id": 0, "Ev Idx": 4033 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120802.852, "dur": 5.359 , "args": { "External id": 4035,"Record function id": 0, "Ev Idx": 4034 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120811.397, "dur": 10.209 , "args": { "External id": 4036,"Record function id": 0, "Ev Idx": 4035 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120811.928, "dur": 0.662 , "args": { "External id": 4037,"Record function id": 0, "Ev Idx": 4036 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120812.089, "dur": 0.410 , "args": { "External id": 4038,"Record function id": 0, "Ev Idx": 4037 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120812.309, "dur": 0.080 , "args": { "External id": 4039,"Record function id": 0, "Ev Idx": 4038 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120812.880, "dur": 0.661 , "args": { "External id": 4040,"Record function id": 0, "Ev Idx": 4039 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120813.050, "dur": 0.391 , "args": { "External id": 4041,"Record function id": 0, "Ev Idx": 4040 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120813.271, "dur": 0.070 , "args": { "External id": 4042,"Record function id": 0, "Ev Idx": 4041 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120813.802, "dur": 0.651 , "args": { "External id": 4043,"Record function id": 0, "Ev Idx": 4042 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120813.972, "dur": 0.381 , "args": { "External id": 4044,"Record function id": 0, "Ev Idx": 4043 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120814.173, "dur": 0.080 , "args": { "External id": 4045,"Record function id": 0, "Ev Idx": 4044 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120815.445, "dur": 1.272 , "args": { "External id": 4046,"Record function id": 0, "Ev Idx": 4045 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120815.645, "dur": 0.972 , "args": { "External id": 4047,"Record function id": 0, "Ev Idx": 4046 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381120826.285, "dur": 4.929 , "args": { "External id": 4048,"Record function id": 0, "Ev Idx": 4047 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381120850.530, "dur": 23.825 , "args": { "External id": 4049,"Record function id": 0, "Ev Idx": 4048 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381120874.765, "dur": 340.003 , "args": { "External id": 4050,"Record function id": 0, "Ev Idx": 4049 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381120879.584, "dur": 10.259 , "args": { "External id": 4051,"Record function id": 0, "Ev Idx": 4050 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120882.880, "dur": 0.381 , "args": { "External id": 4052,"Record function id": 0, "Ev Idx": 4051 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381120883.742, "dur": 0.050 , "args": { "External id": 4053,"Record function id": 0, "Ev Idx": 4052 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381120884.173, "dur": 0.040 , "args": { "External id": 4054,"Record function id": 0, "Ev Idx": 4053 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120885.385, "dur": 0.261 , "args": { "External id": 4055,"Record function id": 0, "Ev Idx": 4054 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381120886.006, "dur": 0.030 , "args": { "External id": 4056,"Record function id": 0, "Ev Idx": 4055 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381120886.377, "dur": 0.020 , "args": { "External id": 4057,"Record function id": 0, "Ev Idx": 4056 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381120892.508, "dur": 1.092 , "args": { "External id": 4058,"Record function id": 0, "Ev Idx": 4057 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381120895.634, "dur": 316.109 , "args": { "External id": 4059,"Record function id": 0, "Ev Idx": 4058 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120904.160, "dur": 6.312 , "args": { "External id": 4060,"Record function id": 0, "Ev Idx": 4059 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120917.244, "dur": 6.993 , "args": { "External id": 4061,"Record function id": 0, "Ev Idx": 4060 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120930.339, "dur": 5.410 , "args": { "External id": 4062,"Record function id": 0, "Ev Idx": 4061 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381120941.369, "dur": 5.240 , "args": { "External id": 4063,"Record function id": 0, "Ev Idx": 4062 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120949.064, "dur": 12.222 , "args": { "External id": 4064,"Record function id": 0, "Ev Idx": 4063 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120949.675, "dur": 1.002 , "args": { "External id": 4065,"Record function id": 0, "Ev Idx": 4064 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120949.905, "dur": 0.661 , "args": { "External id": 4066,"Record function id": 0, "Ev Idx": 4065 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120950.226, "dur": 0.210 , "args": { "External id": 4067,"Record function id": 0, "Ev Idx": 4066 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120950.987, "dur": 0.661 , "args": { "External id": 4068,"Record function id": 0, "Ev Idx": 4067 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120951.178, "dur": 0.380 , "args": { "External id": 4069,"Record function id": 0, "Ev Idx": 4068 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120951.378, "dur": 0.070 , "args": { "External id": 4070,"Record function id": 0, "Ev Idx": 4069 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120951.929, "dur": 0.631 , "args": { "External id": 4071,"Record function id": 0, "Ev Idx": 4070 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120952.089, "dur": 0.381 , "args": { "External id": 4072,"Record function id": 0, "Ev Idx": 4071 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120952.280, "dur": 0.090 , "args": { "External id": 4073,"Record function id": 0, "Ev Idx": 4072 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120953.873, "dur": 1.392 , "args": { "External id": 4074,"Record function id": 0, "Ev Idx": 4073 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120954.073, "dur": 1.072 , "args": { "External id": 4075,"Record function id": 0, "Ev Idx": 4074 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381120966.676, "dur": 4.749 , "args": { "External id": 4076,"Record function id": 0, "Ev Idx": 4075 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120973.599, "dur": 0.932 , "args": { "External id": 4077,"Record function id": 0, "Ev Idx": 4076 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381120976.304, "dur": 0.581 , "args": { "External id": 4078,"Record function id": 0, "Ev Idx": 4077 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381120980.713, "dur": 0.460 , "args": { "External id": 4079,"Record function id": 0, "Ev Idx": 4078 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381120980.993, "dur": 0.060 , "args": { "External id": 4080,"Record function id": 0, "Ev Idx": 4079 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381120984.299, "dur": 6.172 , "args": { "External id": 4081,"Record function id": 0, "Ev Idx": 4080 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381120992.665, "dur": 10.700 , "args": { "External id": 4082,"Record function id": 0, "Ev Idx": 4081 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120993.216, "dur": 0.741 , "args": { "External id": 4083,"Record function id": 0, "Ev Idx": 4082 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120993.406, "dur": 0.461 , "args": { "External id": 4084,"Record function id": 0, "Ev Idx": 4083 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120993.637, "dur": 0.120 , "args": { "External id": 4085,"Record function id": 0, "Ev Idx": 4084 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120994.228, "dur": 0.641 , "args": { "External id": 4086,"Record function id": 0, "Ev Idx": 4085 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120994.398, "dur": 0.381 , "args": { "External id": 4087,"Record function id": 0, "Ev Idx": 4086 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120994.588, "dur": 0.081 , "args": { "External id": 4088,"Record function id": 0, "Ev Idx": 4087 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381120995.129, "dur": 0.612 , "args": { "External id": 4089,"Record function id": 0, "Ev Idx": 4088 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381120995.280, "dur": 0.370 , "args": { "External id": 4090,"Record function id": 0, "Ev Idx": 4089 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381120995.470, "dur": 0.080 , "args": { "External id": 4091,"Record function id": 0, "Ev Idx": 4090 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120996.863, "dur": 1.312 , "args": { "External id": 4092,"Record function id": 0, "Ev Idx": 4091 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381120997.053, "dur": 1.022 , "args": { "External id": 4093,"Record function id": 0, "Ev Idx": 4092 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121006.551, "dur": 3.847 , "args": { "External id": 4094,"Record function id": 0, "Ev Idx": 4093 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121011.430, "dur": 0.601 , "args": { "External id": 4095,"Record function id": 0, "Ev Idx": 4094 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121013.343, "dur": 0.591 , "args": { "External id": 4096,"Record function id": 0, "Ev Idx": 4095 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381121020.497, "dur": 0.360 , "args": { "External id": 4097,"Record function id": 0, "Ev Idx": 4096 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381121020.707, "dur": 0.040 , "args": { "External id": 4098,"Record function id": 0, "Ev Idx": 4097 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381121023.572, "dur": 5.941 , "args": { "External id": 4099,"Record function id": 0, "Ev Idx": 4098 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121031.728, "dur": 10.389 , "args": { "External id": 4100,"Record function id": 0, "Ev Idx": 4099 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121032.289, "dur": 0.711 , "args": { "External id": 4101,"Record function id": 0, "Ev Idx": 4100 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121032.479, "dur": 0.431 , "args": { "External id": 4102,"Record function id": 0, "Ev Idx": 4101 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121032.709, "dur": 0.091 , "args": { "External id": 4103,"Record function id": 0, "Ev Idx": 4102 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121033.280, "dur": 0.652 , "args": { "External id": 4104,"Record function id": 0, "Ev Idx": 4103 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121033.461, "dur": 0.380 , "args": { "External id": 4105,"Record function id": 0, "Ev Idx": 4104 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121033.661, "dur": 0.080 , "args": { "External id": 4106,"Record function id": 0, "Ev Idx": 4105 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121034.192, "dur": 0.631 , "args": { "External id": 4107,"Record function id": 0, "Ev Idx": 4106 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121034.352, "dur": 0.381 , "args": { "External id": 4108,"Record function id": 0, "Ev Idx": 4107 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121034.543, "dur": 0.090 , "args": { "External id": 4109,"Record function id": 0, "Ev Idx": 4108 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121035.915, "dur": 1.283 , "args": { "External id": 4110,"Record function id": 0, "Ev Idx": 4109 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121036.106, "dur": 0.992 , "args": { "External id": 4111,"Record function id": 0, "Ev Idx": 4110 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381121046.705, "dur": 4.098 , "args": { "External id": 4112,"Record function id": 0, "Ev Idx": 4111 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381121056.133, "dur": 4.659 , "args": { "External id": 4113,"Record function id": 0, "Ev Idx": 4112 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381121065.871, "dur": 4.559 , "args": { "External id": 4114,"Record function id": 0, "Ev Idx": 4113 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381121072.724, "dur": 17.262 , "args": { "External id": 4115,"Record function id": 0, "Ev Idx": 4114 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121073.275, "dur": 0.571 , "args": { "External id": 4116,"Record function id": 0, "Ev Idx": 4115 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121073.595, "dur": 0.131 , "args": { "External id": 4117,"Record function id": 0, "Ev Idx": 4116 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121074.086, "dur": 0.411 , "args": { "External id": 4118,"Record function id": 0, "Ev Idx": 4117 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121074.287, "dur": 0.090 , "args": { "External id": 4119,"Record function id": 0, "Ev Idx": 4118 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121074.748, "dur": 0.380 , "args": { "External id": 4120,"Record function id": 0, "Ev Idx": 4119 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121074.938, "dur": 0.080 , "args": { "External id": 4121,"Record function id": 0, "Ev Idx": 4120 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381121075.489, "dur": 13.064 , "args": { "External id": 4122,"Record function id": 0, "Ev Idx": 4121 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381121076.381, "dur": 1.943 , "args": { "External id": 4123,"Record function id": 0, "Ev Idx": 4122 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121076.671, "dur": 1.543 , "args": { "External id": 4124,"Record function id": 0, "Ev Idx": 4123 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121078.765, "dur": 1.252 , "args": { "External id": 4125,"Record function id": 0, "Ev Idx": 4124 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121080.478, "dur": 1.223 , "args": { "External id": 4126,"Record function id": 0, "Ev Idx": 4125 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121081.971, "dur": 1.202 , "args": { "External id": 4127,"Record function id": 0, "Ev Idx": 4126 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121087.291, "dur": 0.922 , "args": { "External id": 4128,"Record function id": 0, "Ev Idx": 4127 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121088.834, "dur": 0.471 , "args": { "External id": 4129,"Record function id": 0, "Ev Idx": 4128 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121089.044, "dur": 0.151 , "args": { "External id": 4130,"Record function id": 0, "Ev Idx": 4129 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121096.789, "dur": 6.812 , "args": { "External id": 4131,"Record function id": 0, "Ev Idx": 4130 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121105.986, "dur": 10.690 , "args": { "External id": 4132,"Record function id": 0, "Ev Idx": 4131 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121106.577, "dur": 0.701 , "args": { "External id": 4133,"Record function id": 0, "Ev Idx": 4132 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121106.767, "dur": 0.421 , "args": { "External id": 4134,"Record function id": 0, "Ev Idx": 4133 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121106.988, "dur": 0.090 , "args": { "External id": 4135,"Record function id": 0, "Ev Idx": 4134 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121107.579, "dur": 0.681 , "args": { "External id": 4136,"Record function id": 0, "Ev Idx": 4135 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121107.759, "dur": 0.411 , "args": { "External id": 4137,"Record function id": 0, "Ev Idx": 4136 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121107.979, "dur": 0.081 , "args": { "External id": 4138,"Record function id": 0, "Ev Idx": 4137 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121108.510, "dur": 0.672 , "args": { "External id": 4139,"Record function id": 0, "Ev Idx": 4138 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121108.681, "dur": 0.401 , "args": { "External id": 4140,"Record function id": 0, "Ev Idx": 4139 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121108.901, "dur": 0.080 , "args": { "External id": 4141,"Record function id": 0, "Ev Idx": 4140 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121110.274, "dur": 1.282 , "args": { "External id": 4142,"Record function id": 0, "Ev Idx": 4141 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121110.484, "dur": 0.952 , "args": { "External id": 4143,"Record function id": 0, "Ev Idx": 4142 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381121120.703, "dur": 5.170 , "args": { "External id": 4144,"Record function id": 0, "Ev Idx": 4143 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121128.778, "dur": 5.220 , "args": { "External id": 4145,"Record function id": 0, "Ev Idx": 4144 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121137.294, "dur": 5.029 , "args": { "External id": 4146,"Record function id": 0, "Ev Idx": 4145 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121145.860, "dur": 10.299 , "args": { "External id": 4147,"Record function id": 0, "Ev Idx": 4146 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121146.381, "dur": 0.661 , "args": { "External id": 4148,"Record function id": 0, "Ev Idx": 4147 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121146.561, "dur": 0.391 , "args": { "External id": 4149,"Record function id": 0, "Ev Idx": 4148 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121146.772, "dur": 0.080 , "args": { "External id": 4150,"Record function id": 0, "Ev Idx": 4149 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121147.313, "dur": 0.641 , "args": { "External id": 4151,"Record function id": 0, "Ev Idx": 4150 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121147.473, "dur": 0.391 , "args": { "External id": 4152,"Record function id": 0, "Ev Idx": 4151 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121147.693, "dur": 0.071 , "args": { "External id": 4153,"Record function id": 0, "Ev Idx": 4152 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121148.214, "dur": 0.642 , "args": { "External id": 4154,"Record function id": 0, "Ev Idx": 4153 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121148.365, "dur": 0.400 , "args": { "External id": 4155,"Record function id": 0, "Ev Idx": 4154 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121148.585, "dur": 0.080 , "args": { "External id": 4156,"Record function id": 0, "Ev Idx": 4155 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121149.918, "dur": 1.282 , "args": { "External id": 4157,"Record function id": 0, "Ev Idx": 4156 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121150.108, "dur": 0.992 , "args": { "External id": 4158,"Record function id": 0, "Ev Idx": 4157 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121159.756, "dur": 9.988 , "args": { "External id": 4159,"Record function id": 0, "Ev Idx": 4158 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121160.257, "dur": 0.661 , "args": { "External id": 4160,"Record function id": 0, "Ev Idx": 4159 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121160.437, "dur": 0.391 , "args": { "External id": 4161,"Record function id": 0, "Ev Idx": 4160 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121160.648, "dur": 0.080 , "args": { "External id": 4162,"Record function id": 0, "Ev Idx": 4161 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121161.179, "dur": 0.641 , "args": { "External id": 4163,"Record function id": 0, "Ev Idx": 4162 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121161.349, "dur": 0.381 , "args": { "External id": 4164,"Record function id": 0, "Ev Idx": 4163 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121161.549, "dur": 0.080 , "args": { "External id": 4165,"Record function id": 0, "Ev Idx": 4164 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121162.070, "dur": 0.641 , "args": { "External id": 4166,"Record function id": 0, "Ev Idx": 4165 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121162.230, "dur": 0.391 , "args": { "External id": 4167,"Record function id": 0, "Ev Idx": 4166 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121162.441, "dur": 0.080 , "args": { "External id": 4168,"Record function id": 0, "Ev Idx": 4167 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121163.723, "dur": 1.233 , "args": { "External id": 4169,"Record function id": 0, "Ev Idx": 4168 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121163.904, "dur": 0.941 , "args": { "External id": 4170,"Record function id": 0, "Ev Idx": 4169 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381121174.443, "dur": 4.058 , "args": { "External id": 4171,"Record function id": 0, "Ev Idx": 4170 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121182.238, "dur": 5.280 , "args": { "External id": 4172,"Record function id": 0, "Ev Idx": 4171 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121190.724, "dur": 10.178 , "args": { "External id": 4173,"Record function id": 0, "Ev Idx": 4172 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121191.234, "dur": 0.672 , "args": { "External id": 4174,"Record function id": 0, "Ev Idx": 4173 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121191.405, "dur": 0.401 , "args": { "External id": 4175,"Record function id": 0, "Ev Idx": 4174 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121191.625, "dur": 0.080 , "args": { "External id": 4176,"Record function id": 0, "Ev Idx": 4175 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121192.166, "dur": 0.661 , "args": { "External id": 4177,"Record function id": 0, "Ev Idx": 4176 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121192.337, "dur": 0.400 , "args": { "External id": 4178,"Record function id": 0, "Ev Idx": 4177 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121192.567, "dur": 0.070 , "args": { "External id": 4179,"Record function id": 0, "Ev Idx": 4178 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121193.088, "dur": 0.621 , "args": { "External id": 4180,"Record function id": 0, "Ev Idx": 4179 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121193.238, "dur": 0.381 , "args": { "External id": 4181,"Record function id": 0, "Ev Idx": 4180 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121193.439, "dur": 0.070 , "args": { "External id": 4182,"Record function id": 0, "Ev Idx": 4181 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121194.791, "dur": 1.252 , "args": { "External id": 4183,"Record function id": 0, "Ev Idx": 4182 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121194.971, "dur": 0.962 , "args": { "External id": 4184,"Record function id": 0, "Ev Idx": 4183 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381121205.571, "dur": 4.649 , "args": { "External id": 4185,"Record function id": 0, "Ev Idx": 4184 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381121228.945, "dur": 22.121 , "args": { "External id": 4186,"Record function id": 0, "Ev Idx": 4185 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381121251.447, "dur": 342.337 , "args": { "External id": 4187,"Record function id": 0, "Ev Idx": 4186 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381121256.015, "dur": 10.259 , "args": { "External id": 4188,"Record function id": 0, "Ev Idx": 4187 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121259.331, "dur": 0.341 , "args": { "External id": 4189,"Record function id": 0, "Ev Idx": 4188 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381121260.113, "dur": 0.030 , "args": { "External id": 4190,"Record function id": 0, "Ev Idx": 4189 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381121260.513, "dur": 0.051 , "args": { "External id": 4191,"Record function id": 0, "Ev Idx": 4190 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121261.746, "dur": 0.250 , "args": { "External id": 4192,"Record function id": 0, "Ev Idx": 4191 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381121262.337, "dur": 0.040 , "args": { "External id": 4193,"Record function id": 0, "Ev Idx": 4192 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381121262.728, "dur": 0.020 , "args": { "External id": 4194,"Record function id": 0, "Ev Idx": 4193 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381121269.049, "dur": 1.132 , "args": { "External id": 4195,"Record function id": 0, "Ev Idx": 4194 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381121272.245, "dur": 318.593 , "args": { "External id": 4196,"Record function id": 0, "Ev Idx": 4195 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121280.891, "dur": 6.292 , "args": { "External id": 4197,"Record function id": 0, "Ev Idx": 4196 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121293.685, "dur": 7.023 , "args": { "External id": 4198,"Record function id": 0, "Ev Idx": 4197 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121306.669, "dur": 5.521 , "args": { "External id": 4199,"Record function id": 0, "Ev Idx": 4198 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121317.580, "dur": 5.199 , "args": { "External id": 4200,"Record function id": 0, "Ev Idx": 4199 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121325.374, "dur": 12.373 , "args": { "External id": 4201,"Record function id": 0, "Ev Idx": 4200 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121326.005, "dur": 0.982 , "args": { "External id": 4202,"Record function id": 0, "Ev Idx": 4201 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121326.226, "dur": 0.661 , "args": { "External id": 4203,"Record function id": 0, "Ev Idx": 4202 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121326.546, "dur": 0.211 , "args": { "External id": 4204,"Record function id": 0, "Ev Idx": 4203 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121327.298, "dur": 0.721 , "args": { "External id": 4205,"Record function id": 0, "Ev Idx": 4204 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121327.488, "dur": 0.441 , "args": { "External id": 4206,"Record function id": 0, "Ev Idx": 4205 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121327.749, "dur": 0.070 , "args": { "External id": 4207,"Record function id": 0, "Ev Idx": 4206 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121328.300, "dur": 0.631 , "args": { "External id": 4208,"Record function id": 0, "Ev Idx": 4207 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121328.460, "dur": 0.381 , "args": { "External id": 4209,"Record function id": 0, "Ev Idx": 4208 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121328.660, "dur": 0.070 , "args": { "External id": 4210,"Record function id": 0, "Ev Idx": 4209 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121330.313, "dur": 1.443 , "args": { "External id": 4211,"Record function id": 0, "Ev Idx": 4210 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121330.554, "dur": 1.092 , "args": { "External id": 4212,"Record function id": 0, "Ev Idx": 4211 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121343.147, "dur": 4.769 , "args": { "External id": 4213,"Record function id": 0, "Ev Idx": 4212 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121350.090, "dur": 0.962 , "args": { "External id": 4214,"Record function id": 0, "Ev Idx": 4213 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121352.795, "dur": 0.561 , "args": { "External id": 4215,"Record function id": 0, "Ev Idx": 4214 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381121357.193, "dur": 0.451 , "args": { "External id": 4216,"Record function id": 0, "Ev Idx": 4215 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381121357.474, "dur": 0.050 , "args": { "External id": 4217,"Record function id": 0, "Ev Idx": 4216 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381121360.610, "dur": 6.502 , "args": { "External id": 4218,"Record function id": 0, "Ev Idx": 4217 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121369.186, "dur": 10.890 , "args": { "External id": 4219,"Record function id": 0, "Ev Idx": 4218 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121369.727, "dur": 0.761 , "args": { "External id": 4220,"Record function id": 0, "Ev Idx": 4219 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121369.917, "dur": 0.471 , "args": { "External id": 4221,"Record function id": 0, "Ev Idx": 4220 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121370.178, "dur": 0.110 , "args": { "External id": 4222,"Record function id": 0, "Ev Idx": 4221 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121370.769, "dur": 0.621 , "args": { "External id": 4223,"Record function id": 0, "Ev Idx": 4222 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121370.939, "dur": 0.361 , "args": { "External id": 4224,"Record function id": 0, "Ev Idx": 4223 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121371.129, "dur": 0.070 , "args": { "External id": 4225,"Record function id": 0, "Ev Idx": 4224 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121371.650, "dur": 0.752 , "args": { "External id": 4226,"Record function id": 0, "Ev Idx": 4225 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121371.831, "dur": 0.481 , "args": { "External id": 4227,"Record function id": 0, "Ev Idx": 4226 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121372.021, "dur": 0.180 , "args": { "External id": 4228,"Record function id": 0, "Ev Idx": 4227 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121373.514, "dur": 1.302 , "args": { "External id": 4229,"Record function id": 0, "Ev Idx": 4228 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121373.704, "dur": 1.012 , "args": { "External id": 4230,"Record function id": 0, "Ev Idx": 4229 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121383.352, "dur": 4.098 , "args": { "External id": 4231,"Record function id": 0, "Ev Idx": 4230 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121388.542, "dur": 0.631 , "args": { "External id": 4232,"Record function id": 0, "Ev Idx": 4231 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121390.545, "dur": 0.582 , "args": { "External id": 4233,"Record function id": 0, "Ev Idx": 4232 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381121394.803, "dur": 0.351 , "args": { "External id": 4234,"Record function id": 0, "Ev Idx": 4233 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381121395.004, "dur": 0.050 , "args": { "External id": 4235,"Record function id": 0, "Ev Idx": 4234 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381121397.869, "dur": 5.681 , "args": { "External id": 4236,"Record function id": 0, "Ev Idx": 4235 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121405.784, "dur": 10.369 , "args": { "External id": 4237,"Record function id": 0, "Ev Idx": 4236 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121406.325, "dur": 0.791 , "args": { "External id": 4238,"Record function id": 0, "Ev Idx": 4237 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121406.515, "dur": 0.511 , "args": { "External id": 4239,"Record function id": 0, "Ev Idx": 4238 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121406.826, "dur": 0.100 , "args": { "External id": 4240,"Record function id": 0, "Ev Idx": 4239 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121407.397, "dur": 0.631 , "args": { "External id": 4241,"Record function id": 0, "Ev Idx": 4240 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121407.577, "dur": 0.361 , "args": { "External id": 4242,"Record function id": 0, "Ev Idx": 4241 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121407.758, "dur": 0.080 , "args": { "External id": 4243,"Record function id": 0, "Ev Idx": 4242 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121408.278, "dur": 0.702 , "args": { "External id": 4244,"Record function id": 0, "Ev Idx": 4243 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121408.439, "dur": 0.451 , "args": { "External id": 4245,"Record function id": 0, "Ev Idx": 4244 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121408.629, "dur": 0.150 , "args": { "External id": 4246,"Record function id": 0, "Ev Idx": 4245 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121410.042, "dur": 1.282 , "args": { "External id": 4247,"Record function id": 0, "Ev Idx": 4246 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121410.232, "dur": 0.992 , "args": { "External id": 4248,"Record function id": 0, "Ev Idx": 4247 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381121420.631, "dur": 4.298 , "args": { "External id": 4249,"Record function id": 0, "Ev Idx": 4248 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381121430.239, "dur": 4.859 , "args": { "External id": 4250,"Record function id": 0, "Ev Idx": 4249 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381121440.148, "dur": 4.548 , "args": { "External id": 4251,"Record function id": 0, "Ev Idx": 4250 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381121446.750, "dur": 17.122 , "args": { "External id": 4252,"Record function id": 0, "Ev Idx": 4251 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121447.241, "dur": 0.511 , "args": { "External id": 4253,"Record function id": 0, "Ev Idx": 4252 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121447.512, "dur": 0.120 , "args": { "External id": 4254,"Record function id": 0, "Ev Idx": 4253 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121447.992, "dur": 0.411 , "args": { "External id": 4255,"Record function id": 0, "Ev Idx": 4254 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121448.203, "dur": 0.090 , "args": { "External id": 4256,"Record function id": 0, "Ev Idx": 4255 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121448.634, "dur": 0.390 , "args": { "External id": 4257,"Record function id": 0, "Ev Idx": 4256 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121448.824, "dur": 0.090 , "args": { "External id": 4258,"Record function id": 0, "Ev Idx": 4257 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381121449.385, "dur": 13.034 , "args": { "External id": 4259,"Record function id": 0, "Ev Idx": 4258 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381121450.237, "dur": 1.833 , "args": { "External id": 4260,"Record function id": 0, "Ev Idx": 4259 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121450.517, "dur": 1.443 , "args": { "External id": 4261,"Record function id": 0, "Ev Idx": 4260 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121452.521, "dur": 1.262 , "args": { "External id": 4262,"Record function id": 0, "Ev Idx": 4261 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121454.224, "dur": 1.222 , "args": { "External id": 4263,"Record function id": 0, "Ev Idx": 4262 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121455.727, "dur": 1.402 , "args": { "External id": 4264,"Record function id": 0, "Ev Idx": 4263 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121461.127, "dur": 0.942 , "args": { "External id": 4265,"Record function id": 0, "Ev Idx": 4264 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121462.710, "dur": 0.461 , "args": { "External id": 4266,"Record function id": 0, "Ev Idx": 4265 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121462.920, "dur": 0.140 , "args": { "External id": 4267,"Record function id": 0, "Ev Idx": 4266 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121470.494, "dur": 6.462 , "args": { "External id": 4268,"Record function id": 0, "Ev Idx": 4267 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121479.631, "dur": 10.580 , "args": { "External id": 4269,"Record function id": 0, "Ev Idx": 4268 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121480.202, "dur": 0.702 , "args": { "External id": 4270,"Record function id": 0, "Ev Idx": 4269 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121480.393, "dur": 0.420 , "args": { "External id": 4271,"Record function id": 0, "Ev Idx": 4270 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121480.613, "dur": 0.080 , "args": { "External id": 4272,"Record function id": 0, "Ev Idx": 4271 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121481.174, "dur": 0.631 , "args": { "External id": 4273,"Record function id": 0, "Ev Idx": 4272 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121481.324, "dur": 0.401 , "args": { "External id": 4274,"Record function id": 0, "Ev Idx": 4273 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121481.535, "dur": 0.080 , "args": { "External id": 4275,"Record function id": 0, "Ev Idx": 4274 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121482.066, "dur": 0.641 , "args": { "External id": 4276,"Record function id": 0, "Ev Idx": 4275 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121482.226, "dur": 0.391 , "args": { "External id": 4277,"Record function id": 0, "Ev Idx": 4276 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121482.437, "dur": 0.070 , "args": { "External id": 4278,"Record function id": 0, "Ev Idx": 4277 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121483.829, "dur": 1.232 , "args": { "External id": 4279,"Record function id": 0, "Ev Idx": 4278 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121483.999, "dur": 0.952 , "args": { "External id": 4280,"Record function id": 0, "Ev Idx": 4279 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381121494.269, "dur": 5.299 , "args": { "External id": 4281,"Record function id": 0, "Ev Idx": 4280 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121502.454, "dur": 5.450 , "args": { "External id": 4282,"Record function id": 0, "Ev Idx": 4281 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121511.120, "dur": 5.119 , "args": { "External id": 4283,"Record function id": 0, "Ev Idx": 4282 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121519.716, "dur": 11.201 , "args": { "External id": 4284,"Record function id": 0, "Ev Idx": 4283 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121520.227, "dur": 0.681 , "args": { "External id": 4285,"Record function id": 0, "Ev Idx": 4284 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121520.387, "dur": 0.431 , "args": { "External id": 4286,"Record function id": 0, "Ev Idx": 4285 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121520.608, "dur": 0.090 , "args": { "External id": 4287,"Record function id": 0, "Ev Idx": 4286 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121521.169, "dur": 0.661 , "args": { "External id": 4288,"Record function id": 0, "Ev Idx": 4287 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121521.339, "dur": 0.401 , "args": { "External id": 4289,"Record function id": 0, "Ev Idx": 4288 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121521.569, "dur": 0.070 , "args": { "External id": 4290,"Record function id": 0, "Ev Idx": 4289 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121522.080, "dur": 0.641 , "args": { "External id": 4291,"Record function id": 0, "Ev Idx": 4290 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121522.251, "dur": 0.380 , "args": { "External id": 4292,"Record function id": 0, "Ev Idx": 4291 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121522.451, "dur": 0.080 , "args": { "External id": 4293,"Record function id": 0, "Ev Idx": 4292 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121523.773, "dur": 1.263 , "args": { "External id": 4294,"Record function id": 0, "Ev Idx": 4293 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121523.954, "dur": 0.982 , "args": { "External id": 4295,"Record function id": 0, "Ev Idx": 4294 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121534.584, "dur": 13.074 , "args": { "External id": 4296,"Record function id": 0, "Ev Idx": 4295 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121535.095, "dur": 0.661 , "args": { "External id": 4297,"Record function id": 0, "Ev Idx": 4296 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121535.275, "dur": 0.381 , "args": { "External id": 4298,"Record function id": 0, "Ev Idx": 4297 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121535.485, "dur": 0.070 , "args": { "External id": 4299,"Record function id": 0, "Ev Idx": 4298 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121536.006, "dur": 0.621 , "args": { "External id": 4300,"Record function id": 0, "Ev Idx": 4299 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121536.167, "dur": 0.370 , "args": { "External id": 4301,"Record function id": 0, "Ev Idx": 4300 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121536.357, "dur": 0.070 , "args": { "External id": 4302,"Record function id": 0, "Ev Idx": 4301 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121536.878, "dur": 3.637 , "args": { "External id": 4303,"Record function id": 0, "Ev Idx": 4302 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121537.038, "dur": 3.356 , "args": { "External id": 4304,"Record function id": 0, "Ev Idx": 4303 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121540.144, "dur": 0.110 , "args": { "External id": 4305,"Record function id": 0, "Ev Idx": 4304 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121541.547, "dur": 1.222 , "args": { "External id": 4306,"Record function id": 0, "Ev Idx": 4305 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121541.717, "dur": 0.942 , "args": { "External id": 4307,"Record function id": 0, "Ev Idx": 4306 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381121552.527, "dur": 4.358 , "args": { "External id": 4308,"Record function id": 0, "Ev Idx": 4307 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121560.091, "dur": 5.530 , "args": { "External id": 4309,"Record function id": 0, "Ev Idx": 4308 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121569.068, "dur": 10.439 , "args": { "External id": 4310,"Record function id": 0, "Ev Idx": 4309 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121569.569, "dur": 0.721 , "args": { "External id": 4311,"Record function id": 0, "Ev Idx": 4310 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121569.759, "dur": 0.441 , "args": { "External id": 4312,"Record function id": 0, "Ev Idx": 4311 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121570.000, "dur": 0.080 , "args": { "External id": 4313,"Record function id": 0, "Ev Idx": 4312 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121570.551, "dur": 0.681 , "args": { "External id": 4314,"Record function id": 0, "Ev Idx": 4313 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121570.711, "dur": 0.431 , "args": { "External id": 4315,"Record function id": 0, "Ev Idx": 4314 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121570.961, "dur": 0.080 , "args": { "External id": 4316,"Record function id": 0, "Ev Idx": 4315 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121571.482, "dur": 0.621 , "args": { "External id": 4317,"Record function id": 0, "Ev Idx": 4316 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121571.643, "dur": 0.360 , "args": { "External id": 4318,"Record function id": 0, "Ev Idx": 4317 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121571.833, "dur": 0.070 , "args": { "External id": 4319,"Record function id": 0, "Ev Idx": 4318 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121573.115, "dur": 1.253 , "args": { "External id": 4320,"Record function id": 0, "Ev Idx": 4319 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121573.296, "dur": 0.971 , "args": { "External id": 4321,"Record function id": 0, "Ev Idx": 4320 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381121584.336, "dur": 5.020 , "args": { "External id": 4322,"Record function id": 0, "Ev Idx": 4321 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381121608.110, "dur": 23.624 , "args": { "External id": 4323,"Record function id": 0, "Ev Idx": 4322 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381121632.105, "dur": 338.480 , "args": { "External id": 4324,"Record function id": 0, "Ev Idx": 4323 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381121636.784, "dur": 10.650 , "args": { "External id": 4325,"Record function id": 0, "Ev Idx": 4324 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121640.250, "dur": 0.451 , "args": { "External id": 4326,"Record function id": 0, "Ev Idx": 4325 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381121641.192, "dur": 0.040 , "args": { "External id": 4327,"Record function id": 0, "Ev Idx": 4326 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381121641.623, "dur": 0.060 , "args": { "External id": 4328,"Record function id": 0, "Ev Idx": 4327 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121642.915, "dur": 0.281 , "args": { "External id": 4329,"Record function id": 0, "Ev Idx": 4328 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381121643.556, "dur": 0.030 , "args": { "External id": 4330,"Record function id": 0, "Ev Idx": 4329 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381121643.887, "dur": 0.030 , "args": { "External id": 4331,"Record function id": 0, "Ev Idx": 4330 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381121650.289, "dur": 1.122 , "args": { "External id": 4332,"Record function id": 0, "Ev Idx": 4331 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381121653.475, "dur": 314.225 , "args": { "External id": 4333,"Record function id": 0, "Ev Idx": 4332 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121662.091, "dur": 6.322 , "args": { "External id": 4334,"Record function id": 0, "Ev Idx": 4333 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121674.805, "dur": 7.053 , "args": { "External id": 4335,"Record function id": 0, "Ev Idx": 4334 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121687.819, "dur": 5.510 , "args": { "External id": 4336,"Record function id": 0, "Ev Idx": 4335 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121698.769, "dur": 5.150 , "args": { "External id": 4337,"Record function id": 0, "Ev Idx": 4336 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121706.584, "dur": 12.423 , "args": { "External id": 4338,"Record function id": 0, "Ev Idx": 4337 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121707.235, "dur": 0.912 , "args": { "External id": 4339,"Record function id": 0, "Ev Idx": 4338 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121707.445, "dur": 0.591 , "args": { "External id": 4340,"Record function id": 0, "Ev Idx": 4339 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121707.726, "dur": 0.170 , "args": { "External id": 4341,"Record function id": 0, "Ev Idx": 4340 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121708.467, "dur": 0.641 , "args": { "External id": 4342,"Record function id": 0, "Ev Idx": 4341 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121708.638, "dur": 0.380 , "args": { "External id": 4343,"Record function id": 0, "Ev Idx": 4342 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121708.838, "dur": 0.070 , "args": { "External id": 4344,"Record function id": 0, "Ev Idx": 4343 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121709.389, "dur": 0.701 , "args": { "External id": 4345,"Record function id": 0, "Ev Idx": 4344 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121709.539, "dur": 0.461 , "args": { "External id": 4346,"Record function id": 0, "Ev Idx": 4345 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121709.740, "dur": 0.150 , "args": { "External id": 4347,"Record function id": 0, "Ev Idx": 4346 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121711.363, "dur": 1.462 , "args": { "External id": 4348,"Record function id": 0, "Ev Idx": 4347 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121711.563, "dur": 1.132 , "args": { "External id": 4349,"Record function id": 0, "Ev Idx": 4348 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121724.567, "dur": 4.589 , "args": { "External id": 4350,"Record function id": 0, "Ev Idx": 4349 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121731.300, "dur": 0.962 , "args": { "External id": 4351,"Record function id": 0, "Ev Idx": 4350 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121734.095, "dur": 0.571 , "args": { "External id": 4352,"Record function id": 0, "Ev Idx": 4351 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381121738.403, "dur": 0.501 , "args": { "External id": 4353,"Record function id": 0, "Ev Idx": 4352 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381121738.714, "dur": 0.080 , "args": { "External id": 4354,"Record function id": 0, "Ev Idx": 4353 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381121741.729, "dur": 6.292 , "args": { "External id": 4355,"Record function id": 0, "Ev Idx": 4354 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121750.075, "dur": 10.920 , "args": { "External id": 4356,"Record function id": 0, "Ev Idx": 4355 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121750.646, "dur": 0.721 , "args": { "External id": 4357,"Record function id": 0, "Ev Idx": 4356 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121750.816, "dur": 0.461 , "args": { "External id": 4358,"Record function id": 0, "Ev Idx": 4357 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121751.067, "dur": 0.100 , "args": { "External id": 4359,"Record function id": 0, "Ev Idx": 4358 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121751.638, "dur": 0.631 , "args": { "External id": 4360,"Record function id": 0, "Ev Idx": 4359 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121751.798, "dur": 0.381 , "args": { "External id": 4361,"Record function id": 0, "Ev Idx": 4360 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121751.998, "dur": 0.070 , "args": { "External id": 4362,"Record function id": 0, "Ev Idx": 4361 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121752.519, "dur": 0.712 , "args": { "External id": 4363,"Record function id": 0, "Ev Idx": 4362 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121752.690, "dur": 0.450 , "args": { "External id": 4364,"Record function id": 0, "Ev Idx": 4363 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121752.880, "dur": 0.160 , "args": { "External id": 4365,"Record function id": 0, "Ev Idx": 4364 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121754.333, "dur": 1.282 , "args": { "External id": 4366,"Record function id": 0, "Ev Idx": 4365 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121754.533, "dur": 0.972 , "args": { "External id": 4367,"Record function id": 0, "Ev Idx": 4366 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121764.472, "dur": 3.957 , "args": { "External id": 4368,"Record function id": 0, "Ev Idx": 4367 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121769.451, "dur": 0.591 , "args": { "External id": 4369,"Record function id": 0, "Ev Idx": 4368 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381121771.364, "dur": 0.571 , "args": { "External id": 4370,"Record function id": 0, "Ev Idx": 4369 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381121775.763, "dur": 0.320 , "args": { "External id": 4371,"Record function id": 0, "Ev Idx": 4370 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381121775.943, "dur": 0.040 , "args": { "External id": 4372,"Record function id": 0, "Ev Idx": 4371 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381121778.838, "dur": 5.871 , "args": { "External id": 4373,"Record function id": 0, "Ev Idx": 4372 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121786.953, "dur": 10.430 , "args": { "External id": 4374,"Record function id": 0, "Ev Idx": 4373 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121787.474, "dur": 0.762 , "args": { "External id": 4375,"Record function id": 0, "Ev Idx": 4374 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121787.665, "dur": 0.481 , "args": { "External id": 4376,"Record function id": 0, "Ev Idx": 4375 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121787.955, "dur": 0.080 , "args": { "External id": 4377,"Record function id": 0, "Ev Idx": 4376 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121788.526, "dur": 0.611 , "args": { "External id": 4378,"Record function id": 0, "Ev Idx": 4377 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121788.677, "dur": 0.370 , "args": { "External id": 4379,"Record function id": 0, "Ev Idx": 4378 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121788.867, "dur": 0.070 , "args": { "External id": 4380,"Record function id": 0, "Ev Idx": 4379 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121789.388, "dur": 0.631 , "args": { "External id": 4381,"Record function id": 0, "Ev Idx": 4380 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121789.548, "dur": 0.381 , "args": { "External id": 4382,"Record function id": 0, "Ev Idx": 4381 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121789.739, "dur": 0.080 , "args": { "External id": 4383,"Record function id": 0, "Ev Idx": 4382 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121791.081, "dur": 1.292 , "args": { "External id": 4384,"Record function id": 0, "Ev Idx": 4383 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121791.261, "dur": 1.012 , "args": { "External id": 4385,"Record function id": 0, "Ev Idx": 4384 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381121802.001, "dur": 4.258 , "args": { "External id": 4386,"Record function id": 0, "Ev Idx": 4385 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381121811.599, "dur": 4.829 , "args": { "External id": 4387,"Record function id": 0, "Ev Idx": 4386 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381121821.397, "dur": 4.799 , "args": { "External id": 4388,"Record function id": 0, "Ev Idx": 4387 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381121828.441, "dur": 17.813 , "args": { "External id": 4389,"Record function id": 0, "Ev Idx": 4388 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121828.942, "dur": 0.480 , "args": { "External id": 4390,"Record function id": 0, "Ev Idx": 4389 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121829.172, "dur": 0.140 , "args": { "External id": 4391,"Record function id": 0, "Ev Idx": 4390 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121829.653, "dur": 0.411 , "args": { "External id": 4392,"Record function id": 0, "Ev Idx": 4391 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121829.863, "dur": 0.090 , "args": { "External id": 4393,"Record function id": 0, "Ev Idx": 4392 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121830.304, "dur": 0.381 , "args": { "External id": 4394,"Record function id": 0, "Ev Idx": 4393 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121830.494, "dur": 0.091 , "args": { "External id": 4395,"Record function id": 0, "Ev Idx": 4394 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381121831.015, "dur": 13.706 , "args": { "External id": 4396,"Record function id": 0, "Ev Idx": 4395 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381121831.997, "dur": 1.954 , "args": { "External id": 4397,"Record function id": 0, "Ev Idx": 4396 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121832.278, "dur": 1.563 , "args": { "External id": 4398,"Record function id": 0, "Ev Idx": 4397 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121834.422, "dur": 1.242 , "args": { "External id": 4399,"Record function id": 0, "Ev Idx": 4398 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121836.105, "dur": 1.272 , "args": { "External id": 4400,"Record function id": 0, "Ev Idx": 4399 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121837.658, "dur": 1.452 , "args": { "External id": 4401,"Record function id": 0, "Ev Idx": 4400 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121843.348, "dur": 1.022 , "args": { "External id": 4402,"Record function id": 0, "Ev Idx": 4401 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121845.031, "dur": 0.441 , "args": { "External id": 4403,"Record function id": 0, "Ev Idx": 4402 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121845.242, "dur": 0.130 , "args": { "External id": 4404,"Record function id": 0, "Ev Idx": 4403 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121852.916, "dur": 6.292 , "args": { "External id": 4405,"Record function id": 0, "Ev Idx": 4404 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121861.622, "dur": 10.730 , "args": { "External id": 4406,"Record function id": 0, "Ev Idx": 4405 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121862.203, "dur": 0.712 , "args": { "External id": 4407,"Record function id": 0, "Ev Idx": 4406 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121862.384, "dur": 0.441 , "args": { "External id": 4408,"Record function id": 0, "Ev Idx": 4407 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121862.614, "dur": 0.080 , "args": { "External id": 4409,"Record function id": 0, "Ev Idx": 4408 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121863.205, "dur": 0.682 , "args": { "External id": 4410,"Record function id": 0, "Ev Idx": 4409 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121863.386, "dur": 0.410 , "args": { "External id": 4411,"Record function id": 0, "Ev Idx": 4410 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121863.606, "dur": 0.080 , "args": { "External id": 4412,"Record function id": 0, "Ev Idx": 4411 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121864.147, "dur": 0.641 , "args": { "External id": 4413,"Record function id": 0, "Ev Idx": 4412 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121864.307, "dur": 0.401 , "args": { "External id": 4414,"Record function id": 0, "Ev Idx": 4413 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121864.518, "dur": 0.080 , "args": { "External id": 4415,"Record function id": 0, "Ev Idx": 4414 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121865.950, "dur": 1.323 , "args": { "External id": 4416,"Record function id": 0, "Ev Idx": 4415 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121866.131, "dur": 1.032 , "args": { "External id": 4417,"Record function id": 0, "Ev Idx": 4416 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381121876.280, "dur": 5.199 , "args": { "External id": 4418,"Record function id": 0, "Ev Idx": 4417 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121884.244, "dur": 5.521 , "args": { "External id": 4419,"Record function id": 0, "Ev Idx": 4418 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381121892.971, "dur": 4.999 , "args": { "External id": 4420,"Record function id": 0, "Ev Idx": 4419 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121901.446, "dur": 10.520 , "args": { "External id": 4421,"Record function id": 0, "Ev Idx": 4420 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121901.947, "dur": 0.682 , "args": { "External id": 4422,"Record function id": 0, "Ev Idx": 4421 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121902.128, "dur": 0.400 , "args": { "External id": 4423,"Record function id": 0, "Ev Idx": 4422 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121902.338, "dur": 0.080 , "args": { "External id": 4424,"Record function id": 0, "Ev Idx": 4423 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121902.909, "dur": 0.621 , "args": { "External id": 4425,"Record function id": 0, "Ev Idx": 4424 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121903.059, "dur": 0.381 , "args": { "External id": 4426,"Record function id": 0, "Ev Idx": 4425 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121903.260, "dur": 0.070 , "args": { "External id": 4427,"Record function id": 0, "Ev Idx": 4426 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121903.791, "dur": 0.651 , "args": { "External id": 4428,"Record function id": 0, "Ev Idx": 4427 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121903.951, "dur": 0.401 , "args": { "External id": 4429,"Record function id": 0, "Ev Idx": 4428 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121904.162, "dur": 0.090 , "args": { "External id": 4430,"Record function id": 0, "Ev Idx": 4429 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121905.464, "dur": 1.262 , "args": { "External id": 4431,"Record function id": 0, "Ev Idx": 4430 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121905.654, "dur": 0.962 , "args": { "External id": 4432,"Record function id": 0, "Ev Idx": 4431 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121915.593, "dur": 9.988 , "args": { "External id": 4433,"Record function id": 0, "Ev Idx": 4432 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121916.114, "dur": 0.641 , "args": { "External id": 4434,"Record function id": 0, "Ev Idx": 4433 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121916.284, "dur": 0.381 , "args": { "External id": 4435,"Record function id": 0, "Ev Idx": 4434 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121916.484, "dur": 0.071 , "args": { "External id": 4436,"Record function id": 0, "Ev Idx": 4435 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121917.025, "dur": 0.622 , "args": { "External id": 4437,"Record function id": 0, "Ev Idx": 4436 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121917.196, "dur": 0.360 , "args": { "External id": 4438,"Record function id": 0, "Ev Idx": 4437 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121917.386, "dur": 0.070 , "args": { "External id": 4439,"Record function id": 0, "Ev Idx": 4438 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121917.917, "dur": 0.621 , "args": { "External id": 4440,"Record function id": 0, "Ev Idx": 4439 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121918.077, "dur": 0.371 , "args": { "External id": 4441,"Record function id": 0, "Ev Idx": 4440 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121918.268, "dur": 0.080 , "args": { "External id": 4442,"Record function id": 0, "Ev Idx": 4441 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121919.530, "dur": 1.232 , "args": { "External id": 4443,"Record function id": 0, "Ev Idx": 4442 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121919.700, "dur": 0.962 , "args": { "External id": 4444,"Record function id": 0, "Ev Idx": 4443 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381121930.200, "dur": 4.318 , "args": { "External id": 4445,"Record function id": 0, "Ev Idx": 4444 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381121937.754, "dur": 5.661 , "args": { "External id": 4446,"Record function id": 0, "Ev Idx": 4445 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381121946.601, "dur": 10.239 , "args": { "External id": 4447,"Record function id": 0, "Ev Idx": 4446 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121947.101, "dur": 0.652 , "args": { "External id": 4448,"Record function id": 0, "Ev Idx": 4447 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121947.272, "dur": 0.390 , "args": { "External id": 4449,"Record function id": 0, "Ev Idx": 4448 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121947.482, "dur": 0.080 , "args": { "External id": 4450,"Record function id": 0, "Ev Idx": 4449 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121948.023, "dur": 0.681 , "args": { "External id": 4451,"Record function id": 0, "Ev Idx": 4450 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121948.193, "dur": 0.421 , "args": { "External id": 4452,"Record function id": 0, "Ev Idx": 4451 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121948.424, "dur": 0.080 , "args": { "External id": 4453,"Record function id": 0, "Ev Idx": 4452 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381121948.955, "dur": 0.641 , "args": { "External id": 4454,"Record function id": 0, "Ev Idx": 4453 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381121949.125, "dur": 0.381 , "args": { "External id": 4455,"Record function id": 0, "Ev Idx": 4454 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381121949.326, "dur": 0.080 , "args": { "External id": 4456,"Record function id": 0, "Ev Idx": 4455 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121950.658, "dur": 1.282 , "args": { "External id": 4457,"Record function id": 0, "Ev Idx": 4456 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381121950.848, "dur": 1.002 , "args": { "External id": 4458,"Record function id": 0, "Ev Idx": 4457 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381121961.428, "dur": 4.759 , "args": { "External id": 4459,"Record function id": 0, "Ev Idx": 4458 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381121985.623, "dur": 23.133 , "args": { "External id": 4460,"Record function id": 0, "Ev Idx": 4459 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381122009.137, "dur": 334.473 , "args": { "External id": 4461,"Record function id": 0, "Ev Idx": 4460 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381122013.665, "dur": 10.530 , "args": { "External id": 4462,"Record function id": 0, "Ev Idx": 4461 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122017.042, "dur": 0.380 , "args": { "External id": 4463,"Record function id": 0, "Ev Idx": 4462 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381122017.923, "dur": 0.040 , "args": { "External id": 4464,"Record function id": 0, "Ev Idx": 4463 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381122018.334, "dur": 0.040 , "args": { "External id": 4465,"Record function id": 0, "Ev Idx": 4464 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122019.667, "dur": 0.280 , "args": { "External id": 4466,"Record function id": 0, "Ev Idx": 4465 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381122020.298, "dur": 0.030 , "args": { "External id": 4467,"Record function id": 0, "Ev Idx": 4466 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381122020.638, "dur": 0.030 , "args": { "External id": 4468,"Record function id": 0, "Ev Idx": 4467 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381122027.030, "dur": 1.122 , "args": { "External id": 4469,"Record function id": 0, "Ev Idx": 4468 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381122030.156, "dur": 310.578 , "args": { "External id": 4470,"Record function id": 0, "Ev Idx": 4469 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122038.632, "dur": 6.302 , "args": { "External id": 4471,"Record function id": 0, "Ev Idx": 4470 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122051.285, "dur": 7.114 , "args": { "External id": 4472,"Record function id": 0, "Ev Idx": 4471 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122064.179, "dur": 5.911 , "args": { "External id": 4473,"Record function id": 0, "Ev Idx": 4472 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122075.420, "dur": 4.960 , "args": { "External id": 4474,"Record function id": 0, "Ev Idx": 4473 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122082.934, "dur": 12.213 , "args": { "External id": 4475,"Record function id": 0, "Ev Idx": 4474 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122083.555, "dur": 1.022 , "args": { "External id": 4476,"Record function id": 0, "Ev Idx": 4475 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122083.806, "dur": 0.661 , "args": { "External id": 4477,"Record function id": 0, "Ev Idx": 4476 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122084.137, "dur": 0.190 , "args": { "External id": 4478,"Record function id": 0, "Ev Idx": 4477 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122084.888, "dur": 0.611 , "args": { "External id": 4479,"Record function id": 0, "Ev Idx": 4478 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122085.048, "dur": 0.361 , "args": { "External id": 4480,"Record function id": 0, "Ev Idx": 4479 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122085.229, "dur": 0.080 , "args": { "External id": 4481,"Record function id": 0, "Ev Idx": 4480 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122085.770, "dur": 0.711 , "args": { "External id": 4482,"Record function id": 0, "Ev Idx": 4481 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122085.940, "dur": 0.451 , "args": { "External id": 4483,"Record function id": 0, "Ev Idx": 4482 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122086.130, "dur": 0.151 , "args": { "External id": 4484,"Record function id": 0, "Ev Idx": 4483 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122087.763, "dur": 1.433 , "args": { "External id": 4485,"Record function id": 0, "Ev Idx": 4484 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122087.994, "dur": 1.092 , "args": { "External id": 4486,"Record function id": 0, "Ev Idx": 4485 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381122100.487, "dur": 4.418 , "args": { "External id": 4487,"Record function id": 0, "Ev Idx": 4486 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122107.159, "dur": 0.862 , "args": { "External id": 4488,"Record function id": 0, "Ev Idx": 4487 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122109.754, "dur": 0.561 , "args": { "External id": 4489,"Record function id": 0, "Ev Idx": 4488 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381122114.002, "dur": 0.491 , "args": { "External id": 4490,"Record function id": 0, "Ev Idx": 4489 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381122114.283, "dur": 0.090 , "args": { "External id": 4491,"Record function id": 0, "Ev Idx": 4490 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381122117.529, "dur": 6.071 , "args": { "External id": 4492,"Record function id": 0, "Ev Idx": 4491 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122125.784, "dur": 10.980 , "args": { "External id": 4493,"Record function id": 0, "Ev Idx": 4492 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122126.315, "dur": 0.782 , "args": { "External id": 4494,"Record function id": 0, "Ev Idx": 4493 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122126.495, "dur": 0.511 , "args": { "External id": 4495,"Record function id": 0, "Ev Idx": 4494 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122126.766, "dur": 0.110 , "args": { "External id": 4496,"Record function id": 0, "Ev Idx": 4495 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122127.367, "dur": 0.631 , "args": { "External id": 4497,"Record function id": 0, "Ev Idx": 4496 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122127.527, "dur": 0.381 , "args": { "External id": 4498,"Record function id": 0, "Ev Idx": 4497 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122127.718, "dur": 0.090 , "args": { "External id": 4499,"Record function id": 0, "Ev Idx": 4498 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122128.259, "dur": 1.002 , "args": { "External id": 4500,"Record function id": 0, "Ev Idx": 4499 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122128.419, "dur": 0.751 , "args": { "External id": 4501,"Record function id": 0, "Ev Idx": 4500 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122128.689, "dur": 0.371 , "args": { "External id": 4502,"Record function id": 0, "Ev Idx": 4501 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122130.333, "dur": 1.302 , "args": { "External id": 4503,"Record function id": 0, "Ev Idx": 4502 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122130.533, "dur": 0.992 , "args": { "External id": 4504,"Record function id": 0, "Ev Idx": 4503 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381122140.181, "dur": 3.667 , "args": { "External id": 4505,"Record function id": 0, "Ev Idx": 4504 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122144.910, "dur": 0.661 , "args": { "External id": 4506,"Record function id": 0, "Ev Idx": 4505 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122146.903, "dur": 0.581 , "args": { "External id": 4507,"Record function id": 0, "Ev Idx": 4506 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381122151.352, "dur": 0.310 , "args": { "External id": 4508,"Record function id": 0, "Ev Idx": 4507 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381122151.532, "dur": 0.030 , "args": { "External id": 4509,"Record function id": 0, "Ev Idx": 4508 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381122154.207, "dur": 5.560 , "args": { "External id": 4510,"Record function id": 0, "Ev Idx": 4509 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122161.891, "dur": 10.430 , "args": { "External id": 4511,"Record function id": 0, "Ev Idx": 4510 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122162.412, "dur": 0.722 , "args": { "External id": 4512,"Record function id": 0, "Ev Idx": 4511 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122162.613, "dur": 0.430 , "args": { "External id": 4513,"Record function id": 0, "Ev Idx": 4512 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122162.843, "dur": 0.090 , "args": { "External id": 4514,"Record function id": 0, "Ev Idx": 4513 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122163.404, "dur": 0.621 , "args": { "External id": 4515,"Record function id": 0, "Ev Idx": 4514 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122163.564, "dur": 0.371 , "args": { "External id": 4516,"Record function id": 0, "Ev Idx": 4515 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122163.755, "dur": 0.070 , "args": { "External id": 4517,"Record function id": 0, "Ev Idx": 4516 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122164.286, "dur": 0.631 , "args": { "External id": 4518,"Record function id": 0, "Ev Idx": 4517 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122164.436, "dur": 0.391 , "args": { "External id": 4519,"Record function id": 0, "Ev Idx": 4518 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122164.626, "dur": 0.091 , "args": { "External id": 4520,"Record function id": 0, "Ev Idx": 4519 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122165.979, "dur": 1.262 , "args": { "External id": 4521,"Record function id": 0, "Ev Idx": 4520 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122166.159, "dur": 0.972 , "args": { "External id": 4522,"Record function id": 0, "Ev Idx": 4521 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381122176.789, "dur": 4.208 , "args": { "External id": 4523,"Record function id": 0, "Ev Idx": 4522 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381122186.186, "dur": 4.579 , "args": { "External id": 4524,"Record function id": 0, "Ev Idx": 4523 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381122195.754, "dur": 4.388 , "args": { "External id": 4525,"Record function id": 0, "Ev Idx": 4524 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381122202.317, "dur": 17.853 , "args": { "External id": 4526,"Record function id": 0, "Ev Idx": 4525 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122202.787, "dur": 0.672 , "args": { "External id": 4527,"Record function id": 0, "Ev Idx": 4526 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122203.128, "dur": 0.220 , "args": { "External id": 4528,"Record function id": 0, "Ev Idx": 4527 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122203.699, "dur": 0.401 , "args": { "External id": 4529,"Record function id": 0, "Ev Idx": 4528 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122203.889, "dur": 0.101 , "args": { "External id": 4530,"Record function id": 0, "Ev Idx": 4529 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122204.350, "dur": 0.401 , "args": { "External id": 4531,"Record function id": 0, "Ev Idx": 4530 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122204.561, "dur": 0.080 , "args": { "External id": 4532,"Record function id": 0, "Ev Idx": 4531 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381122205.092, "dur": 13.555 , "args": { "External id": 4533,"Record function id": 0, "Ev Idx": 4532 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381122206.003, "dur": 1.904 , "args": { "External id": 4534,"Record function id": 0, "Ev Idx": 4533 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122206.364, "dur": 1.423 , "args": { "External id": 4535,"Record function id": 0, "Ev Idx": 4534 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122208.348, "dur": 1.202 , "args": { "External id": 4536,"Record function id": 0, "Ev Idx": 4535 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122210.001, "dur": 1.342 , "args": { "External id": 4537,"Record function id": 0, "Ev Idx": 4536 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122211.634, "dur": 1.372 , "args": { "External id": 4538,"Record function id": 0, "Ev Idx": 4537 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122217.314, "dur": 1.012 , "args": { "External id": 4539,"Record function id": 0, "Ev Idx": 4538 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122218.947, "dur": 0.461 , "args": { "External id": 4540,"Record function id": 0, "Ev Idx": 4539 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122219.158, "dur": 0.130 , "args": { "External id": 4541,"Record function id": 0, "Ev Idx": 4540 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122226.782, "dur": 6.472 , "args": { "External id": 4542,"Record function id": 0, "Ev Idx": 4541 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122235.608, "dur": 10.630 , "args": { "External id": 4543,"Record function id": 0, "Ev Idx": 4542 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122236.200, "dur": 0.701 , "args": { "External id": 4544,"Record function id": 0, "Ev Idx": 4543 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122236.380, "dur": 0.431 , "args": { "External id": 4545,"Record function id": 0, "Ev Idx": 4544 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122236.600, "dur": 0.111 , "args": { "External id": 4546,"Record function id": 0, "Ev Idx": 4545 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122237.171, "dur": 0.642 , "args": { "External id": 4547,"Record function id": 0, "Ev Idx": 4546 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122237.332, "dur": 0.390 , "args": { "External id": 4548,"Record function id": 0, "Ev Idx": 4547 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122237.552, "dur": 0.070 , "args": { "External id": 4549,"Record function id": 0, "Ev Idx": 4548 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122238.083, "dur": 0.631 , "args": { "External id": 4550,"Record function id": 0, "Ev Idx": 4549 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122238.253, "dur": 0.371 , "args": { "External id": 4551,"Record function id": 0, "Ev Idx": 4550 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122238.444, "dur": 0.070 , "args": { "External id": 4552,"Record function id": 0, "Ev Idx": 4551 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122239.826, "dur": 1.293 , "args": { "External id": 4553,"Record function id": 0, "Ev Idx": 4552 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122240.047, "dur": 0.972 , "args": { "External id": 4554,"Record function id": 0, "Ev Idx": 4553 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381122250.276, "dur": 5.230 , "args": { "External id": 4555,"Record function id": 0, "Ev Idx": 4554 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122258.351, "dur": 5.079 , "args": { "External id": 4556,"Record function id": 0, "Ev Idx": 4555 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122266.646, "dur": 4.949 , "args": { "External id": 4557,"Record function id": 0, "Ev Idx": 4556 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122275.122, "dur": 10.490 , "args": { "External id": 4558,"Record function id": 0, "Ev Idx": 4557 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122275.653, "dur": 0.822 , "args": { "External id": 4559,"Record function id": 0, "Ev Idx": 4558 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122275.843, "dur": 0.541 , "args": { "External id": 4560,"Record function id": 0, "Ev Idx": 4559 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122276.074, "dur": 0.200 , "args": { "External id": 4561,"Record function id": 0, "Ev Idx": 4560 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122276.745, "dur": 0.671 , "args": { "External id": 4562,"Record function id": 0, "Ev Idx": 4561 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122276.905, "dur": 0.421 , "args": { "External id": 4563,"Record function id": 0, "Ev Idx": 4562 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122277.146, "dur": 0.070 , "args": { "External id": 4564,"Record function id": 0, "Ev Idx": 4563 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122277.667, "dur": 0.651 , "args": { "External id": 4565,"Record function id": 0, "Ev Idx": 4564 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122277.827, "dur": 0.401 , "args": { "External id": 4566,"Record function id": 0, "Ev Idx": 4565 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122278.047, "dur": 0.071 , "args": { "External id": 4567,"Record function id": 0, "Ev Idx": 4566 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122279.360, "dur": 1.262 , "args": { "External id": 4568,"Record function id": 0, "Ev Idx": 4567 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122279.550, "dur": 0.972 , "args": { "External id": 4569,"Record function id": 0, "Ev Idx": 4568 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122289.308, "dur": 10.229 , "args": { "External id": 4570,"Record function id": 0, "Ev Idx": 4569 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122289.819, "dur": 0.632 , "args": { "External id": 4571,"Record function id": 0, "Ev Idx": 4570 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122289.990, "dur": 0.370 , "args": { "External id": 4572,"Record function id": 0, "Ev Idx": 4571 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122290.180, "dur": 0.070 , "args": { "External id": 4573,"Record function id": 0, "Ev Idx": 4572 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122290.721, "dur": 0.651 , "args": { "External id": 4574,"Record function id": 0, "Ev Idx": 4573 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122290.901, "dur": 0.371 , "args": { "External id": 4575,"Record function id": 0, "Ev Idx": 4574 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122291.102, "dur": 0.070 , "args": { "External id": 4576,"Record function id": 0, "Ev Idx": 4575 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122291.633, "dur": 0.661 , "args": { "External id": 4577,"Record function id": 0, "Ev Idx": 4576 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122291.813, "dur": 0.391 , "args": { "External id": 4578,"Record function id": 0, "Ev Idx": 4577 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122292.024, "dur": 0.070 , "args": { "External id": 4579,"Record function id": 0, "Ev Idx": 4578 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122293.286, "dur": 1.252 , "args": { "External id": 4580,"Record function id": 0, "Ev Idx": 4579 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122293.456, "dur": 0.972 , "args": { "External id": 4581,"Record function id": 0, "Ev Idx": 4580 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381122304.116, "dur": 4.007 , "args": { "External id": 4582,"Record function id": 0, "Ev Idx": 4581 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381122311.339, "dur": 5.070 , "args": { "External id": 4583,"Record function id": 0, "Ev Idx": 4582 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122319.505, "dur": 10.389 , "args": { "External id": 4584,"Record function id": 0, "Ev Idx": 4583 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122319.986, "dur": 0.671 , "args": { "External id": 4585,"Record function id": 0, "Ev Idx": 4584 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122320.166, "dur": 0.401 , "args": { "External id": 4586,"Record function id": 0, "Ev Idx": 4585 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122320.376, "dur": 0.080 , "args": { "External id": 4587,"Record function id": 0, "Ev Idx": 4586 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122320.937, "dur": 0.662 , "args": { "External id": 4588,"Record function id": 0, "Ev Idx": 4587 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122321.128, "dur": 0.380 , "args": { "External id": 4589,"Record function id": 0, "Ev Idx": 4588 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122321.328, "dur": 0.070 , "args": { "External id": 4590,"Record function id": 0, "Ev Idx": 4589 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122321.859, "dur": 0.651 , "args": { "External id": 4591,"Record function id": 0, "Ev Idx": 4590 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122322.039, "dur": 0.381 , "args": { "External id": 4592,"Record function id": 0, "Ev Idx": 4591 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122322.240, "dur": 0.070 , "args": { "External id": 4593,"Record function id": 0, "Ev Idx": 4592 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122323.562, "dur": 1.222 , "args": { "External id": 4594,"Record function id": 0, "Ev Idx": 4593 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122323.733, "dur": 0.951 , "args": { "External id": 4595,"Record function id": 0, "Ev Idx": 4594 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381122334.553, "dur": 4.668 , "args": { "External id": 4596,"Record function id": 0, "Ev Idx": 4595 } }, { "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1728719, "tid": 1728719, "ts": 6950381122358.728, "dur": 23.493 , "args": { "External id": 4597,"Record function id": 0, "Ev Idx": 4596 } }, { "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/3", "pid": 1728719, "tid": 1728719, "ts": 6950381122382.572, "dur": 345.163 , "args": { "External id": 4598,"Record function id": 0, "Ev Idx": 4597 } }, { "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 1728719, "tid": 1728719, "ts": 6950381122387.201, "dur": 10.499 , "args": { "External id": 4599,"Record function id": 0, "Ev Idx": 4598 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122390.597, "dur": 0.421 , "args": { "External id": 4600,"Record function id": 0, "Ev Idx": 4599 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381122391.469, "dur": 0.030 , "args": { "External id": 4601,"Record function id": 0, "Ev Idx": 4600 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381122391.869, "dur": 0.040 , "args": { "External id": 4602,"Record function id": 0, "Ev Idx": 4601 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122393.142, "dur": 0.280 , "args": { "External id": 4603,"Record function id": 0, "Ev Idx": 4602 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381122393.773, "dur": 0.030 , "args": { "External id": 4604,"Record function id": 0, "Ev Idx": 4603 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1728719, "tid": 1728719, "ts": 6950381122394.154, "dur": 0.030 , "args": { "External id": 4605,"Record function id": 0, "Ev Idx": 4604 } }, { "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 1728719, "tid": 1728719, "ts": 6950381122400.375, "dur": 1.142 , "args": { "External id": 4606,"Record function id": 0, "Ev Idx": 4605 } }, { "ph": "X", "cat": "cpu_op", "name": "## Call CompiledFxGraph fhlbzairthosomedh2bm4xkbc6y43i43ehjulsy7hid3ogbzqpba ##", "pid": 1728719, "tid": 1728719, "ts": 6950381122403.531, "dur": 321.258 , "args": { "External id": 4607,"Record function id": 0, "Ev Idx": 4606 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122411.927, "dur": 6.031 , "args": { "External id": 4608,"Record function id": 0, "Ev Idx": 4607 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122427.426, "dur": 7.103 , "args": { "External id": 4609,"Record function id": 0, "Ev Idx": 4608 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122440.279, "dur": 5.150 , "args": { "External id": 4610,"Record function id": 0, "Ev Idx": 4609 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122450.819, "dur": 5.120 , "args": { "External id": 4611,"Record function id": 0, "Ev Idx": 4610 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122458.423, "dur": 12.704 , "args": { "External id": 4612,"Record function id": 0, "Ev Idx": 4611 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122459.064, "dur": 0.982 , "args": { "External id": 4613,"Record function id": 0, "Ev Idx": 4612 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122459.295, "dur": 0.651 , "args": { "External id": 4614,"Record function id": 0, "Ev Idx": 4613 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122459.615, "dur": 0.191 , "args": { "External id": 4615,"Record function id": 0, "Ev Idx": 4614 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122460.357, "dur": 0.661 , "args": { "External id": 4616,"Record function id": 0, "Ev Idx": 4615 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122460.537, "dur": 0.381 , "args": { "External id": 4617,"Record function id": 0, "Ev Idx": 4616 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122460.748, "dur": 0.070 , "args": { "External id": 4618,"Record function id": 0, "Ev Idx": 4617 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122461.269, "dur": 0.691 , "args": { "External id": 4619,"Record function id": 0, "Ev Idx": 4618 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122461.429, "dur": 0.441 , "args": { "External id": 4620,"Record function id": 0, "Ev Idx": 4619 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122461.649, "dur": 0.110 , "args": { "External id": 4621,"Record function id": 0, "Ev Idx": 4620 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122463.352, "dur": 1.513 , "args": { "External id": 4622,"Record function id": 0, "Ev Idx": 4621 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122463.603, "dur": 1.122 , "args": { "External id": 4623,"Record function id": 0, "Ev Idx": 4622 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381122476.427, "dur": 4.478 , "args": { "External id": 4624,"Record function id": 0, "Ev Idx": 4623 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122483.089, "dur": 0.882 , "args": { "External id": 4625,"Record function id": 0, "Ev Idx": 4624 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122485.744, "dur": 0.571 , "args": { "External id": 4626,"Record function id": 0, "Ev Idx": 4625 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381122490.012, "dur": 0.451 , "args": { "External id": 4627,"Record function id": 0, "Ev Idx": 4626 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381122490.283, "dur": 0.070 , "args": { "External id": 4628,"Record function id": 0, "Ev Idx": 4627 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381122493.408, "dur": 6.142 , "args": { "External id": 4629,"Record function id": 0, "Ev Idx": 4628 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122501.564, "dur": 10.740 , "args": { "External id": 4630,"Record function id": 0, "Ev Idx": 4629 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122502.095, "dur": 0.741 , "args": { "External id": 4631,"Record function id": 0, "Ev Idx": 4630 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122502.265, "dur": 0.481 , "args": { "External id": 4632,"Record function id": 0, "Ev Idx": 4631 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122502.535, "dur": 0.101 , "args": { "External id": 4633,"Record function id": 0, "Ev Idx": 4632 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122503.116, "dur": 0.652 , "args": { "External id": 4634,"Record function id": 0, "Ev Idx": 4633 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122503.297, "dur": 0.380 , "args": { "External id": 4635,"Record function id": 0, "Ev Idx": 4634 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122503.497, "dur": 0.070 , "args": { "External id": 4636,"Record function id": 0, "Ev Idx": 4635 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122504.028, "dur": 0.641 , "args": { "External id": 4637,"Record function id": 0, "Ev Idx": 4636 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122504.188, "dur": 0.391 , "args": { "External id": 4638,"Record function id": 0, "Ev Idx": 4637 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122504.389, "dur": 0.080 , "args": { "External id": 4639,"Record function id": 0, "Ev Idx": 4638 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122505.751, "dur": 1.343 , "args": { "External id": 4640,"Record function id": 0, "Ev Idx": 4639 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122505.962, "dur": 1.022 , "args": { "External id": 4641,"Record function id": 0, "Ev Idx": 4640 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 1728719, "tid": 1728719, "ts": 6950381122515.660, "dur": 3.827 , "args": { "External id": 4642,"Record function id": 0, "Ev Idx": 4641 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122520.549, "dur": 0.651 , "args": { "External id": 4643,"Record function id": 0, "Ev Idx": 4642 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view_as_real", "pid": 1728719, "tid": 1728719, "ts": 6950381122522.553, "dur": 0.611 , "args": { "External id": 4644,"Record function id": 0, "Ev Idx": 4643 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1728719, "tid": 1728719, "ts": 6950381122527.852, "dur": 0.361 , "args": { "External id": 4645,"Record function id": 0, "Ev Idx": 4644 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1728719, "tid": 1728719, "ts": 6950381122528.063, "dur": 0.050 , "args": { "External id": 4646,"Record function id": 0, "Ev Idx": 4645 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 1728719, "tid": 1728719, "ts": 6950381122530.728, "dur": 5.650 , "args": { "External id": 4647,"Record function id": 0, "Ev Idx": 4646 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122542.039, "dur": 10.840 , "args": { "External id": 4648,"Record function id": 0, "Ev Idx": 4647 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122542.630, "dur": 0.822 , "args": { "External id": 4649,"Record function id": 0, "Ev Idx": 4648 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122542.830, "dur": 0.531 , "args": { "External id": 4650,"Record function id": 0, "Ev Idx": 4649 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122543.081, "dur": 0.170 , "args": { "External id": 4651,"Record function id": 0, "Ev Idx": 4650 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122543.762, "dur": 0.621 , "args": { "External id": 4652,"Record function id": 0, "Ev Idx": 4651 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122543.922, "dur": 0.361 , "args": { "External id": 4653,"Record function id": 0, "Ev Idx": 4652 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122544.113, "dur": 0.070 , "args": { "External id": 4654,"Record function id": 0, "Ev Idx": 4653 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122544.644, "dur": 0.641 , "args": { "External id": 4655,"Record function id": 0, "Ev Idx": 4654 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122544.804, "dur": 0.381 , "args": { "External id": 4656,"Record function id": 0, "Ev Idx": 4655 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122544.994, "dur": 0.091 , "args": { "External id": 4657,"Record function id": 0, "Ev Idx": 4656 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122546.347, "dur": 1.342 , "args": { "External id": 4658,"Record function id": 0, "Ev Idx": 4657 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122546.577, "dur": 1.012 , "args": { "External id": 4659,"Record function id": 0, "Ev Idx": 4658 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 1728719, "tid": 1728719, "ts": 6950381122557.658, "dur": 4.288 , "args": { "External id": 4660,"Record function id": 0, "Ev Idx": 4659 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 1728719, "tid": 1728719, "ts": 6950381122567.246, "dur": 5.069 , "args": { "External id": 4661,"Record function id": 0, "Ev Idx": 4660 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 1728719, "tid": 1728719, "ts": 6950381122577.274, "dur": 4.339 , "args": { "External id": 4662,"Record function id": 0, "Ev Idx": 4661 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 1728719, "tid": 1728719, "ts": 6950381122584.037, "dur": 17.883 , "args": { "External id": 4663,"Record function id": 0, "Ev Idx": 4662 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122584.548, "dur": 0.511 , "args": { "External id": 4664,"Record function id": 0, "Ev Idx": 4663 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122584.809, "dur": 0.130 , "args": { "External id": 4665,"Record function id": 0, "Ev Idx": 4664 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122585.299, "dur": 0.391 , "args": { "External id": 4666,"Record function id": 0, "Ev Idx": 4665 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122585.490, "dur": 0.090 , "args": { "External id": 4667,"Record function id": 0, "Ev Idx": 4666 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122585.951, "dur": 0.390 , "args": { "External id": 4668,"Record function id": 0, "Ev Idx": 4667 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122586.161, "dur": 0.070 , "args": { "External id": 4669,"Record function id": 0, "Ev Idx": 4668 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 1728719, "tid": 1728719, "ts": 6950381122586.712, "dur": 13.615 , "args": { "External id": 4670,"Record function id": 0, "Ev Idx": 4669 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1728719, "tid": 1728719, "ts": 6950381122587.744, "dur": 1.944 , "args": { "External id": 4671,"Record function id": 0, "Ev Idx": 4670 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122588.105, "dur": 1.462 , "args": { "External id": 4672,"Record function id": 0, "Ev Idx": 4671 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122590.128, "dur": 1.273 , "args": { "External id": 4673,"Record function id": 0, "Ev Idx": 4672 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122591.862, "dur": 1.292 , "args": { "External id": 4674,"Record function id": 0, "Ev Idx": 4673 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122593.445, "dur": 1.442 , "args": { "External id": 4675,"Record function id": 0, "Ev Idx": 4674 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122598.985, "dur": 1.012 , "args": { "External id": 4676,"Record function id": 0, "Ev Idx": 4675 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122600.648, "dur": 0.531 , "args": { "External id": 4677,"Record function id": 0, "Ev Idx": 4676 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122600.858, "dur": 0.201 , "args": { "External id": 4678,"Record function id": 0, "Ev Idx": 4677 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122608.623, "dur": 6.853 , "args": { "External id": 4679,"Record function id": 0, "Ev Idx": 4678 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122618.070, "dur": 10.770 , "args": { "External id": 4680,"Record function id": 0, "Ev Idx": 4679 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122618.692, "dur": 0.711 , "args": { "External id": 4681,"Record function id": 0, "Ev Idx": 4680 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122618.872, "dur": 0.431 , "args": { "External id": 4682,"Record function id": 0, "Ev Idx": 4681 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122619.092, "dur": 0.090 , "args": { "External id": 4683,"Record function id": 0, "Ev Idx": 4682 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122619.683, "dur": 0.662 , "args": { "External id": 4684,"Record function id": 0, "Ev Idx": 4683 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122619.834, "dur": 0.420 , "args": { "External id": 4685,"Record function id": 0, "Ev Idx": 4684 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122620.064, "dur": 0.080 , "args": { "External id": 4686,"Record function id": 0, "Ev Idx": 4685 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122620.605, "dur": 0.631 , "args": { "External id": 4687,"Record function id": 0, "Ev Idx": 4686 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122620.765, "dur": 0.381 , "args": { "External id": 4688,"Record function id": 0, "Ev Idx": 4687 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122620.966, "dur": 0.080 , "args": { "External id": 4689,"Record function id": 0, "Ev Idx": 4688 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122622.388, "dur": 1.313 , "args": { "External id": 4690,"Record function id": 0, "Ev Idx": 4689 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122622.599, "dur": 0.992 , "args": { "External id": 4691,"Record function id": 0, "Ev Idx": 4690 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 1728719, "tid": 1728719, "ts": 6950381122632.958, "dur": 5.240 , "args": { "External id": 4692,"Record function id": 0, "Ev Idx": 4691 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122641.093, "dur": 5.330 , "args": { "External id": 4693,"Record function id": 0, "Ev Idx": 4692 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_0", "pid": 1728719, "tid": 1728719, "ts": 6950381122649.689, "dur": 5.130 , "args": { "External id": 4694,"Record function id": 0, "Ev Idx": 4693 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122658.345, "dur": 10.370 , "args": { "External id": 4695,"Record function id": 0, "Ev Idx": 4694 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122658.866, "dur": 0.672 , "args": { "External id": 4696,"Record function id": 0, "Ev Idx": 4695 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122659.027, "dur": 0.410 , "args": { "External id": 4697,"Record function id": 0, "Ev Idx": 4696 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122659.237, "dur": 0.090 , "args": { "External id": 4698,"Record function id": 0, "Ev Idx": 4697 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122659.808, "dur": 0.671 , "args": { "External id": 4699,"Record function id": 0, "Ev Idx": 4698 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122659.978, "dur": 0.411 , "args": { "External id": 4700,"Record function id": 0, "Ev Idx": 4699 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122660.209, "dur": 0.070 , "args": { "External id": 4701,"Record function id": 0, "Ev Idx": 4700 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122660.730, "dur": 0.651 , "args": { "External id": 4702,"Record function id": 0, "Ev Idx": 4701 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122660.900, "dur": 0.391 , "args": { "External id": 4703,"Record function id": 0, "Ev Idx": 4702 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122661.121, "dur": 0.070 , "args": { "External id": 4704,"Record function id": 0, "Ev Idx": 4703 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122662.413, "dur": 1.272 , "args": { "External id": 4705,"Record function id": 0, "Ev Idx": 4704 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122662.613, "dur": 0.972 , "args": { "External id": 4706,"Record function id": 0, "Ev Idx": 4705 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122672.361, "dur": 10.119 , "args": { "External id": 4707,"Record function id": 0, "Ev Idx": 4706 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122672.862, "dur": 0.652 , "args": { "External id": 4708,"Record function id": 0, "Ev Idx": 4707 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122673.033, "dur": 0.390 , "args": { "External id": 4709,"Record function id": 0, "Ev Idx": 4708 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122673.243, "dur": 0.080 , "args": { "External id": 4710,"Record function id": 0, "Ev Idx": 4709 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122673.784, "dur": 0.631 , "args": { "External id": 4711,"Record function id": 0, "Ev Idx": 4710 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122673.944, "dur": 0.371 , "args": { "External id": 4712,"Record function id": 0, "Ev Idx": 4711 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122674.145, "dur": 0.070 , "args": { "External id": 4713,"Record function id": 0, "Ev Idx": 4712 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122674.676, "dur": 0.641 , "args": { "External id": 4714,"Record function id": 0, "Ev Idx": 4713 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122674.846, "dur": 0.381 , "args": { "External id": 4715,"Record function id": 0, "Ev Idx": 4714 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122675.046, "dur": 0.071 , "args": { "External id": 4716,"Record function id": 0, "Ev Idx": 4715 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122676.309, "dur": 1.272 , "args": { "External id": 4717,"Record function id": 0, "Ev Idx": 4716 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122676.499, "dur": 0.972 , "args": { "External id": 4718,"Record function id": 0, "Ev Idx": 4717 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_mul_silu_view_7", "pid": 1728719, "tid": 1728719, "ts": 6950381122687.349, "dur": 4.218 , "args": { "External id": 4719,"Record function id": 0, "Ev Idx": 4718 } }, { "ph": "X", "cat": "cpu_op", "name": "_quantize_rows_1", "pid": 1728719, "tid": 1728719, "ts": 6950381122694.863, "dur": 5.320 , "args": { "External id": 4720,"Record function id": 0, "Ev Idx": 4719 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122703.550, "dur": 10.279 , "args": { "External id": 4721,"Record function id": 0, "Ev Idx": 4720 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122704.050, "dur": 0.692 , "args": { "External id": 4722,"Record function id": 0, "Ev Idx": 4721 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122704.231, "dur": 0.411 , "args": { "External id": 4723,"Record function id": 0, "Ev Idx": 4722 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122704.461, "dur": 0.080 , "args": { "External id": 4724,"Record function id": 0, "Ev Idx": 4723 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122704.992, "dur": 0.671 , "args": { "External id": 4725,"Record function id": 0, "Ev Idx": 4724 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122705.163, "dur": 0.410 , "args": { "External id": 4726,"Record function id": 0, "Ev Idx": 4725 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122705.403, "dur": 0.070 , "args": { "External id": 4727,"Record function id": 0, "Ev Idx": 4726 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122705.924, "dur": 0.611 , "args": { "External id": 4728,"Record function id": 0, "Ev Idx": 4727 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122706.074, "dur": 0.371 , "args": { "External id": 4729,"Record function id": 0, "Ev Idx": 4728 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122706.275, "dur": 0.070 , "args": { "External id": 4730,"Record function id": 0, "Ev Idx": 4729 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122707.547, "dur": 1.292 , "args": { "External id": 4731,"Record function id": 0, "Ev Idx": 4730 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122707.757, "dur": 0.972 , "args": { "External id": 4732,"Record function id": 0, "Ev Idx": 4731 } }, { "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 1728719, "tid": 1728719, "ts": 6950381122718.457, "dur": 4.709 , "args": { "External id": 4733,"Record function id": 0, "Ev Idx": 4732 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 1728719, "tid": 1728719, "ts": 6950381122739.516, "dur": 7.334 , "args": { "External id": 4734,"Record function id": 0, "Ev Idx": 4733 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1728719, "tid": 1728719, "ts": 6950381122749.846, "dur": 0.260 , "args": { "External id": 4735,"Record function id": 0, "Ev Idx": 4734 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381122752.841, "dur": 22.953 , "args": { "External id": 4736,"Record function id": 0, "Ev Idx": 4735 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122753.142, "dur": 0.881 , "args": { "External id": 4737,"Record function id": 0, "Ev Idx": 4736 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122753.342, "dur": 0.581 , "args": { "External id": 4738,"Record function id": 0, "Ev Idx": 4737 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122753.593, "dur": 0.180 , "args": { "External id": 4739,"Record function id": 0, "Ev Idx": 4738 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381122754.394, "dur": 21.180 , "args": { "External id": 4740,"Record function id": 0, "Ev Idx": 4739 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122754.755, "dur": 20.618 , "args": { "External id": 4741,"Record function id": 0, "Ev Idx": 4740 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1728719, "tid": 1728719, "ts": 6950381122778.920, "dur": 1.072 , "args": { "External id": 4742,"Record function id": 0, "Ev Idx": 4741 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122779.491, "dur": 0.230 , "args": { "External id": 4743,"Record function id": 0, "Ev Idx": 4742 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1728719, "tid": 1728719, "ts": 6950381122783.017, "dur": 1.103 , "args": { "External id": 4744,"Record function id": 0, "Ev Idx": 4743 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122783.669, "dur": 0.230 , "args": { "External id": 4745,"Record function id": 0, "Ev Idx": 4744 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1728719, "tid": 1728719, "ts": 6950381122785.161, "dur": 0.541 , "args": { "External id": 4746,"Record function id": 0, "Ev Idx": 4745 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122785.442, "dur": 0.140 , "args": { "External id": 4747,"Record function id": 0, "Ev Idx": 4746 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1728719, "tid": 1728719, "ts": 6950381122786.554, "dur": 0.531 , "args": { "External id": 4748,"Record function id": 0, "Ev Idx": 4747 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122786.845, "dur": 0.120 , "args": { "External id": 4749,"Record function id": 0, "Ev Idx": 4748 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381122788.798, "dur": 0.651 , "args": { "External id": 4750,"Record function id": 0, "Ev Idx": 4749 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::where", "pid": 1728719, "tid": 1728719, "ts": 6950381122790.692, "dur": 13.174 , "args": { "External id": 4751,"Record function id": 0, "Ev Idx": 4750 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122791.824, "dur": 1.072 , "args": { "External id": 4752,"Record function id": 0, "Ev Idx": 4751 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1728719, "tid": 1728719, "ts": 6950381122796.002, "dur": 1.643 , "args": { "External id": 4753,"Record function id": 0, "Ev Idx": 4752 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::layer_norm", "pid": 1728719, "tid": 1728719, "ts": 6950381122816.069, "dur": 14.798 , "args": { "External id": 4754,"Record function id": 0, "Ev Idx": 4753 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::native_layer_norm", "pid": 1728719, "tid": 1728719, "ts": 6950381122817.261, "dur": 12.964 , "args": { "External id": 4755,"Record function id": 0, "Ev Idx": 4754 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122818.934, "dur": 1.573 , "args": { "External id": 4756,"Record function id": 0, "Ev Idx": 4755 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122821.058, "dur": 1.233 , "args": { "External id": 4757,"Record function id": 0, "Ev Idx": 4756 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1728719, "tid": 1728719, "ts": 6950381122822.571, "dur": 1.393 , "args": { "External id": 4758,"Record function id": 0, "Ev Idx": 4757 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381122829.093, "dur": 0.140 , "args": { "External id": 4759,"Record function id": 0, "Ev Idx": 4758 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381122829.574, "dur": 0.080 , "args": { "External id": 4760,"Record function id": 0, "Ev Idx": 4759 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1728719, "tid": 1728719, "ts": 6950381122834.533, "dur": 7.564 , "args": { "External id": 4761,"Record function id": 0, "Ev Idx": 4760 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 1728719, "tid": 1728719, "ts": 6950381122843.410, "dur": 6.452 , "args": { "External id": 4762,"Record function id": 0, "Ev Idx": 4761 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1728719, "tid": 1728719, "ts": 6950381122855.793, "dur": 19.797 , "args": { "External id": 4763,"Record function id": 0, "Ev Idx": 4762 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1728719, "tid": 1728719, "ts": 6950381122856.023, "dur": 0.832 , "args": { "External id": 4764,"Record function id": 0, "Ev Idx": 4763 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1728719, "tid": 1728719, "ts": 6950381122856.244, "dur": 0.511 , "args": { "External id": 4765,"Record function id": 0, "Ev Idx": 4764 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1728719, "tid": 1728719, "ts": 6950381122856.514, "dur": 0.131 , "args": { "External id": 4766,"Record function id": 0, "Ev Idx": 4765 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1728719, "tid": 1728719, "ts": 6950381122857.105, "dur": 18.334 , "args": { "External id": 4767,"Record function id": 0, "Ev Idx": 4766 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1728719, "tid": 1728719, "ts": 6950381122858.728, "dur": 0.712 , "args": { "External id": 4768,"Record function id": 0, "Ev Idx": 4767 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1728719, "tid": 1728719, "ts": 6950381122859.059, "dur": 0.250 , "args": { "External id": 4769,"Record function id": 0, "Ev Idx": 4768 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1728719, "tid": 1728719, "ts": 6950381122859.790, "dur": 14.277 , "args": { "External id": 4770,"Record function id": 0, "Ev Idx": 4769 } }, { "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1728719, "tid": 1728719, "ts": 6950381122874.638, "dur": 0.461 , "args": { "External id": 4771,"Record function id": 0, "Ev Idx": 4770 } }, { "ph": "X", "cat": "overhead", "name": "Activity Buffer Request", "pid": -1, "tid": 0, "ts": 6950381108284.449, "dur": 77.575 }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108365.170, "dur": 0.711 , "args": { "External id": 8,"cbid": 200, "correlation": 12 } }, { "ph": "f", "id": 12, "pid": 1728719, "tid": 1728719, "ts": 6950381108365.170, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tt_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381108389.468, "dur": 16.000 , "args": { "External id": 8,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15, "registers per thread": 218, "shared memory": 73728, "blocks per SM": 2.7234041690826416, "warps per SM": 21.787233352661133, "grid": [128, 4, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 57344, "allocatedSharedMemPerBlock": 74752}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15, "pid": 0, "tid": 7, "ts": 6950381108389.468, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108368.015, "dur": 14.677 , "args": { "External id": 8,"cbid": 307, "correlation": 15 } }, { "ph": "s", "id": 15, "pid": 1728719, "tid": 1728719, "ts": 6950381108368.015, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, 4, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, "ts": 6950381108423.132, "dur": 3.008 , "args": { "External id": 13,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 28, "registers per thread": 31, "shared memory": 0, "blocks per SM": 1.4042552709579468, "warps per SM": 5.617021083831787, "grid": [264, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 12, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 28, "pid": 0, "tid": 7, "ts": 6950381108423.132, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108410.754, "dur": 4.719 , "args": { "External id": 13,"cbid": 211, "correlation": 28 } }, { "ph": "s", "id": 28, "pid": 1728719, "tid": 1728719, "ts": 6950381108410.754, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array, false>(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381108434.683, "dur": 0.865 , "args": { "External id": 14,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 40, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.7021276354789734, "warps per SM": 2.8085105419158936, "grid": [132, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 6, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 40, "pid": 0, "tid": 7, "ts": 6950381108434.683, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108429.790, "dur": 2.805 , "args": { "External id": 14,"cbid": 211, "correlation": 40 } }, { "ph": "s", "id": 40, "pid": 1728719, "tid": 1728719, "ts": 6950381108429.790, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, float, 4, 4> >(at::native::ReduceOp, unsigned int, float, 4, 4>)", "pid": 0, "tid": 7, "ts": 6950381108447.771, "dur": 5.984 , "args": { "External id": 17,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 55, "registers per thread": 27, "shared memory": 16, "blocks per SM": 0.01595744676887989, "warps per SM": 0.25531914830207825, "grid": [3, 1, 1], "block": [32, 16, 1], "est. achieved occupancy %": 1, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 55, "pid": 0, "tid": 7, "ts": 6950381108447.771, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108443.245, "dur": 2.565 , "args": { "External id": 17,"cbid": 211, "correlation": 55 } }, { "ph": "s", "id": 55, "pid": 1728719, "tid": 1728719, "ts": 6950381108443.245, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, std::array, false>(int, at::native::CUDAFunctorOnSelf_add, std::array)", "pid": 0, "tid": 7, "ts": 6950381108462.331, "dur": 1.088 , "args": { "External id": 18,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 66, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 66, "pid": 0, "tid": 7, "ts": 6950381108462.331, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108457.872, "dur": 2.244 , "args": { "External id": 18,"cbid": 211, "correlation": 66 } }, { "ph": "s", "id": 66, "pid": 1728719, "tid": 1728719, "ts": 6950381108457.872, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array, false>(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381108470.939, "dur": 1.088 , "args": { "External id": 19,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 77, "registers per thread": 32, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 77, "pid": 0, "tid": 7, "ts": 6950381108470.939, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108466.658, "dur": 2.305 , "args": { "External id": 19,"cbid": 211, "correlation": 77 } }, { "ph": "s", "id": 77, "pid": 1728719, "tid": 1728719, "ts": 6950381108466.658, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381108479.931, "dur": 1.344 , "args": { "External id": 20,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 88, "registers per thread": 21, "shared memory": 0, "blocks per SM": 2.8085105419158936, "warps per SM": 11.234042167663574, "grid": [528, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 23, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 88, "pid": 0, "tid": 7, "ts": 6950381108479.931, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108475.495, "dur": 2.695 , "args": { "External id": 20,"cbid": 211, "correlation": 88 } }, { "ph": "s", "id": 88, "pid": 1728719, "tid": 1728719, "ts": 6950381108475.495, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, 4, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, "ts": 6950381108488.219, "dur": 2.656 , "args": { "External id": 24,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 102, "registers per thread": 31, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 102, "pid": 0, "tid": 7, "ts": 6950381108488.219, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108484.121, "dur": 1.894 , "args": { "External id": 24,"cbid": 211, "correlation": 102 } }, { "ph": "s", "id": 102, "pid": 1728719, "tid": 1728719, "ts": 6950381108484.121, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, std::array, false>(int, at::native::CUDAFunctorOnSelf_add, std::array)", "pid": 0, "tid": 7, "ts": 6950381108495.611, "dur": 0.800 , "args": { "External id": 25,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 114, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.021276595070958138, "warps per SM": 0.08510638028383255, "grid": [4, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 114, "pid": 0, "tid": 7, "ts": 6950381108495.611, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108491.244, "dur": 1.854 , "args": { "External id": 25,"cbid": 211, "correlation": 114 } }, { "ph": "s", "id": 114, "pid": 1728719, "tid": 1728719, "ts": 6950381108491.244, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381108500.731, "dur": 1.184 , "args": { "External id": 26,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 125, "registers per thread": 21, "shared memory": 0, "blocks per SM": 2.8085105419158936, "warps per SM": 11.234042167663574, "grid": [528, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 23, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 125, "pid": 0, "tid": 7, "ts": 6950381108500.731, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108496.765, "dur": 1.743 , "args": { "External id": 26,"cbid": 211, "correlation": 125 } }, { "ph": "s", "id": 125, "pid": 1728719, "tid": 1728719, "ts": 6950381108496.765, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::bfloat16_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda(float)#1}, std::array, false>(int, at::native::bfloat16_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda(float)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381108510.939, "dur": 0.864 , "args": { "External id": 30,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 139, "registers per thread": 40, "shared memory": 0, "blocks per SM": 0.7021276354789734, "warps per SM": 2.8085105419158936, "grid": [132, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 6, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 139, "pid": 0, "tid": 7, "ts": 6950381108510.939, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108506.943, "dur": 1.944 , "args": { "External id": 30,"cbid": 211, "correlation": 139 } }, { "ph": "s", "id": 139, "pid": 1728719, "tid": 1728719, "ts": 6950381108506.943, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108526.109, "dur": 0.201 , "args": { "External id": 38,"cbid": 200, "correlation": 155 } }, { "ph": "f", "id": 155, "pid": 1728719, "tid": 1728719, "ts": 6950381108526.109, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_32x6_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381108532.570, "dur": 23.488 , "args": { "External id": 38,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 158, "registers per thread": 140, "shared memory": 73728, "blocks per SM": 0.6808510422706604, "warps per SM": 2.7234041690826416, "grid": [8, 4, 4], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 18432, "allocatedSharedMemPerBlock": 74752}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 158, "pid": 0, "tid": 7, "ts": 6950381108532.570, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108528.153, "dur": 2.414 , "args": { "External id": 38,"cbid": 307, "correlation": 158 } }, { "ph": "s", "id": 158, "pid": 1728719, "tid": 1728719, "ts": 6950381108528.153, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7, "ts": 6950381108556.346, "dur": 1.760 , "args": { "External id": 38,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 160, "registers per thread": 42, "shared memory": 0, "blocks per SM": 2.042553186416626, "warps per SM": 32.680850982666016, "grid": [128, 3, 1], "block": [32, 16, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 2, "limitingFactors": "REGS", "blockLimitRegs": 2, "blockLimitSharedMem": 100, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 24576, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 160, "pid": 0, "tid": 7, "ts": 6950381108556.346, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernelExC", "pid": 1728719, "tid": 1728719, "ts": 6950381108531.409, "dur": 3.396 , "args": { "External id": 38,"cbid": 430, "correlation": 160 } }, { "ph": "s", "id": 160, "pid": 1728719, "tid": 1728719, "ts": 6950381108531.409, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::GeluCUDAKernelImpl(at::TensorIteratorBase&, at::native::GeluType)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(c10::BFloat16)#1}, std::array, false>(int, at::native::GeluCUDAKernelImpl(at::TensorIteratorBase&, at::native::GeluType)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(c10::BFloat16)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381108558.394, "dur": 0.928 , "args": { "External id": 40,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 171, "registers per thread": 38, "shared memory": 0, "blocks per SM": 0.7021276354789734, "warps per SM": 2.8085105419158936, "grid": [132, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 6, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 171, "pid": 0, "tid": 7, "ts": 6950381108558.394, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108553.560, "dur": 2.866 , "args": { "External id": 40,"cbid": 211, "correlation": 171 } }, { "ph": "s", "id": 171, "pid": 1728719, "tid": 1728719, "ts": 6950381108553.560, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108566.584, "dur": 0.161 , "args": { "External id": 48,"cbid": 200, "correlation": 186 } }, { "ph": "f", "id": 186, "pid": 1728719, "tid": 1728719, "ts": 6950381108566.584, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_32x6_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381108571.162, "dur": 23.584 , "args": { "External id": 48,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 189, "registers per thread": 140, "shared memory": 73728, "blocks per SM": 0.6808510422706604, "warps per SM": 2.7234041690826416, "grid": [8, 4, 4], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 18432, "allocatedSharedMemPerBlock": 74752}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 189, "pid": 0, "tid": 7, "ts": 6950381108571.162, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108567.166, "dur": 1.803 , "args": { "External id": 48,"cbid": 307, "correlation": 189 } }, { "ph": "s", "id": 189, "pid": 1728719, "tid": 1728719, "ts": 6950381108567.166, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7, "ts": 6950381108595.002, "dur": 1.632 , "args": { "External id": 48,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 191, "registers per thread": 42, "shared memory": 0, "blocks per SM": 2.042553186416626, "warps per SM": 32.680850982666016, "grid": [128, 3, 1], "block": [32, 16, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 2, "limitingFactors": "REGS", "blockLimitRegs": 2, "blockLimitSharedMem": 100, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 24576, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 191, "pid": 0, "tid": 7, "ts": 6950381108595.002, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernelExC", "pid": 1728719, "tid": 1728719, "ts": 6950381108569.219, "dur": 1.934 , "args": { "External id": 48,"cbid": 430, "correlation": 191 } }, { "ph": "s", "id": 191, "pid": 1728719, "tid": 1728719, "ts": 6950381108569.219, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<2, at::native::FillFunctor, std::array, false>(int, at::native::FillFunctor, std::array)", "pid": 0, "tid": 7, "ts": 6950381108596.922, "dur": 0.768 , "args": { "External id": 53,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 205, "registers per thread": 22, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 205, "pid": 0, "tid": 7, "ts": 6950381108596.922, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108583.256, "dur": 2.564 , "args": { "External id": 53,"cbid": 211, "correlation": 205 } }, { "ph": "s", "id": 205, "pid": 1728719, "tid": 1728719, "ts": 6950381108583.256, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<2, at::native::FillFunctor, std::array, false>(int, at::native::FillFunctor, std::array)", "pid": 0, "tid": 7, "ts": 6950381108597.946, "dur": 0.768 , "args": { "External id": 56,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 219, "registers per thread": 22, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 219, "pid": 0, "tid": 7, "ts": 6950381108597.946, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108589.718, "dur": 1.863 , "args": { "External id": 56,"cbid": 211, "correlation": 219 } }, { "ph": "s", "id": 219, "pid": 1728719, "tid": 1728719, "ts": 6950381108589.718, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::(anonymous namespace)::where_kernel_impl(at::TensorIterator&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(bool, unsigned long, unsigned long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::(anonymous namespace)::where_kernel_impl(at::TensorIterator&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(bool, unsigned long, unsigned long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381108611.610, "dur": 1.664 , "args": { "External id": 57,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 237, "registers per thread": 23, "shared memory": 0, "blocks per SM": 0.026595745235681534, "warps per SM": 0.10638298094272614, "grid": [5, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 237, "pid": 0, "tid": 7, "ts": 6950381108611.610, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108607.040, "dur": 2.955 , "args": { "External id": 57,"cbid": 211, "correlation": 237 } }, { "ph": "s", "id": 237, "pid": 1728719, "tid": 1728719, "ts": 6950381108607.040, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceScanInitKernel >(at_cuda_detail::cub::ScanTileState, int)", "pid": 0, "tid": 7, "ts": 6950381108646.489, "dur": 0.672 , "args": { "External id": 67,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 275, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 275, "pid": 0, "tid": 7, "ts": 6950381108646.489, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108636.304, "dur": 3.086 , "args": { "External id": 67,"cbid": 211, "correlation": 275 } }, { "ph": "s", "id": 275, "pid": 1728719, "tid": 1728719, "ts": 6950381108636.304, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108639.741, "dur": 0.090 , "args": { "External id": 67,"cbid": 11, "correlation": 276 } }, { "ph": "f", "id": 276, "pid": 1728719, "tid": 1728719, "ts": 6950381108639.741, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108639.911, "dur": 0.170 , "args": { "External id": 67,"cbid": 11, "correlation": 277 } }, { "ph": "f", "id": 277, "pid": 1728719, "tid": 1728719, "ts": 6950381108639.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381108640.913, "dur": 1.122 , "args": { "External id": 67,"cbid": 251, "correlation": 280 } }, { "ph": "f", "id": 280, "pid": 1728719, "tid": 1728719, "ts": 6950381108640.913, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108642.546, "dur": 0.260 , "args": { "External id": 67,"cbid": 200, "correlation": 284 } }, { "ph": "f", "id": 284, "pid": 1728719, "tid": 1728719, "ts": 6950381108642.546, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceScanKernel >::Policy1000, long const*, long*, at_cuda_detail::cub::ScanTileState, std::plus, at_cuda_detail::cub::NullType, unsigned int, long, false, at_cuda_detail::cub::NullType>(long const*, long*, at_cuda_detail::cub::ScanTileState, int, std::plus, at_cuda_detail::cub::NullType, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381108648.121, "dur": 1.696 , "args": { "External id": 67,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 287, "registers per thread": 40, "shared memory": 7184, "blocks per SM": 0.010638297535479069, "warps per SM": 0.042553190141916275, "grid": [2, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS|SMEM", "blockLimitRegs": 12, "blockLimitSharedMem": 12, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 8320}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 287, "pid": 0, "tid": 7, "ts": 6950381108648.121, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108643.377, "dur": 2.375 , "args": { "External id": 67,"cbid": 211, "correlation": 287 } }, { "ph": "s", "id": 287, "pid": 1728719, "tid": 1728719, "ts": 6950381108643.377, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108645.892, "dur": 0.030 , "args": { "External id": 67,"cbid": 11, "correlation": 288 } }, { "ph": "f", "id": 288, "pid": 1728719, "tid": 1728719, "ts": 6950381108645.892, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108645.992, "dur": 0.020 , "args": { "External id": 67,"cbid": 11, "correlation": 289 } }, { "ph": "f", "id": 289, "pid": 1728719, "tid": 1728719, "ts": 6950381108645.992, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381108670.521, "dur": 0.672 , "args": { "External id": 72,"device": 0, "context": 1, "stream": 7, "correlation": 298, "bytes": 8, "memory bandwidth (GB/s)": 0.011904761904761904, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 298, "pid": 0, "tid": 7, "ts": 6950381108670.521, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381108652.845, "dur": 9.418 , "args": { "External id": 72,"cbid": 41, "correlation": 298 } }, { "ph": "s", "id": 298, "pid": 1728719, "tid": 1728719, "ts": 6950381108652.845, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381108662.503, "dur": 10.329 , "args": { "External id": 72,"cbid": 131, "correlation": 299 } }, { "ph": "s", "id": 299, "pid": 1728719, "tid": 1728719, "ts": 6950381108662.503, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::compare_scalar_kernel(at::TensorIteratorBase&, at::native::(anonymous namespace)::OpType, long)::{lambda(long)#1}, std::array, false>(int, at::native::compare_scalar_kernel(at::TensorIteratorBase&, at::native::(anonymous namespace)::OpType, long)::{lambda(long)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381108684.409, "dur": 0.928 , "args": { "External id": 73,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 309, "registers per thread": 40, "shared memory": 0, "blocks per SM": 0.010638297535479069, "warps per SM": 0.042553190141916275, "grid": [2, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 309, "pid": 0, "tid": 7, "ts": 6950381108684.409, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108679.865, "dur": 2.705 , "args": { "External id": 73,"cbid": 211, "correlation": 309 } }, { "ph": "s", "id": 309, "pid": 1728719, "tid": 1728719, "ts": 6950381108679.865, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, bool, 4, 4> >(at::native::ReduceOp, unsigned int, bool, 4, 4>)", "pid": 0, "tid": 7, "ts": 6950381108693.561, "dur": 1.888 , "args": { "External id": 74,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 324, "registers per thread": 26, "shared memory": 272, "blocks per SM": 0.005319148767739534, "warps per SM": 0.042553190141916275, "grid": [1, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 72, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1408}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 324, "pid": 0, "tid": 7, "ts": 6950381108693.561, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108689.073, "dur": 2.735 , "args": { "External id": 74,"cbid": 211, "correlation": 324 } }, { "ph": "s", "id": 324, "pid": 1728719, "tid": 1728719, "ts": 6950381108689.073, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381108699.161, "dur": 0.544 , "args": { "External id": 77,"device": 0, "context": 1, "stream": 7, "correlation": 330, "bytes": 1, "memory bandwidth (GB/s)": 0.001838235294117647, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 330, "pid": 0, "tid": 7, "ts": 6950381108699.161, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381108694.342, "dur": 2.846 , "args": { "External id": 77,"cbid": 41, "correlation": 330 } }, { "ph": "s", "id": 330, "pid": 1728719, "tid": 1728719, "ts": 6950381108694.342, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381108697.348, "dur": 3.356 , "args": { "External id": 77,"cbid": 131, "correlation": 331 } }, { "ph": "s", "id": 331, "pid": 1728719, "tid": 1728719, "ts": 6950381108697.348, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void compute_cuda_kernel(long const*, long const*, long*, long, long)", "pid": 0, "tid": 7, "ts": 6950381108710.809, "dur": 1.184 , "args": { "External id": 66,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 344, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.35638296604156494, "warps per SM": 5.702127456665039, "grid": [67, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 12, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 100, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 344, "pid": 0, "tid": 7, "ts": 6950381108710.809, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108705.643, "dur": 3.116 , "args": { "External id": 66,"cbid": 211, "correlation": 344 } }, { "ph": "s", "id": 344, "pid": 1728719, "tid": 1728719, "ts": 6950381108705.643, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::_scatter_gather_elementwise_kernel<128, 8, at::native::_cuda_scatter_gather_internal_kernel, long>::operator()(at::TensorIterator&, long, long, long, at::native::TensorAssign const&)::{lambda(int)#1}>(int, at::native::_cuda_scatter_gather_internal_kernel, long>::operator()(at::TensorIterator&, long, long, long, at::native::TensorAssign const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381108727.896, "dur": 3.712 , "args": { "External id": 85,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 366, "registers per thread": 28, "shared memory": 0, "blocks per SM": 0.026595745235681534, "warps per SM": 0.10638298094272614, "grid": [5, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 366, "pid": 0, "tid": 7, "ts": 6950381108727.896, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108723.356, "dur": 2.605 , "args": { "External id": 85,"cbid": 211, "correlation": 366 } }, { "ph": "s", "id": 366, "pid": 1728719, "tid": 1728719, "ts": 6950381108723.356, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, std::array, false>(int, at::native::FillFunctor, std::array)", "pid": 0, "tid": 7, "ts": 6950381108744.248, "dur": 2.432 , "args": { "External id": 92,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 383, "registers per thread": 16, "shared memory": 0, "blocks per SM": 21.787233352661133, "warps per SM": 87.14893341064453, "grid": [4096, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 383, "pid": 0, "tid": 7, "ts": 6950381108744.248, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108739.807, "dur": 2.795 , "args": { "External id": 92,"cbid": 211, "correlation": 383 } }, { "ph": "s", "id": 383, "pid": 1728719, "tid": 1728719, "ts": 6950381108739.807, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_vectorized, unsigned int, 2, 128, 1, 16, 8>(char*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381108756.120, "dur": 4.384 , "args": { "External id": 93,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 397, "registers per thread": 19, "shared memory": 0, "blocks per SM": 43.574466705322266, "warps per SM": 174.29786682128906, "grid": [4096, 2, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 397, "pid": 0, "tid": 7, "ts": 6950381108756.120, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108751.559, "dur": 2.735 , "args": { "External id": 93,"cbid": 211, "correlation": 397 } }, { "ph": "s", "id": 397, "pid": 1728719, "tid": 1728719, "ts": 6950381108751.559, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceScanInitKernel >(at_cuda_detail::cub::ScanTileState, int)", "pid": 0, "tid": 7, "ts": 6950381108767.064, "dur": 0.704 , "args": { "External id": 96,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 429, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 429, "pid": 0, "tid": 7, "ts": 6950381108767.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108763.140, "dur": 1.734 , "args": { "External id": 96,"cbid": 211, "correlation": 429 } }, { "ph": "s", "id": 429, "pid": 1728719, "tid": 1728719, "ts": 6950381108763.140, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108764.994, "dur": 0.050 , "args": { "External id": 96,"cbid": 11, "correlation": 430 } }, { "ph": "f", "id": 430, "pid": 1728719, "tid": 1728719, "ts": 6950381108764.994, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108765.104, "dur": 0.020 , "args": { "External id": 96,"cbid": 11, "correlation": 431 } }, { "ph": "f", "id": 431, "pid": 1728719, "tid": 1728719, "ts": 6950381108765.104, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381108765.385, "dur": 0.320 , "args": { "External id": 96,"cbid": 251, "correlation": 434 } }, { "ph": "f", "id": 434, "pid": 1728719, "tid": 1728719, "ts": 6950381108765.385, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108766.116, "dur": 0.190 , "args": { "External id": 96,"cbid": 200, "correlation": 438 } }, { "ph": "f", "id": 438, "pid": 1728719, "tid": 1728719, "ts": 6950381108766.116, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceScanKernel >::Policy1000, long const*, long*, at_cuda_detail::cub::ScanTileState, std::plus, at_cuda_detail::cub::NullType, unsigned int, long, false, at_cuda_detail::cub::NullType>(long const*, long*, at_cuda_detail::cub::ScanTileState, int, std::plus, at_cuda_detail::cub::NullType, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381108770.584, "dur": 1.472 , "args": { "External id": 96,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 441, "registers per thread": 40, "shared memory": 7184, "blocks per SM": 0.010638297535479069, "warps per SM": 0.042553190141916275, "grid": [2, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS|SMEM", "blockLimitRegs": 12, "blockLimitSharedMem": 12, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 8320}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 441, "pid": 0, "tid": 7, "ts": 6950381108770.584, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108766.597, "dur": 2.084 , "args": { "External id": 96,"cbid": 211, "correlation": 441 } }, { "ph": "s", "id": 441, "pid": 1728719, "tid": 1728719, "ts": 6950381108766.597, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108768.831, "dur": 0.030 , "args": { "External id": 96,"cbid": 11, "correlation": 442 } }, { "ph": "f", "id": 442, "pid": 1728719, "tid": 1728719, "ts": 6950381108768.831, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108768.911, "dur": 0.050 , "args": { "External id": 96,"cbid": 11, "correlation": 443 } }, { "ph": "f", "id": 443, "pid": 1728719, "tid": 1728719, "ts": 6950381108768.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381108776.696, "dur": 0.352 , "args": { "External id": 101,"device": 0, "context": 1, "stream": 7, "correlation": 452, "bytes": 8, "memory bandwidth (GB/s)": 0.022727272727272728, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 452, "pid": 0, "tid": 7, "ts": 6950381108776.696, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381108772.548, "dur": 2.304 , "args": { "External id": 101,"cbid": 41, "correlation": 452 } }, { "ph": "s", "id": 452, "pid": 1728719, "tid": 1728719, "ts": 6950381108772.548, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381108774.962, "dur": 3.066 , "args": { "External id": 101,"cbid": 131, "correlation": 453 } }, { "ph": "s", "id": 453, "pid": 1728719, "tid": 1728719, "ts": 6950381108774.962, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::compare_scalar_kernel(at::TensorIteratorBase&, at::native::(anonymous namespace)::OpType, long)::{lambda(long)#1}, std::array, false>(int, at::native::compare_scalar_kernel(at::TensorIteratorBase&, at::native::(anonymous namespace)::OpType, long)::{lambda(long)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381108785.976, "dur": 0.832 , "args": { "External id": 102,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 463, "registers per thread": 40, "shared memory": 0, "blocks per SM": 0.010638297535479069, "warps per SM": 0.042553190141916275, "grid": [2, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 463, "pid": 0, "tid": 7, "ts": 6950381108785.976, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108781.885, "dur": 1.834 , "args": { "External id": 102,"cbid": 211, "correlation": 463 } }, { "ph": "s", "id": 463, "pid": 1728719, "tid": 1728719, "ts": 6950381108781.885, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, bool, 4, 4> >(at::native::ReduceOp, unsigned int, bool, 4, 4>)", "pid": 0, "tid": 7, "ts": 6950381108791.640, "dur": 1.856 , "args": { "External id": 103,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 478, "registers per thread": 26, "shared memory": 272, "blocks per SM": 0.005319148767739534, "warps per SM": 0.042553190141916275, "grid": [1, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 72, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1408}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 478, "pid": 0, "tid": 7, "ts": 6950381108791.640, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108787.436, "dur": 1.933 , "args": { "External id": 103,"cbid": 211, "correlation": 478 } }, { "ph": "s", "id": 478, "pid": 1728719, "tid": 1728719, "ts": 6950381108787.436, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381108794.712, "dur": 0.544 , "args": { "External id": 106,"device": 0, "context": 1, "stream": 7, "correlation": 484, "bytes": 1, "memory bandwidth (GB/s)": 0.001838235294117647, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 484, "pid": 0, "tid": 7, "ts": 6950381108794.712, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381108790.962, "dur": 2.264 , "args": { "External id": 106,"cbid": 41, "correlation": 484 } }, { "ph": "s", "id": 484, "pid": 1728719, "tid": 1728719, "ts": 6950381108790.962, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381108793.317, "dur": 2.875 , "args": { "External id": 106,"cbid": 131, "correlation": 485 } }, { "ph": "s", "id": 485, "pid": 1728719, "tid": 1728719, "ts": 6950381108793.317, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void compute_cuda_kernel(long const*, long const*, long*, long, long)", "pid": 0, "tid": 7, "ts": 6950381108803.352, "dur": 1.151 , "args": { "External id": 95,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 498, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.35638296604156494, "warps per SM": 5.702127456665039, "grid": [67, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 12, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 100, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 498, "pid": 0, "tid": 7, "ts": 6950381108803.352, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108799.308, "dur": 1.883 , "args": { "External id": 95,"cbid": 211, "correlation": 498 } }, { "ph": "s", "id": 498, "pid": 1728719, "tid": 1728719, "ts": 6950381108799.308, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_gather_kernel<16, long>(char*, char*, long*, int, long, long, long, long, bool)", "pid": 0, "tid": 7, "ts": 6950381108817.239, "dur": 9.216 , "args": { "External id": 114,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 522, "registers per thread": 24, "shared memory": 0, "blocks per SM": 43.925533294677734, "warps per SM": 351.4042663574219, "grid": [4129, 2, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 522, "pid": 0, "tid": 7, "ts": 6950381108817.239, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108812.512, "dur": 2.555 , "args": { "External id": 114,"cbid": 211, "correlation": 522 } }, { "ph": "s", "id": 522, "pid": 1728719, "tid": 1728719, "ts": 6950381108812.512, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108829.334, "dur": 0.190 , "args": { "External id": 120,"cbid": 200, "correlation": 543 } }, { "ph": "f", "id": 543, "pid": 1728719, "tid": 1728719, "ts": 6950381108829.334, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381108829.754, "dur": 0.762 , "args": { "External id": 120,"cbid": 251, "correlation": 545 } }, { "ph": "f", "id": 545, "pid": 1728719, "tid": 1728719, "ts": 6950381108829.754, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381108831.948, "dur": 0.111 , "args": { "External id": 120,"cbid": 200, "correlation": 556 } }, { "ph": "f", "id": 556, "pid": 1728719, "tid": 1728719, "ts": 6950381108831.948, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381108832.189, "dur": 0.170 , "args": { "External id": 120,"cbid": 251, "correlation": 558 } }, { "ph": "f", "id": 558, "pid": 1728719, "tid": 1728719, "ts": 6950381108832.189, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceKernel >::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, unsigned long long, cuda::std::__4::plus, int, cuda::std::__4::__identity>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, int*, unsigned long long, at_cuda_detail::cub::GridEvenShare, cuda::std::__4::plus, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381108837.175, "dur": 1.248 , "args": { "External id": 120,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 562, "registers per thread": 34, "shared memory": 44, "blocks per SM": 0.010638297535479069, "warps per SM": 0.08510638028383255, "grid": [2, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 562, "pid": 0, "tid": 7, "ts": 6950381108837.175, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108832.900, "dur": 2.505 , "args": { "External id": 120,"cbid": 211, "correlation": 562 } }, { "ph": "s", "id": 562, "pid": 1728719, "tid": 1728719, "ts": 6950381108832.900, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108835.515, "dur": 0.030 , "args": { "External id": 120,"cbid": 11, "correlation": 563 } }, { "ph": "f", "id": 563, "pid": 1728719, "tid": 1728719, "ts": 6950381108835.515, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108835.615, "dur": 0.020 , "args": { "External id": 120,"cbid": 11, "correlation": 564 } }, { "ph": "f", "id": 564, "pid": 1728719, "tid": 1728719, "ts": 6950381108835.615, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceSingleTileKernel >::Policy1000, int*, int*, int, cuda::std::__4::plus, int, int, cuda::std::__4::__identity>(int*, int*, int, cuda::std::__4::plus, int, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381108840.247, "dur": 0.992 , "args": { "External id": 120,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 568, "registers per thread": 30, "shared memory": 44, "blocks per SM": 0.005319148767739534, "warps per SM": 0.042553190141916275, "grid": [1, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 568, "pid": 0, "tid": 7, "ts": 6950381108840.247, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108836.287, "dur": 1.973 , "args": { "External id": 120,"cbid": 211, "correlation": 568 } }, { "ph": "s", "id": 568, "pid": 1728719, "tid": 1728719, "ts": 6950381108836.287, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108838.380, "dur": 0.030 , "args": { "External id": 120,"cbid": 11, "correlation": 569 } }, { "ph": "f", "id": 569, "pid": 1728719, "tid": 1728719, "ts": 6950381108838.380, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108838.481, "dur": 0.030 , "args": { "External id": 120,"cbid": 11, "correlation": 570 } }, { "ph": "f", "id": 570, "pid": 1728719, "tid": 1728719, "ts": 6950381108838.481, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381108843.031, "dur": 0.352 , "args": { "External id": 120,"device": 0, "context": 1, "stream": 7, "correlation": 574, "bytes": 4, "memory bandwidth (GB/s)": 0.011363636363636364, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 574, "pid": 0, "tid": 7, "ts": 6950381108843.031, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381108839.523, "dur": 1.973 , "args": { "External id": 120,"cbid": 41, "correlation": 574 } }, { "ph": "s", "id": 574, "pid": 1728719, "tid": 1728719, "ts": 6950381108839.523, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381108841.677, "dur": 2.634 , "args": { "External id": 120,"cbid": 131, "correlation": 575 } }, { "ph": "s", "id": 575, "pid": 1728719, "tid": 1728719, "ts": 6950381108841.677, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceCompactInitKernel, int*>(at_cuda_detail::cub::ScanTileState, int, int*)", "pid": 0, "tid": 7, "ts": 6950381108853.239, "dur": 0.704 , "args": { "External id": 120,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 600, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 600, "pid": 0, "tid": 7, "ts": 6950381108853.239, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108848.740, "dur": 2.484 , "args": { "External id": 120,"cbid": 211, "correlation": 600 } }, { "ph": "s", "id": 600, "pid": 1728719, "tid": 1728719, "ts": 6950381108848.740, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108851.335, "dur": 0.030 , "args": { "External id": 120,"cbid": 11, "correlation": 601 } }, { "ph": "f", "id": 601, "pid": 1728719, "tid": 1728719, "ts": 6950381108851.335, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108851.425, "dur": 0.020 , "args": { "External id": 120,"cbid": 11, "correlation": 602 } }, { "ph": "f", "id": 602, "pid": 1728719, "tid": 1728719, "ts": 6950381108851.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::select::DeviceSelectSweepKernel::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, at_cuda_detail::cub::detail::select::streaming_context_t, (at_cuda_detail::cub::SelectImpl)0>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, int, at_cuda_detail::cub::detail::select::streaming_context_t, at_cuda_detail::cub::detail::vsmem_t)", "pid": 0, "tid": 7, "ts": 6950381108858.455, "dur": 3.360 , "args": { "External id": 120,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 606, "registers per thread": 64, "shared memory": 28160, "blocks per SM": 0.010638297535479069, "warps per SM": 0.10638297349214554, "grid": [2, 1, 1], "block": [320, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "REGS|SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 3, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 29184}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 606, "pid": 0, "tid": 7, "ts": 6950381108858.455, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108852.096, "dur": 2.615 , "args": { "External id": 120,"cbid": 211, "correlation": 606 } }, { "ph": "s", "id": 606, "pid": 1728719, "tid": 1728719, "ts": 6950381108852.096, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108854.821, "dur": 0.030 , "args": { "External id": 120,"cbid": 11, "correlation": 607 } }, { "ph": "f", "id": 607, "pid": 1728719, "tid": 1728719, "ts": 6950381108854.821, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381108854.901, "dur": 0.030 , "args": { "External id": 120,"cbid": 11, "correlation": 608 } }, { "ph": "f", "id": 608, "pid": 1728719, "tid": 1728719, "ts": 6950381108854.901, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381108872.599, "dur": 44.319 , "args": { "External id": 119,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 616, "registers per thread": 48, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 697.1914672851562, "grid": [32768, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 83, "occupancy": {"activeBlocksPerMultiprocessor": 10, "limitingFactors": "REGS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 616, "pid": 0, "tid": 7, "ts": 6950381108872.599, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381108868.116, "dur": 2.534 , "args": { "External id": 119,"cbid": 211, "correlation": 616 } }, { "ph": "s", "id": 616, "pid": 1728719, "tid": 1728719, "ts": 6950381108868.116, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 1728719, "tid": 1728719, "ts": 6950381108896.218, "dur": 0.992 , "args": { "External id": 140,"cbid": 317, "correlation": 626 } }, { "ph": "f", "id": 626, "pid": 1728719, "tid": 1728719, "ts": 6950381108896.218, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, "ts": 6950381108918.806, "dur": 0.704 , "args": { "External id": 140,"device": 0, "context": 1, "stream": 7, "correlation": 628, "bytes": 4129, "memory bandwidth (GB/s)": 5.865056818181818, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 628, "pid": 0, "tid": 7, "ts": 6950381108918.806, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381108897.631, "dur": 25.748 , "args": { "External id": 140,"cbid": 41, "correlation": 628 } }, { "ph": "s", "id": 628, "pid": 1728719, "tid": 1728719, "ts": 6950381108897.631, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381108923.549, "dur": 0.581 , "args": { "External id": 140,"cbid": 131, "correlation": 629 } }, { "ph": "s", "id": 629, "pid": 1728719, "tid": 1728719, "ts": 6950381108923.549, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 1728719, "tid": 1728719, "ts": 6950381109146.514, "dur": 0.170 , "args": { "External id": 145,"cbid": 317, "correlation": 643 } }, { "ph": "f", "id": 643, "pid": 1728719, "tid": 1728719, "ts": 6950381109146.514, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, "ts": 6950381109159.252, "dur": 1.184 , "args": { "External id": 145,"device": 0, "context": 1, "stream": 7, "correlation": 645, "bytes": 33032, "memory bandwidth (GB/s)": 27.89864864864865, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 645, "pid": 0, "tid": 7, "ts": 6950381109159.252, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109147.015, "dur": 5.290 , "args": { "External id": 145,"cbid": 41, "correlation": 645 } }, { "ph": "s", "id": 645, "pid": 1728719, "tid": 1728719, "ts": 6950381109147.015, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109152.405, "dur": 8.997 , "args": { "External id": 145,"cbid": 131, "correlation": 646 } }, { "ph": "s", "id": 646, "pid": 1728719, "tid": 1728719, "ts": 6950381109152.405, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, "ts": 6950381109177.971, "dur": 0.800 , "args": { "External id": 150,"device": 0, "context": 1, "stream": 7, "correlation": 657, "bytes": 33032, "memory bandwidth (GB/s)": 41.29, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 657, "pid": 0, "tid": 7, "ts": 6950381109177.971, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109169.918, "dur": 6.191 , "args": { "External id": 150,"cbid": 41, "correlation": 657 } }, { "ph": "s", "id": 657, "pid": 1728719, "tid": 1728719, "ts": 6950381109169.918, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, "ts": 6950381109184.947, "dur": 0.768 , "args": { "External id": 153,"device": 0, "context": 1, "stream": 7, "correlation": 669, "bytes": 33032, "memory bandwidth (GB/s)": 43.010416666666664, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 669, "pid": 0, "tid": 7, "ts": 6950381109184.947, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109179.766, "dur": 2.655 , "args": { "External id": 153,"cbid": 41, "correlation": 669 } }, { "ph": "s", "id": 669, "pid": 1728719, "tid": 1728719, "ts": 6950381109179.766, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 1728719, "tid": 1728719, "ts": 6950381109302.074, "dur": 0.110 , "args": { "External id": 158,"cbid": 317, "correlation": 684 } }, { "ph": "f", "id": 684, "pid": 1728719, "tid": 1728719, "ts": 6950381109302.074, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, "ts": 6950381109312.306, "dur": 1.184 , "args": { "External id": 158,"device": 0, "context": 1, "stream": 7, "correlation": 686, "bytes": 32768, "memory bandwidth (GB/s)": 27.675675675675677, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 686, "pid": 0, "tid": 7, "ts": 6950381109312.306, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109302.384, "dur": 3.246 , "args": { "External id": 158,"cbid": 41, "correlation": 686 } }, { "ph": "s", "id": 686, "pid": 1728719, "tid": 1728719, "ts": 6950381109302.384, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109305.730, "dur": 8.687 , "args": { "External id": 158,"cbid": 131, "correlation": 687 } }, { "ph": "s", "id": 687, "pid": 1728719, "tid": 1728719, "ts": 6950381109305.730, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109321.580, "dur": 0.190 , "args": { "External id": 163,"cbid": 200, "correlation": 705 } }, { "ph": "f", "id": 705, "pid": 1728719, "tid": 1728719, "ts": 6950381109321.580, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109321.971, "dur": 0.410 , "args": { "External id": 163,"cbid": 251, "correlation": 707 } }, { "ph": "f", "id": 707, "pid": 1728719, "tid": 1728719, "ts": 6950381109321.971, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109324.024, "dur": 0.111 , "args": { "External id": 163,"cbid": 200, "correlation": 718 } }, { "ph": "f", "id": 718, "pid": 1728719, "tid": 1728719, "ts": 6950381109324.024, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109324.265, "dur": 0.170 , "args": { "External id": 163,"cbid": 251, "correlation": 720 } }, { "ph": "f", "id": 720, "pid": 1728719, "tid": 1728719, "ts": 6950381109324.265, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceKernel >::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, unsigned long long, cuda::std::__4::plus, int, cuda::std::__4::__identity>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, int*, unsigned long long, at_cuda_detail::cub::GridEvenShare, cuda::std::__4::plus, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381109330.002, "dur": 0.992 , "args": { "External id": 163,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 724, "registers per thread": 34, "shared memory": 44, "blocks per SM": 0.010638297535479069, "warps per SM": 0.08510638028383255, "grid": [2, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 724, "pid": 0, "tid": 7, "ts": 6950381109330.002, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109325.006, "dur": 2.856 , "args": { "External id": 163,"cbid": 211, "correlation": 724 } }, { "ph": "s", "id": 724, "pid": 1728719, "tid": 1728719, "ts": 6950381109325.006, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109327.972, "dur": 0.050 , "args": { "External id": 163,"cbid": 11, "correlation": 725 } }, { "ph": "f", "id": 725, "pid": 1728719, "tid": 1728719, "ts": 6950381109327.972, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109328.082, "dur": 0.020 , "args": { "External id": 163,"cbid": 11, "correlation": 726 } }, { "ph": "f", "id": 726, "pid": 1728719, "tid": 1728719, "ts": 6950381109328.082, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceSingleTileKernel >::Policy1000, int*, int*, int, cuda::std::__4::plus, int, int, cuda::std::__4::__identity>(int*, int*, int, cuda::std::__4::plus, int, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381109331.858, "dur": 1.248 , "args": { "External id": 163,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 730, "registers per thread": 30, "shared memory": 44, "blocks per SM": 0.005319148767739534, "warps per SM": 0.042553190141916275, "grid": [1, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 730, "pid": 0, "tid": 7, "ts": 6950381109331.858, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109328.503, "dur": 1.823 , "args": { "External id": 163,"cbid": 211, "correlation": 730 } }, { "ph": "s", "id": 730, "pid": 1728719, "tid": 1728719, "ts": 6950381109328.503, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109330.426, "dur": 0.030 , "args": { "External id": 163,"cbid": 11, "correlation": 731 } }, { "ph": "f", "id": 731, "pid": 1728719, "tid": 1728719, "ts": 6950381109330.426, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109330.527, "dur": 0.030 , "args": { "External id": 163,"cbid": 11, "correlation": 732 } }, { "ph": "f", "id": 732, "pid": 1728719, "tid": 1728719, "ts": 6950381109330.527, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381109336.370, "dur": 0.352 , "args": { "External id": 163,"device": 0, "context": 1, "stream": 7, "correlation": 736, "bytes": 4, "memory bandwidth (GB/s)": 0.011363636363636364, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 736, "pid": 0, "tid": 7, "ts": 6950381109336.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109331.839, "dur": 2.334 , "args": { "External id": 163,"cbid": 41, "correlation": 736 } }, { "ph": "s", "id": 736, "pid": 1728719, "tid": 1728719, "ts": 6950381109331.839, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109334.263, "dur": 3.367 , "args": { "External id": 163,"cbid": 131, "correlation": 737 } }, { "ph": "s", "id": 737, "pid": 1728719, "tid": 1728719, "ts": 6950381109334.263, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceCompactInitKernel, int*>(at_cuda_detail::cub::ScanTileState, int, int*)", "pid": 0, "tid": 7, "ts": 6950381109346.546, "dur": 0.704 , "args": { "External id": 163,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 762, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 762, "pid": 0, "tid": 7, "ts": 6950381109346.546, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109342.158, "dur": 1.954 , "args": { "External id": 163,"cbid": 211, "correlation": 762 } }, { "ph": "s", "id": 762, "pid": 1728719, "tid": 1728719, "ts": 6950381109342.158, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109344.222, "dur": 0.030 , "args": { "External id": 163,"cbid": 11, "correlation": 763 } }, { "ph": "f", "id": 763, "pid": 1728719, "tid": 1728719, "ts": 6950381109344.222, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109344.302, "dur": 0.030 , "args": { "External id": 163,"cbid": 11, "correlation": 764 } }, { "ph": "f", "id": 764, "pid": 1728719, "tid": 1728719, "ts": 6950381109344.302, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::select::DeviceSelectSweepKernel::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, at_cuda_detail::cub::detail::select::streaming_context_t, (at_cuda_detail::cub::SelectImpl)0>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, int, at_cuda_detail::cub::detail::select::streaming_context_t, at_cuda_detail::cub::detail::vsmem_t)", "pid": 0, "tid": 7, "ts": 6950381109348.370, "dur": 2.880 , "args": { "External id": 163,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 768, "registers per thread": 64, "shared memory": 28160, "blocks per SM": 0.010638297535479069, "warps per SM": 0.10638297349214554, "grid": [2, 1, 1], "block": [320, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "REGS|SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 3, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 29184}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 768, "pid": 0, "tid": 7, "ts": 6950381109348.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109344.823, "dur": 1.924 , "args": { "External id": 163,"cbid": 211, "correlation": 768 } }, { "ph": "s", "id": 768, "pid": 1728719, "tid": 1728719, "ts": 6950381109344.823, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109346.827, "dur": 0.030 , "args": { "External id": 163,"cbid": 11, "correlation": 769 } }, { "ph": "f", "id": 769, "pid": 1728719, "tid": 1728719, "ts": 6950381109346.827, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109346.907, "dur": 0.020 , "args": { "External id": 163,"cbid": 11, "correlation": 770 } }, { "ph": "f", "id": 770, "pid": 1728719, "tid": 1728719, "ts": 6950381109346.907, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381109360.113, "dur": 5.281 , "args": { "External id": 162,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 778, "registers per thread": 48, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 10, "limitingFactors": "REGS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 778, "pid": 0, "tid": 7, "ts": 6950381109360.113, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109355.643, "dur": 2.395 , "args": { "External id": 162,"cbid": 211, "correlation": 778 } }, { "ph": "s", "id": 778, "pid": 1728719, "tid": 1728719, "ts": 6950381109355.643, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 1728719, "tid": 1728719, "ts": 6950381109481.447, "dur": 0.161 , "args": { "External id": 179,"cbid": 317, "correlation": 795 } }, { "ph": "f", "id": 795, "pid": 1728719, "tid": 1728719, "ts": 6950381109481.447, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, "ts": 6950381109492.112, "dur": 1.184 , "args": { "External id": 179,"device": 0, "context": 1, "stream": 7, "correlation": 797, "bytes": 32768, "memory bandwidth (GB/s)": 27.675675675675677, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 797, "pid": 0, "tid": 7, "ts": 6950381109492.112, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109481.848, "dur": 3.226 , "args": { "External id": 179,"cbid": 41, "correlation": 797 } }, { "ph": "s", "id": 797, "pid": 1728719, "tid": 1728719, "ts": 6950381109481.848, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109485.164, "dur": 9.067 , "args": { "External id": 179,"cbid": 131, "correlation": 798 } }, { "ph": "s", "id": 798, "pid": 1728719, "tid": 1728719, "ts": 6950381109485.164, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109499.591, "dur": 0.130 , "args": { "External id": 184,"cbid": 200, "correlation": 816 } }, { "ph": "f", "id": 816, "pid": 1728719, "tid": 1728719, "ts": 6950381109499.591, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109499.892, "dur": 0.280 , "args": { "External id": 184,"cbid": 251, "correlation": 818 } }, { "ph": "f", "id": 818, "pid": 1728719, "tid": 1728719, "ts": 6950381109499.892, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109501.555, "dur": 0.090 , "args": { "External id": 184,"cbid": 200, "correlation": 829 } }, { "ph": "f", "id": 829, "pid": 1728719, "tid": 1728719, "ts": 6950381109501.555, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109501.775, "dur": 0.181 , "args": { "External id": 184,"cbid": 251, "correlation": 831 } }, { "ph": "f", "id": 831, "pid": 1728719, "tid": 1728719, "ts": 6950381109501.775, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceKernel >::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, unsigned long long, cuda::std::__4::plus, int, cuda::std::__4::__identity>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, int*, unsigned long long, at_cuda_detail::cub::GridEvenShare, cuda::std::__4::plus, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381109506.736, "dur": 1.024 , "args": { "External id": 184,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 835, "registers per thread": 34, "shared memory": 44, "blocks per SM": 0.010638297535479069, "warps per SM": 0.08510638028383255, "grid": [2, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 835, "pid": 0, "tid": 7, "ts": 6950381109506.736, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109502.386, "dur": 2.004 , "args": { "External id": 184,"cbid": 211, "correlation": 835 } }, { "ph": "s", "id": 835, "pid": 1728719, "tid": 1728719, "ts": 6950381109502.386, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109504.470, "dur": 0.030 , "args": { "External id": 184,"cbid": 11, "correlation": 836 } }, { "ph": "f", "id": 836, "pid": 1728719, "tid": 1728719, "ts": 6950381109504.470, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109504.550, "dur": 0.020 , "args": { "External id": 184,"cbid": 11, "correlation": 837 } }, { "ph": "f", "id": 837, "pid": 1728719, "tid": 1728719, "ts": 6950381109504.550, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceSingleTileKernel >::Policy1000, int*, int*, int, cuda::std::__4::plus, int, int, cuda::std::__4::__identity>(int*, int*, int, cuda::std::__4::plus, int, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381109508.080, "dur": 0.992 , "args": { "External id": 184,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 841, "registers per thread": 30, "shared memory": 44, "blocks per SM": 0.005319148767739534, "warps per SM": 0.042553190141916275, "grid": [1, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 841, "pid": 0, "tid": 7, "ts": 6950381109508.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109504.931, "dur": 1.673 , "args": { "External id": 184,"cbid": 211, "correlation": 841 } }, { "ph": "s", "id": 841, "pid": 1728719, "tid": 1728719, "ts": 6950381109504.931, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109506.684, "dur": 0.030 , "args": { "External id": 184,"cbid": 11, "correlation": 842 } }, { "ph": "f", "id": 842, "pid": 1728719, "tid": 1728719, "ts": 6950381109506.684, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109506.795, "dur": 0.020 , "args": { "External id": 184,"cbid": 11, "correlation": 843 } }, { "ph": "f", "id": 843, "pid": 1728719, "tid": 1728719, "ts": 6950381109506.795, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381109511.088, "dur": 0.352 , "args": { "External id": 184,"device": 0, "context": 1, "stream": 7, "correlation": 847, "bytes": 4, "memory bandwidth (GB/s)": 0.011363636363636364, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 847, "pid": 0, "tid": 7, "ts": 6950381109511.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109507.706, "dur": 2.024 , "args": { "External id": 184,"cbid": 41, "correlation": 847 } }, { "ph": "s", "id": 847, "pid": 1728719, "tid": 1728719, "ts": 6950381109507.706, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109509.810, "dur": 2.525 , "args": { "External id": 184,"cbid": 131, "correlation": 848 } }, { "ph": "s", "id": 848, "pid": 1728719, "tid": 1728719, "ts": 6950381109509.810, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceCompactInitKernel, int*>(at_cuda_detail::cub::ScanTileState, int, int*)", "pid": 0, "tid": 7, "ts": 6950381109519.760, "dur": 0.704 , "args": { "External id": 184,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 873, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 873, "pid": 0, "tid": 7, "ts": 6950381109519.760, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109515.862, "dur": 1.653 , "args": { "External id": 184,"cbid": 211, "correlation": 873 } }, { "ph": "s", "id": 873, "pid": 1728719, "tid": 1728719, "ts": 6950381109515.862, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109517.595, "dur": 0.030 , "args": { "External id": 184,"cbid": 11, "correlation": 874 } }, { "ph": "f", "id": 874, "pid": 1728719, "tid": 1728719, "ts": 6950381109517.595, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109517.675, "dur": 0.020 , "args": { "External id": 184,"cbid": 11, "correlation": 875 } }, { "ph": "f", "id": 875, "pid": 1728719, "tid": 1728719, "ts": 6950381109517.675, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::select::DeviceSelectSweepKernel::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, at_cuda_detail::cub::detail::select::streaming_context_t, (at_cuda_detail::cub::SelectImpl)0>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, int, at_cuda_detail::cub::detail::select::streaming_context_t, at_cuda_detail::cub::detail::vsmem_t)", "pid": 0, "tid": 7, "ts": 6950381109521.104, "dur": 2.176 , "args": { "External id": 184,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 879, "registers per thread": 64, "shared memory": 28160, "blocks per SM": 0.010638297535479069, "warps per SM": 0.10638297349214554, "grid": [2, 1, 1], "block": [320, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "REGS|SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 3, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 29184}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 879, "pid": 0, "tid": 7, "ts": 6950381109521.104, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109518.046, "dur": 1.823 , "args": { "External id": 184,"cbid": 211, "correlation": 879 } }, { "ph": "s", "id": 879, "pid": 1728719, "tid": 1728719, "ts": 6950381109518.046, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109519.939, "dur": 0.030 , "args": { "External id": 184,"cbid": 11, "correlation": 880 } }, { "ph": "f", "id": 880, "pid": 1728719, "tid": 1728719, "ts": 6950381109519.939, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109520.029, "dur": 0.020 , "args": { "External id": 184,"cbid": 11, "correlation": 881 } }, { "ph": "f", "id": 881, "pid": 1728719, "tid": 1728719, "ts": 6950381109520.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381109536.528, "dur": 2.560 , "args": { "External id": 183,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 889, "registers per thread": 48, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 10, "limitingFactors": "REGS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 889, "pid": 0, "tid": 7, "ts": 6950381109536.528, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109526.181, "dur": 7.724 , "args": { "External id": 183,"cbid": 211, "correlation": 889 } }, { "ph": "s", "id": 889, "pid": 1728719, "tid": 1728719, "ts": 6950381109526.181, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_gather_kernel<16, long>(char*, char*, long*, int, long, long, long, long, bool)", "pid": 0, "tid": 7, "ts": 6950381109548.975, "dur": 2.689 , "args": { "External id": 196,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 905, "registers per thread": 24, "shared memory": 0, "blocks per SM": 21.962766647338867, "warps per SM": 21.962766647338867, "grid": [4129, 1, 1], "block": [32, 1, 1], "est. achieved occupancy %": 46, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "BLOCKS", "blockLimitRegs": 84, "blockLimitSharedMem": 100, "blockLimitWarps": 48, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 768, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 905, "pid": 0, "tid": 7, "ts": 6950381109548.975, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109544.775, "dur": 1.994 , "args": { "External id": 196,"cbid": 211, "correlation": 905 } }, { "ph": "s", "id": 905, "pid": 1728719, "tid": 1728719, "ts": 6950381109544.775, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_gather_kernel<16, long>(char*, char*, long*, int, long, long, long, long, bool)", "pid": 0, "tid": 7, "ts": 6950381109557.007, "dur": 2.688 , "args": { "External id": 200,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 921, "registers per thread": 24, "shared memory": 0, "blocks per SM": 21.962766647338867, "warps per SM": 21.962766647338867, "grid": [4129, 1, 1], "block": [32, 1, 1], "est. achieved occupancy %": 46, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "BLOCKS", "blockLimitRegs": 84, "blockLimitSharedMem": 100, "blockLimitWarps": 48, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 768, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 921, "pid": 0, "tid": 7, "ts": 6950381109557.007, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109552.830, "dur": 1.774 , "args": { "External id": 200,"cbid": 211, "correlation": 921 } }, { "ph": "s", "id": 921, "pid": 1728719, "tid": 1728719, "ts": 6950381109552.830, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_gather_kernel<16, long>(char*, char*, long*, int, long, long, long, long, bool)", "pid": 0, "tid": 7, "ts": 6950381109563.663, "dur": 2.688 , "args": { "External id": 204,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 937, "registers per thread": 24, "shared memory": 0, "blocks per SM": 21.962766647338867, "warps per SM": 21.962766647338867, "grid": [4129, 1, 1], "block": [32, 1, 1], "est. achieved occupancy %": 46, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "BLOCKS", "blockLimitRegs": 84, "blockLimitSharedMem": 100, "blockLimitWarps": 48, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 768, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 937, "pid": 0, "tid": 7, "ts": 6950381109563.663, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109559.503, "dur": 1.963 , "args": { "External id": 204,"cbid": 211, "correlation": 937 } }, { "ph": "s", "id": 937, "pid": 1728719, "tid": 1728719, "ts": 6950381109559.503, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_vectorized, unsigned int, 2, 128, 1, 16, 2>(char*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381109576.143, "dur": 1.472 , "args": { "External id": 208,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 951, "registers per thread": 19, "shared memory": 0, "blocks per SM": 7.212766170501709, "warps per SM": 28.851064682006836, "grid": [452, 3, 1], "block": [128, 1, 1], "est. achieved occupancy %": 60, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 951, "pid": 0, "tid": 7, "ts": 6950381109576.143, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109568.009, "dur": 2.634 , "args": { "External id": 208,"cbid": 211, "correlation": 951 } }, { "ph": "s", "id": 951, "pid": 1728719, "tid": 1728719, "ts": 6950381109568.009, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109589.779, "dur": 0.130 , "args": { "External id": 210,"cbid": 200, "correlation": 969 } }, { "ph": "f", "id": 969, "pid": 1728719, "tid": 1728719, "ts": 6950381109589.779, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109590.070, "dur": 0.340 , "args": { "External id": 210,"cbid": 251, "correlation": 971 } }, { "ph": "f", "id": 971, "pid": 1728719, "tid": 1728719, "ts": 6950381109590.070, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109591.833, "dur": 0.080 , "args": { "External id": 210,"cbid": 200, "correlation": 982 } }, { "ph": "f", "id": 982, "pid": 1728719, "tid": 1728719, "ts": 6950381109591.833, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109592.043, "dur": 0.181 , "args": { "External id": 210,"cbid": 251, "correlation": 984 } }, { "ph": "f", "id": 984, "pid": 1728719, "tid": 1728719, "ts": 6950381109592.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceKernel >::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, unsigned long long, cuda::std::__4::plus, int, cuda::std::__4::__identity>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, int*, unsigned long long, at_cuda_detail::cub::GridEvenShare, cuda::std::__4::plus, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381109596.847, "dur": 0.992 , "args": { "External id": 210,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 988, "registers per thread": 34, "shared memory": 44, "blocks per SM": 0.010638297535479069, "warps per SM": 0.08510638028383255, "grid": [2, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 988, "pid": 0, "tid": 7, "ts": 6950381109596.847, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109592.624, "dur": 1.984 , "args": { "External id": 210,"cbid": 211, "correlation": 988 } }, { "ph": "s", "id": 988, "pid": 1728719, "tid": 1728719, "ts": 6950381109592.624, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109594.688, "dur": 0.040 , "args": { "External id": 210,"cbid": 11, "correlation": 989 } }, { "ph": "f", "id": 989, "pid": 1728719, "tid": 1728719, "ts": 6950381109594.688, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109594.788, "dur": 0.020 , "args": { "External id": 210,"cbid": 11, "correlation": 990 } }, { "ph": "f", "id": 990, "pid": 1728719, "tid": 1728719, "ts": 6950381109594.788, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::reduce::DeviceReduceSingleTileKernel >::Policy1000, int*, int*, int, cuda::std::__4::plus, int, int, cuda::std::__4::__identity>(int*, int*, int, cuda::std::__4::plus, int, cuda::std::__4::__identity)", "pid": 0, "tid": 7, "ts": 6950381109598.351, "dur": 0.960 , "args": { "External id": 210,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 994, "registers per thread": 30, "shared memory": 44, "blocks per SM": 0.005319148767739534, "warps per SM": 0.042553190141916275, "grid": [1, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 994, "pid": 0, "tid": 7, "ts": 6950381109598.351, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109595.199, "dur": 1.683 , "args": { "External id": 210,"cbid": 211, "correlation": 994 } }, { "ph": "s", "id": 994, "pid": 1728719, "tid": 1728719, "ts": 6950381109595.199, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109597.013, "dur": 0.030 , "args": { "External id": 210,"cbid": 11, "correlation": 995 } }, { "ph": "f", "id": 995, "pid": 1728719, "tid": 1728719, "ts": 6950381109597.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109597.123, "dur": 0.030 , "args": { "External id": 210,"cbid": 11, "correlation": 996 } }, { "ph": "f", "id": 996, "pid": 1728719, "tid": 1728719, "ts": 6950381109597.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381109603.215, "dur": 0.352 , "args": { "External id": 210,"device": 0, "context": 1, "stream": 7, "correlation": 1000, "bytes": 4, "memory bandwidth (GB/s)": 0.011363636363636364, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 1000, "pid": 0, "tid": 7, "ts": 6950381109603.215, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109598.225, "dur": 3.146 , "args": { "External id": 210,"cbid": 41, "correlation": 1000 } }, { "ph": "s", "id": 1000, "pid": 1728719, "tid": 1728719, "ts": 6950381109598.225, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109601.461, "dur": 3.106 , "args": { "External id": 210,"cbid": 131, "correlation": 1001 } }, { "ph": "s", "id": 1001, "pid": 1728719, "tid": 1728719, "ts": 6950381109601.461, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceCompactInitKernel, int*>(at_cuda_detail::cub::ScanTileState, int, int*)", "pid": 0, "tid": 7, "ts": 6950381109612.143, "dur": 0.672 , "args": { "External id": 210,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1026, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1026, "pid": 0, "tid": 7, "ts": 6950381109612.143, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109608.233, "dur": 1.674 , "args": { "External id": 210,"cbid": 211, "correlation": 1026 } }, { "ph": "s", "id": 1026, "pid": 1728719, "tid": 1728719, "ts": 6950381109608.233, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109609.987, "dur": 0.030 , "args": { "External id": 210,"cbid": 11, "correlation": 1027 } }, { "ph": "f", "id": 1027, "pid": 1728719, "tid": 1728719, "ts": 6950381109609.987, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109610.067, "dur": 0.020 , "args": { "External id": 210,"cbid": 11, "correlation": 1028 } }, { "ph": "f", "id": 1028, "pid": 1728719, "tid": 1728719, "ts": 6950381109610.067, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::select::DeviceSelectSweepKernel::Policy1000, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, at_cuda_detail::cub::detail::select::streaming_context_t, (at_cuda_detail::cub::SelectImpl)0>(thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::counting_iterator, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::transform_iterator, bool const*, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default, thrust::THRUST_300001_SM_750_800_860_900_1000_1200_NS::use_default>, long*, int*, at_cuda_detail::cub::ScanTileState, at_cuda_detail::cub::NullType, at_cuda_detail::cub::NullType, int, int, at_cuda_detail::cub::detail::select::streaming_context_t, at_cuda_detail::cub::detail::vsmem_t)", "pid": 0, "tid": 7, "ts": 6950381109613.743, "dur": 3.744 , "args": { "External id": 210,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1032, "registers per thread": 64, "shared memory": 28160, "blocks per SM": 0.010638297535479069, "warps per SM": 0.10638297349214554, "grid": [2, 1, 1], "block": [320, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "REGS|SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 3, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 29184}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1032, "pid": 0, "tid": 7, "ts": 6950381109613.743, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109610.458, "dur": 1.753 , "args": { "External id": 210,"cbid": 211, "correlation": 1032 } }, { "ph": "s", "id": 1032, "pid": 1728719, "tid": 1728719, "ts": 6950381109610.458, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109612.291, "dur": 0.030 , "args": { "External id": 210,"cbid": 11, "correlation": 1033 } }, { "ph": "f", "id": 1033, "pid": 1728719, "tid": 1728719, "ts": 6950381109612.291, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109612.381, "dur": 0.020 , "args": { "External id": 210,"cbid": 11, "correlation": 1034 } }, { "ph": "f", "id": 1034, "pid": 1728719, "tid": 1728719, "ts": 6950381109612.381, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<2, at::native::FillFunctor, std::array, false>(int, at::native::FillFunctor, std::array)", "pid": 0, "tid": 7, "ts": 6950381109634.959, "dur": 0.703 , "args": { "External id": 223,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1052, "registers per thread": 22, "shared memory": 0, "blocks per SM": 0.026595745235681534, "warps per SM": 0.10638298094272614, "grid": [5, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1052, "pid": 0, "tid": 7, "ts": 6950381109634.959, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109630.615, "dur": 2.024 , "args": { "External id": 223,"cbid": 211, "correlation": 1052 } }, { "ph": "s", "id": 1052, "pid": 1728719, "tid": 1728719, "ts": 6950381109630.615, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::elementwise_kernel_with_index(int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, function_traits::result_type*)", "pid": 0, "tid": 7, "ts": 6950381109646.574, "dur": 0.672 , "args": { "External id": 226,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1070, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.010638297535479069, "grid": [1, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 64, "blockLimitSharedMem": 100, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1024, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1070, "pid": 0, "tid": 7, "ts": 6950381109646.574, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109642.046, "dur": 2.916 , "args": { "External id": 226,"cbid": 211, "correlation": 1070 } }, { "ph": "s", "id": 1070, "pid": 1728719, "tid": 1728719, "ts": 6950381109642.046, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 1728719, "tid": 1728719, "ts": 6950381109653.197, "dur": 0.251 , "args": { "External id": 232,"cbid": 317, "correlation": 1085 } }, { "ph": "f", "id": 1085, "pid": 1728719, "tid": 1728719, "ts": 6950381109653.197, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, "ts": 6950381109658.030, "dur": 0.256 , "args": { "External id": 232,"device": 0, "context": 1, "stream": 7, "correlation": 1087, "bytes": 8, "memory bandwidth (GB/s)": 0.03125, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1087, "pid": 0, "tid": 7, "ts": 6950381109658.030, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109653.828, "dur": 2.234 , "args": { "External id": 232,"cbid": 41, "correlation": 1087 } }, { "ph": "s", "id": 1087, "pid": 1728719, "tid": 1728719, "ts": 6950381109653.828, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109656.203, "dur": 3.055 , "args": { "External id": 232,"cbid": 131, "correlation": 1088 } }, { "ph": "s", "id": 1088, "pid": 1728719, "tid": 1728719, "ts": 6950381109656.203, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceScanInitKernel >(at_cuda_detail::cub::ScanTileState, int)", "pid": 0, "tid": 7, "ts": 6950381109675.214, "dur": 0.736 , "args": { "External id": 242,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1119, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1119, "pid": 0, "tid": 7, "ts": 6950381109675.214, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109670.970, "dur": 1.854 , "args": { "External id": 242,"cbid": 211, "correlation": 1119 } }, { "ph": "s", "id": 1119, "pid": 1728719, "tid": 1728719, "ts": 6950381109670.970, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109672.944, "dur": 0.040 , "args": { "External id": 242,"cbid": 11, "correlation": 1120 } }, { "ph": "f", "id": 1120, "pid": 1728719, "tid": 1728719, "ts": 6950381109672.944, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109673.064, "dur": 0.020 , "args": { "External id": 242,"cbid": 11, "correlation": 1121 } }, { "ph": "f", "id": 1121, "pid": 1728719, "tid": 1728719, "ts": 6950381109673.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "pid": 1728719, "tid": 1728719, "ts": 6950381109673.345, "dur": 0.300 , "args": { "External id": 242,"cbid": 251, "correlation": 1124 } }, { "ph": "f", "id": 1124, "pid": 1728719, "tid": 1728719, "ts": 6950381109673.345, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109674.056, "dur": 0.160 , "args": { "External id": 242,"cbid": 200, "correlation": 1128 } }, { "ph": "f", "id": 1128, "pid": 1728719, "tid": 1728719, "ts": 6950381109674.056, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void at_cuda_detail::cub::detail::scan::DeviceScanKernel >::Policy1000, long const*, long*, at_cuda_detail::cub::ScanTileState, std::plus, at_cuda_detail::cub::NullType, unsigned int, long, false, at_cuda_detail::cub::NullType>(long const*, long*, at_cuda_detail::cub::ScanTileState, int, std::plus, at_cuda_detail::cub::NullType, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381109678.734, "dur": 1.152 , "args": { "External id": 242,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1131, "registers per thread": 40, "shared memory": 7184, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS|SMEM", "blockLimitRegs": 12, "blockLimitSharedMem": 12, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 8320}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1131, "pid": 0, "tid": 7, "ts": 6950381109678.734, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109674.657, "dur": 2.114 , "args": { "External id": 242,"cbid": 211, "correlation": 1131 } }, { "ph": "s", "id": 1131, "pid": 1728719, "tid": 1728719, "ts": 6950381109674.657, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109676.861, "dur": 0.030 , "args": { "External id": 242,"cbid": 11, "correlation": 1132 } }, { "ph": "f", "id": 1132, "pid": 1728719, "tid": 1728719, "ts": 6950381109676.861, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaPeekAtLastError", "pid": 1728719, "tid": 1728719, "ts": 6950381109676.951, "dur": 0.030 , "args": { "External id": 242,"cbid": 11, "correlation": 1133 } }, { "ph": "f", "id": 1133, "pid": 1728719, "tid": 1728719, "ts": 6950381109676.951, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381109685.486, "dur": 0.352 , "args": { "External id": 247,"device": 0, "context": 1, "stream": 7, "correlation": 1142, "bytes": 8, "memory bandwidth (GB/s)": 0.022727272727272728, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 1142, "pid": 0, "tid": 7, "ts": 6950381109685.486, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109681.410, "dur": 2.064 , "args": { "External id": 247,"cbid": 41, "correlation": 1142 } }, { "ph": "s", "id": 1142, "pid": 1728719, "tid": 1728719, "ts": 6950381109681.410, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109683.604, "dur": 3.176 , "args": { "External id": 247,"cbid": 131, "correlation": 1143 } }, { "ph": "s", "id": 1143, "pid": 1728719, "tid": 1728719, "ts": 6950381109683.604, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::compare_scalar_kernel(at::TensorIteratorBase&, at::native::(anonymous namespace)::OpType, long)::{lambda(long)#1}, std::array, false>(int, at::native::compare_scalar_kernel(at::TensorIteratorBase&, at::native::(anonymous namespace)::OpType, long)::{lambda(long)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381109695.822, "dur": 0.928 , "args": { "External id": 248,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1153, "registers per thread": 40, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1153, "pid": 0, "tid": 7, "ts": 6950381109695.822, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109691.699, "dur": 1.823 , "args": { "External id": 248,"cbid": 211, "correlation": 1153 } }, { "ph": "s", "id": 1153, "pid": 1728719, "tid": 1728719, "ts": 6950381109691.699, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, "ts": 6950381109704.494, "dur": 0.960 , "args": { "External id": 253,"device": 0, "context": 1, "stream": 7, "correlation": 1163, "bytes": 1, "memory bandwidth (GB/s)": 0.0010416666666666667, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1163, "pid": 0, "tid": 7, "ts": 6950381109704.494, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109699.002, "dur": 3.367 , "args": { "External id": 253,"cbid": 41, "correlation": 1163 } }, { "ph": "s", "id": 1163, "pid": 1728719, "tid": 1728719, "ts": 6950381109699.002, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381109707.854, "dur": 0.352 , "args": { "External id": 255,"device": 0, "context": 1, "stream": 7, "correlation": 1171, "bytes": 1, "memory bandwidth (GB/s)": 0.002840909090909091, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 1171, "pid": 0, "tid": 7, "ts": 6950381109707.854, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109704.483, "dur": 1.783 , "args": { "External id": 255,"cbid": 41, "correlation": 1171 } }, { "ph": "s", "id": 1171, "pid": 1728719, "tid": 1728719, "ts": 6950381109704.483, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109706.406, "dur": 2.725 , "args": { "External id": 255,"cbid": 131, "correlation": 1172 } }, { "ph": "s", "id": 1172, "pid": 1728719, "tid": 1728719, "ts": 6950381109706.406, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void compute_cuda_kernel(long const*, long const*, long*, long, long)", "pid": 0, "tid": 7, "ts": 6950381109716.942, "dur": 1.472 , "args": { "External id": 241,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1185, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.08510638028383255, "grid": [1, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 100, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1185, "pid": 0, "tid": 7, "ts": 6950381109716.942, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109712.678, "dur": 1.933 , "args": { "External id": 241,"cbid": 211, "correlation": 1185 } }, { "ph": "s", "id": 1185, "pid": 1728719, "tid": 1728719, "ts": 6950381109712.678, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::_scatter_gather_elementwise_kernel<128, 8, at::native::_cuda_scatter_gather_internal_kernel, long>::operator()(at::TensorIterator&, long, long, long, at::native::TensorAssign const&)::{lambda(int)#1}>(int, at::native::_cuda_scatter_gather_internal_kernel, long>::operator()(at::TensorIterator&, long, long, long, at::native::TensorAssign const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381109729.197, "dur": 2.945 , "args": { "External id": 263,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1207, "registers per thread": 28, "shared memory": 0, "blocks per SM": 0.021276595070958138, "warps per SM": 0.08510638028383255, "grid": [4, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1207, "pid": 0, "tid": 7, "ts": 6950381109729.197, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109724.901, "dur": 2.514 , "args": { "External id": 263,"cbid": 211, "correlation": 1207 } }, { "ph": "s", "id": 1207, "pid": 1728719, "tid": 1728719, "ts": 6950381109724.901, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381109739.053, "dur": 3.072 , "args": { "External id": 267,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1217, "registers per thread": 48, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 10, "limitingFactors": "REGS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1217, "pid": 0, "tid": 7, "ts": 6950381109739.053, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109734.989, "dur": 1.964 , "args": { "External id": 267,"cbid": 211, "correlation": 1217 } }, { "ph": "s", "id": 1217, "pid": 1728719, "tid": 1728719, "ts": 6950381109734.989, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, std::array, false>(int, at::native::FillFunctor, std::array)", "pid": 0, "tid": 7, "ts": 6950381109749.261, "dur": 0.704 , "args": { "External id": 275,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1234, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.01595744676887989, "warps per SM": 0.06382978707551956, "grid": [3, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1234, "pid": 0, "tid": 7, "ts": 6950381109749.261, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109744.978, "dur": 2.274 , "args": { "External id": 275,"cbid": 211, "correlation": 1234 } }, { "ph": "s", "id": 1234, "pid": 1728719, "tid": 1728719, "ts": 6950381109744.978, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 1728719, "tid": 1728719, "ts": 6950381109759.816, "dur": 0.180 , "args": { "External id": 284,"cbid": 317, "correlation": 1256 } }, { "ph": "f", "id": 1256, "pid": 1728719, "tid": 1728719, "ts": 6950381109759.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, "ts": 6950381109764.429, "dur": 0.256 , "args": { "External id": 284,"device": 0, "context": 1, "stream": 7, "correlation": 1258, "bytes": 1, "memory bandwidth (GB/s)": 0.00390625, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1258, "pid": 0, "tid": 7, "ts": 6950381109764.429, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109760.256, "dur": 2.084 , "args": { "External id": 284,"cbid": 41, "correlation": 1258 } }, { "ph": "s", "id": 1258, "pid": 1728719, "tid": 1728719, "ts": 6950381109760.256, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109762.481, "dur": 3.135 , "args": { "External id": 284,"cbid": 131, "correlation": 1259 } }, { "ph": "s", "id": 1259, "pid": 1728719, "tid": 1728719, "ts": 6950381109762.481, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_put_kernel_impl >(at::TensorIterator&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&, bool)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381109773.197, "dur": 4.928 , "args": { "External id": 280,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1264, "registers per thread": 48, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 10, "limitingFactors": "REGS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1264, "pid": 0, "tid": 7, "ts": 6950381109773.197, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109769.033, "dur": 2.264 , "args": { "External id": 280,"cbid": 211, "correlation": 1264 } }, { "ph": "s", "id": 1264, "pid": 1728719, "tid": 1728719, "ts": 6950381109769.033, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, std::array, false>(int, at::native::FillFunctor, std::array)", "pid": 0, "tid": 7, "ts": 6950381109788.301, "dur": 0.736 , "args": { "External id": 293,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1281, "registers per thread": 16, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 32, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 2048, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1281, "pid": 0, "tid": 7, "ts": 6950381109788.301, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109784.021, "dur": 2.194 , "args": { "External id": 293,"cbid": 211, "correlation": 1281 } }, { "ph": "s", "id": 1281, "pid": 1728719, "tid": 1728719, "ts": 6950381109784.021, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_alignedK_contig, unsigned int, 1, 128, 1, 8>(at::native::(anonymous namespace)::OpaqueType<2u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381109804.685, "dur": 0.864 , "args": { "External id": 294,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1295, "registers per thread": 34, "shared memory": 0, "blocks per SM": 0.010638297535479069, "warps per SM": 0.042553190141916275, "grid": [1, 2, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1295, "pid": 0, "tid": 7, "ts": 6950381109804.685, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109791.845, "dur": 2.555 , "args": { "External id": 294,"cbid": 211, "correlation": 1295 } }, { "ph": "s", "id": 1295, "pid": 1728719, "tid": 1728719, "ts": 6950381109791.845, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, 4, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, "ts": 6950381109810.509, "dur": 1.088 , "args": { "External id": 298,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1309, "registers per thread": 31, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1309, "pid": 0, "tid": 7, "ts": 6950381109810.509, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109805.921, "dur": 2.004 , "args": { "External id": 298,"cbid": 211, "correlation": 1309 } }, { "ph": "s", "id": 1309, "pid": 1728719, "tid": 1728719, "ts": 6950381109805.921, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, std::array, false>(int, at::native::AUnaryFunctor >, std::array)", "pid": 0, "tid": 7, "ts": 6950381109818.861, "dur": 0.928 , "args": { "External id": 299,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1321, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1321, "pid": 0, "tid": 7, "ts": 6950381109818.861, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109814.708, "dur": 2.214 , "args": { "External id": 299,"cbid": 211, "correlation": 1321 } }, { "ph": "s", "id": 1321, "pid": 1728719, "tid": 1728719, "ts": 6950381109814.708, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381109833.068, "dur": 1.248 , "args": { "External id": 305,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1338, "registers per thread": 21, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1338, "pid": 0, "tid": 7, "ts": 6950381109833.068, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109828.954, "dur": 1.914 , "args": { "External id": 305,"cbid": 211, "correlation": 1338 } }, { "ph": "s", "id": 1338, "pid": 1728719, "tid": 1728719, "ts": 6950381109828.954, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::cos_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array, false>(int, at::native::cos_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381109840.812, "dur": 1.440 , "args": { "External id": 306,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1349, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1349, "pid": 0, "tid": 7, "ts": 6950381109840.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109836.378, "dur": 2.294 , "args": { "External id": 306,"cbid": 211, "correlation": 1349 } }, { "ph": "s", "id": 1349, "pid": 1728719, "tid": 1728719, "ts": 6950381109836.378, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::sin_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array, false>(int, at::native::sin_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381109847.532, "dur": 1.408 , "args": { "External id": 307,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1360, "registers per thread": 36, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1360, "pid": 0, "tid": 7, "ts": 6950381109847.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109843.301, "dur": 2.084 , "args": { "External id": 307,"cbid": 211, "correlation": 1360 } }, { "ph": "s", "id": 1360, "pid": 1728719, "tid": 1728719, "ts": 6950381109843.301, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_vectorized, unsigned int, 2, 128, 1, 16, 4>(char*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, "ts": 6950381109855.500, "dur": 0.864 , "args": { "External id": 308,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1372, "registers per thread": 19, "shared memory": 0, "blocks per SM": 0.010638297535479069, "warps per SM": 0.042553190141916275, "grid": [1, 2, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1372, "pid": 0, "tid": 7, "ts": 6950381109855.500, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109850.534, "dur": 2.966 , "args": { "External id": 308,"cbid": 211, "correlation": 1372 } }, { "ph": "s", "id": 1372, "pid": 1728719, "tid": 1728719, "ts": 6950381109850.534, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::bfloat16_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda(float)#1}, std::array, false>(int, at::native::bfloat16_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda(float)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381109867.052, "dur": 0.960 , "args": { "External id": 313,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1386, "registers per thread": 40, "shared memory": 0, "blocks per SM": 0.005319148767739534, "warps per SM": 0.021276595070958138, "grid": [1, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1386, "pid": 0, "tid": 7, "ts": 6950381109867.052, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109863.128, "dur": 1.773 , "args": { "External id": 313,"cbid": 211, "correlation": 1386 } }, { "ph": "s", "id": 1386, "pid": 1728719, "tid": 1728719, "ts": 6950381109863.128, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109883.085, "dur": 0.301 , "args": { "External id": 319,"cbid": 200, "correlation": 1402 } }, { "ph": "f", "id": 1402, "pid": 1728719, "tid": 1728719, "ts": 6950381109883.085, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381109888.684, "dur": 3.456 , "args": { "External id": 319,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1404, "registers per thread": 84, "shared memory": 33792, "blocks per SM": 0.6808510422706604, "warps per SM": 2.7234041690826416, "grid": [8, 16, 1], "block": [128, 1, 1], "est. achieved occupancy %": 6, "occupancy": {"activeBlocksPerMultiprocessor": 2, "limitingFactors": "SMEM", "blockLimitRegs": 5, "blockLimitSharedMem": 2, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 11264, "allocatedSharedMemPerBlock": 34816}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1404, "pid": 0, "tid": 7, "ts": 6950381109888.684, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109884.057, "dur": 2.635 , "args": { "External id": 319,"cbid": 307, "correlation": 1404 } }, { "ph": "s", "id": 1404, "pid": 1728719, "tid": 1728719, "ts": 6950381109884.057, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#6}::operator()() const::{lambda(c10::BFloat16)#1}, std::array, false>(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#6}::operator()() const::{lambda(c10::BFloat16)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381109899.372, "dur": 1.120 , "args": { "External id": 320,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1414, "registers per thread": 38, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1414, "pid": 0, "tid": 7, "ts": 6950381109899.372, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109894.727, "dur": 2.514 , "args": { "External id": 320,"cbid": 211, "correlation": 1414 } }, { "ph": "s", "id": 1414, "pid": 1728719, "tid": 1728719, "ts": 6950381109894.727, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109907.150, "dur": 0.100 , "args": { "External id": 326,"cbid": 200, "correlation": 1429 } }, { "ph": "f", "id": 1429, "pid": 1728719, "tid": 1728719, "ts": 6950381109907.150, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7, "ts": 6950381109915.659, "dur": 0.321 , "args": { "External id": 326,"device": 0, "context": 1, "stream": 7, "correlation": 1431, "bytes": 512, "memory bandwidth (GB/s)": 1.5950155763239875, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1431, "pid": 0, "tid": 7, "ts": 6950381109915.659, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemsetAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109907.811, "dur": 6.091 , "args": { "External id": 326,"cbid": 51, "correlation": 1431 } }, { "ph": "s", "id": 1431, "pid": 1728719, "tid": 1728719, "ts": 6950381109907.811, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x1_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381109917.163, "dur": 25.824 , "args": { "External id": 326,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1432, "registers per thread": 72, "shared memory": 17408, "blocks per SM": 2.7234041690826416, "warps per SM": 10.893616676330566, "grid": [8, 16, 4], "block": [128, 1, 1], "est. achieved occupancy %": 23, "occupancy": {"activeBlocksPerMultiprocessor": 5, "limitingFactors": "SMEM", "blockLimitRegs": 7, "blockLimitSharedMem": 5, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 9216, "allocatedSharedMemPerBlock": 18432}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1432, "pid": 0, "tid": 7, "ts": 6950381109917.163, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109914.163, "dur": 2.174 , "args": { "External id": 326,"cbid": 307, "correlation": 1432 } }, { "ph": "s", "id": 1432, "pid": 1728719, "tid": 1728719, "ts": 6950381109914.163, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#6}::operator()() const::{lambda(c10::BFloat16)#1}, std::array, false>(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#6}::operator()() const::{lambda(c10::BFloat16)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381109943.243, "dur": 1.152 , "args": { "External id": 327,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1442, "registers per thread": 38, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1442, "pid": 0, "tid": 7, "ts": 6950381109943.243, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109925.294, "dur": 1.913 , "args": { "External id": 327,"cbid": 211, "correlation": 1442 } }, { "ph": "s", "id": 1442, "pid": 1728719, "tid": 1728719, "ts": 6950381109925.294, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381109934.751, "dur": 0.110 , "args": { "External id": 333,"cbid": 200, "correlation": 1457 } }, { "ph": "f", "id": 1457, "pid": 1728719, "tid": 1728719, "ts": 6950381109934.751, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7, "ts": 6950381109945.611, "dur": 0.480 , "args": { "External id": 333,"device": 0, "context": 1, "stream": 7, "correlation": 1459, "bytes": 2048, "memory bandwidth (GB/s)": 4.266666666666667, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1459, "pid": 0, "tid": 7, "ts": 6950381109945.611, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemsetAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109935.342, "dur": 1.503 , "args": { "External id": 333,"cbid": 51, "correlation": 1459 } }, { "ph": "s", "id": 1459, "pid": 1728719, "tid": 1728719, "ts": 6950381109935.342, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381109946.763, "dur": 86.623 , "args": { "External id": 333,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1460, "registers per thread": 56, "shared memory": 9216, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [8, 64, 2], "block": [128, 1, 1], "est. achieved occupancy %": 45, "occupancy": {"activeBlocksPerMultiprocessor": 9, "limitingFactors": "REGS", "blockLimitRegs": 9, "blockLimitSharedMem": 10, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 7168, "allocatedSharedMemPerBlock": 10240}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1460, "pid": 0, "tid": 7, "ts": 6950381109946.763, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109937.005, "dur": 1.944 , "args": { "External id": 333,"cbid": 307, "correlation": 1460 } }, { "ph": "s", "id": 1460, "pid": 1728719, "tid": 1728719, "ts": 6950381109937.005, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::bitwise_not_kernel_cuda(at::TensorIteratorBase&)::{lambda(bool)#1}, std::array, false>(int, at::native::bitwise_not_kernel_cuda(at::TensorIteratorBase&)::{lambda(bool)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381110033.642, "dur": 1.984 , "args": { "External id": 334,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1471, "registers per thread": 30, "shared memory": 0, "blocks per SM": 0.026595745235681534, "warps per SM": 0.10638298094272614, "grid": [5, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1471, "pid": 0, "tid": 7, "ts": 6950381110033.642, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109945.231, "dur": 2.344 , "args": { "External id": 334,"cbid": 211, "correlation": 1471 } }, { "ph": "s", "id": 1471, "pid": 1728719, "tid": 1728719, "ts": 6950381109945.231, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, 4, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, std::array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, "ts": 6950381110035.914, "dur": 2.272 , "args": { "External id": 341,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1492, "registers per thread": 33, "shared memory": 0, "blocks per SM": 0.04787234216928482, "warps per SM": 0.19148936867713928, "grid": [9, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1492, "pid": 0, "tid": 7, "ts": 6950381110035.914, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109957.063, "dur": 2.504 , "args": { "External id": 341,"cbid": 211, "correlation": 1492 } }, { "ph": "s", "id": 1492, "pid": 1728719, "tid": 1728719, "ts": 6950381109957.063, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp::operator()(at::TensorIterator&)::{lambda(long, long)#1}>, unsigned int, long, 4, 4> >(at::native::ReduceOp::operator()(at::TensorIterator&)::{lambda(long, long)#1}>, unsigned int, long, 4, 4>)", "pid": 0, "tid": 7, "ts": 6950381110038.474, "dur": 2.368 , "args": { "External id": 336,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1503, "registers per thread": 38, "shared memory": 4112, "blocks per SM": 0.005319148767739534, "warps per SM": 0.08510638028383255, "grid": [1, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 19, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 5248}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1503, "pid": 0, "tid": 7, "ts": 6950381110038.474, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381109963.365, "dur": 2.544 , "args": { "External id": 336,"cbid": 211, "correlation": 1503 } }, { "ph": "s", "id": 1503, "pid": 1728719, "tid": 1728719, "ts": 6950381109963.365, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pinned)", "pid": 0, "tid": 7, "ts": 6950381110043.178, "dur": 0.320 , "args": { "External id": 343,"device": 0, "context": 1, "stream": 7, "correlation": 1510, "bytes": 8, "memory bandwidth (GB/s)": 0.025, "graph id": 0, "graph node id": 0, "channel": 12, "channel_type": 2 } }, { "ph": "f", "id": 1510, "pid": 0, "tid": 7, "ts": 6950381110043.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381109970.337, "dur": 3.788 , "args": { "External id": 343,"cbid": 41, "correlation": 1510 } }, { "ph": "s", "id": 1510, "pid": 1728719, "tid": 1728719, "ts": 6950381109970.337, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381109974.285, "dur": 70.331 , "args": { "External id": 343,"cbid": 131, "correlation": 1511 } }, { "ph": "s", "id": 1511, "pid": 1728719, "tid": 1728719, "ts": 6950381109974.285, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381110270.632, "dur": 20.735 , "args": { "External id": 361,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1535, "registers per thread": 29, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS", "blockLimitRegs": 4, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1535, "pid": 0, "tid": 7, "ts": 6950381110270.632, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110258.945, "dur": 4.228 , "args": { "External id": 361,"cbid": 307, "correlation": 1535 } }, { "ph": "s", "id": 1535, "pid": 1728719, "tid": 1728719, "ts": 6950381110258.945, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381110298.471, "dur": 9.344 , "args": { "External id": 362,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1549, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1549, "pid": 0, "tid": 7, "ts": 6950381110298.471, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110294.230, "dur": 2.275 , "args": { "External id": 362,"cbid": 652, "correlation": 1549 } }, { "ph": "s", "id": 1549, "pid": 1728719, "tid": 1728719, "ts": 6950381110294.230, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381110316.967, "dur": 11.648 , "args": { "External id": 363,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1563, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1563, "pid": 0, "tid": 7, "ts": 6950381110316.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110312.474, "dur": 1.794 , "args": { "External id": 363,"cbid": 652, "correlation": 1563 } }, { "ph": "s", "id": 1563, "pid": 1728719, "tid": 1728719, "ts": 6950381110312.474, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381110331.143, "dur": 10.656 , "args": { "External id": 364,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1577, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1577, "pid": 0, "tid": 7, "ts": 6950381110331.143, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110326.330, "dur": 1.793 , "args": { "External id": 364,"cbid": 652, "correlation": 1577 } }, { "ph": "s", "id": 1577, "pid": 1728719, "tid": 1728719, "ts": 6950381110326.330, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110344.865, "dur": 0.360 , "args": { "External id": 365,"cbid": 200, "correlation": 1592 } }, { "ph": "f", "id": 1592, "pid": 1728719, "tid": 1728719, "ts": 6950381110344.865, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110345.816, "dur": 0.662 , "args": { "External id": 365,"cbid": 468, "correlation": 1593 } }, { "ph": "f", "id": 1593, "pid": 1728719, "tid": 1728719, "ts": 6950381110345.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110347.750, "dur": 0.180 , "args": { "External id": 365,"cbid": 1, "correlation": 1595 } }, { "ph": "f", "id": 1595, "pid": 1728719, "tid": 1728719, "ts": 6950381110347.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110348.111, "dur": 0.170 , "args": { "External id": 365,"cbid": 468, "correlation": 1596 } }, { "ph": "f", "id": 1596, "pid": 1728719, "tid": 1728719, "ts": 6950381110348.111, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110348.451, "dur": 0.020 , "args": { "External id": 365,"cbid": 1, "correlation": 1598 } }, { "ph": "f", "id": 1598, "pid": 1728719, "tid": 1728719, "ts": 6950381110348.451, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110348.792, "dur": 0.100 , "args": { "External id": 365,"cbid": 468, "correlation": 1599 } }, { "ph": "f", "id": 1599, "pid": 1728719, "tid": 1728719, "ts": 6950381110348.792, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110349.123, "dur": 0.020 , "args": { "External id": 365,"cbid": 1, "correlation": 1601 } }, { "ph": "f", "id": 1601, "pid": 1728719, "tid": 1728719, "ts": 6950381110349.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110349.443, "dur": 1.954 , "args": { "External id": 365,"cbid": 273, "correlation": 1602 } }, { "ph": "f", "id": 1602, "pid": 1728719, "tid": 1728719, "ts": 6950381110349.443, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381110357.543, "dur": 212.861 , "args": { "External id": 365,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1605, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1605, "pid": 0, "tid": 7, "ts": 6950381110357.543, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110352.228, "dur": 3.337 , "args": { "External id": 365,"cbid": 211, "correlation": 1605 } }, { "ph": "s", "id": 1605, "pid": 1728719, "tid": 1728719, "ts": 6950381110352.228, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381110570.692, "dur": 8.672 , "args": { "External id": 377,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1618, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1618, "pid": 0, "tid": 7, "ts": 6950381110570.692, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110374.830, "dur": 2.425 , "args": { "External id": 377,"cbid": 307, "correlation": 1618 } }, { "ph": "s", "id": 1618, "pid": 1728719, "tid": 1728719, "ts": 6950381110374.830, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381110579.652, "dur": 33.280 , "args": { "External id": 382,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1641, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1641, "pid": 0, "tid": 7, "ts": 6950381110579.652, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110415.927, "dur": 2.475 , "args": { "External id": 382,"cbid": 307, "correlation": 1641 } }, { "ph": "s", "id": 1641, "pid": 1728719, "tid": 1728719, "ts": 6950381110415.927, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110428.841, "dur": 0.120 , "args": { "External id": 383,"cbid": 200, "correlation": 1656 } }, { "ph": "f", "id": 1656, "pid": 1728719, "tid": 1728719, "ts": 6950381110428.841, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110429.071, "dur": 0.161 , "args": { "External id": 383,"cbid": 468, "correlation": 1657 } }, { "ph": "f", "id": 1657, "pid": 1728719, "tid": 1728719, "ts": 6950381110429.071, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110429.532, "dur": 0.030 , "args": { "External id": 383,"cbid": 1, "correlation": 1659 } }, { "ph": "f", "id": 1659, "pid": 1728719, "tid": 1728719, "ts": 6950381110429.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110429.683, "dur": 0.100 , "args": { "External id": 383,"cbid": 468, "correlation": 1660 } }, { "ph": "f", "id": 1660, "pid": 1728719, "tid": 1728719, "ts": 6950381110429.683, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110429.953, "dur": 0.020 , "args": { "External id": 383,"cbid": 1, "correlation": 1662 } }, { "ph": "f", "id": 1662, "pid": 1728719, "tid": 1728719, "ts": 6950381110429.953, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110430.073, "dur": 0.100 , "args": { "External id": 383,"cbid": 468, "correlation": 1663 } }, { "ph": "f", "id": 1663, "pid": 1728719, "tid": 1728719, "ts": 6950381110430.073, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110430.354, "dur": 0.020 , "args": { "External id": 383,"cbid": 1, "correlation": 1665 } }, { "ph": "f", "id": 1665, "pid": 1728719, "tid": 1728719, "ts": 6950381110430.354, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110430.514, "dur": 0.611 , "args": { "External id": 383,"cbid": 273, "correlation": 1666 } }, { "ph": "f", "id": 1666, "pid": 1728719, "tid": 1728719, "ts": 6950381110430.514, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381110613.220, "dur": 232.829 , "args": { "External id": 383,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1669, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1669, "pid": 0, "tid": 7, "ts": 6950381110613.220, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110431.556, "dur": 2.034 , "args": { "External id": 383,"cbid": 211, "correlation": 1669 } }, { "ph": "s", "id": 1669, "pid": 1728719, "tid": 1728719, "ts": 6950381110431.556, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381110846.370, "dur": 8.927 , "args": { "External id": 395,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1678, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1678, "pid": 0, "tid": 7, "ts": 6950381110846.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110441.054, "dur": 1.863 , "args": { "External id": 395,"cbid": 307, "correlation": 1678 } }, { "ph": "s", "id": 1678, "pid": 1728719, "tid": 1728719, "ts": 6950381110441.054, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381110855.585, "dur": 25.760 , "args": { "External id": 400,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1701, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1701, "pid": 0, "tid": 7, "ts": 6950381110855.585, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110461.532, "dur": 1.743 , "args": { "External id": 400,"cbid": 307, "correlation": 1701 } }, { "ph": "s", "id": 1701, "pid": 1728719, "tid": 1728719, "ts": 6950381110461.532, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110472.522, "dur": 0.121 , "args": { "External id": 401,"cbid": 200, "correlation": 1716 } }, { "ph": "f", "id": 1716, "pid": 1728719, "tid": 1728719, "ts": 6950381110472.522, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110472.733, "dur": 0.130 , "args": { "External id": 401,"cbid": 468, "correlation": 1717 } }, { "ph": "f", "id": 1717, "pid": 1728719, "tid": 1728719, "ts": 6950381110472.733, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110473.063, "dur": 0.030 , "args": { "External id": 401,"cbid": 1, "correlation": 1719 } }, { "ph": "f", "id": 1719, "pid": 1728719, "tid": 1728719, "ts": 6950381110473.063, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110473.244, "dur": 0.100 , "args": { "External id": 401,"cbid": 468, "correlation": 1720 } }, { "ph": "f", "id": 1720, "pid": 1728719, "tid": 1728719, "ts": 6950381110473.244, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110473.484, "dur": 0.030 , "args": { "External id": 401,"cbid": 1, "correlation": 1722 } }, { "ph": "f", "id": 1722, "pid": 1728719, "tid": 1728719, "ts": 6950381110473.484, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110473.614, "dur": 0.100 , "args": { "External id": 401,"cbid": 468, "correlation": 1723 } }, { "ph": "f", "id": 1723, "pid": 1728719, "tid": 1728719, "ts": 6950381110473.614, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110473.875, "dur": 0.030 , "args": { "External id": 401,"cbid": 1, "correlation": 1725 } }, { "ph": "f", "id": 1725, "pid": 1728719, "tid": 1728719, "ts": 6950381110473.875, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110473.995, "dur": 0.471 , "args": { "External id": 401,"cbid": 273, "correlation": 1726 } }, { "ph": "f", "id": 1726, "pid": 1728719, "tid": 1728719, "ts": 6950381110473.995, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381110881.633, "dur": 234.558 , "args": { "External id": 401,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1729, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1729, "pid": 0, "tid": 7, "ts": 6950381110881.633, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110474.827, "dur": 1.813 , "args": { "External id": 401,"cbid": 211, "correlation": 1729 } }, { "ph": "s", "id": 1729, "pid": 1728719, "tid": 1728719, "ts": 6950381110474.827, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381111116.479, "dur": 58.015 , "args": { "External id": 413,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1743, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1743, "pid": 0, "tid": 7, "ts": 6950381111116.479, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110490.536, "dur": 2.194 , "args": { "External id": 413,"cbid": 307, "correlation": 1743 } }, { "ph": "s", "id": 1743, "pid": 1728719, "tid": 1728719, "ts": 6950381110490.536, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381111174.750, "dur": 64.287 , "args": { "External id": 414,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1753, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1753, "pid": 0, "tid": 7, "ts": 6950381111174.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110508.700, "dur": 2.244 , "args": { "External id": 414,"cbid": 307, "correlation": 1753 } }, { "ph": "s", "id": 1753, "pid": 1728719, "tid": 1728719, "ts": 6950381110508.700, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381111239.325, "dur": 39.328 , "args": { "External id": 415,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1762, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1762, "pid": 0, "tid": 7, "ts": 6950381111239.325, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110524.709, "dur": 1.904 , "args": { "External id": 415,"cbid": 307, "correlation": 1762 } }, { "ph": "s", "id": 1762, "pid": 1728719, "tid": 1728719, "ts": 6950381110524.709, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110553.333, "dur": 0.190 , "args": { "External id": 423,"cbid": 200, "correlation": 1797 } }, { "ph": "f", "id": 1797, "pid": 1728719, "tid": 1728719, "ts": 6950381110553.333, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110553.593, "dur": 0.100 , "args": { "External id": 423,"cbid": 200, "correlation": 1798 } }, { "ph": "f", "id": 1798, "pid": 1728719, "tid": 1728719, "ts": 6950381110553.593, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110553.904, "dur": 0.791 , "args": { "External id": 423,"cbid": 273, "correlation": 1799 } }, { "ph": "f", "id": 1799, "pid": 1728719, "tid": 1728719, "ts": 6950381110553.904, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381111278.909, "dur": 865.942 , "args": { "External id": 423,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1801, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1801, "pid": 0, "tid": 7, "ts": 6950381111278.909, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110555.146, "dur": 2.725 , "args": { "External id": 423,"cbid": 211, "correlation": 1801 } }, { "ph": "s", "id": 1801, "pid": 1728719, "tid": 1728719, "ts": 6950381110555.146, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381112145.139, "dur": 9.760 , "args": { "External id": 432,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1821, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1821, "pid": 0, "tid": 7, "ts": 6950381112145.139, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110579.421, "dur": 1.844 , "args": { "External id": 432,"cbid": 652, "correlation": 1821 } }, { "ph": "s", "id": 1821, "pid": 1728719, "tid": 1728719, "ts": 6950381110579.421, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110591.894, "dur": 0.181 , "args": { "External id": 433,"cbid": 200, "correlation": 1836 } }, { "ph": "f", "id": 1836, "pid": 1728719, "tid": 1728719, "ts": 6950381110591.894, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110592.205, "dur": 0.200 , "args": { "External id": 433,"cbid": 468, "correlation": 1837 } }, { "ph": "f", "id": 1837, "pid": 1728719, "tid": 1728719, "ts": 6950381110592.205, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110592.776, "dur": 0.040 , "args": { "External id": 433,"cbid": 1, "correlation": 1839 } }, { "ph": "f", "id": 1839, "pid": 1728719, "tid": 1728719, "ts": 6950381110592.776, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110592.896, "dur": 0.111 , "args": { "External id": 433,"cbid": 468, "correlation": 1840 } }, { "ph": "f", "id": 1840, "pid": 1728719, "tid": 1728719, "ts": 6950381110592.896, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110593.157, "dur": 0.020 , "args": { "External id": 433,"cbid": 1, "correlation": 1842 } }, { "ph": "f", "id": 1842, "pid": 1728719, "tid": 1728719, "ts": 6950381110593.157, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110593.287, "dur": 0.090 , "args": { "External id": 433,"cbid": 468, "correlation": 1843 } }, { "ph": "f", "id": 1843, "pid": 1728719, "tid": 1728719, "ts": 6950381110593.287, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110593.588, "dur": 0.020 , "args": { "External id": 433,"cbid": 1, "correlation": 1845 } }, { "ph": "f", "id": 1845, "pid": 1728719, "tid": 1728719, "ts": 6950381110593.588, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110593.748, "dur": 0.771 , "args": { "External id": 433,"cbid": 273, "correlation": 1846 } }, { "ph": "f", "id": 1846, "pid": 1728719, "tid": 1728719, "ts": 6950381110593.748, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381112155.187, "dur": 212.574 , "args": { "External id": 433,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1849, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1849, "pid": 0, "tid": 7, "ts": 6950381112155.187, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110594.910, "dur": 1.954 , "args": { "External id": 433,"cbid": 211, "correlation": 1849 } }, { "ph": "s", "id": 1849, "pid": 1728719, "tid": 1728719, "ts": 6950381110594.910, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381112368.049, "dur": 47.615 , "args": { "External id": 445,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1861, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1861, "pid": 0, "tid": 7, "ts": 6950381112368.049, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110611.701, "dur": 1.914 , "args": { "External id": 445,"cbid": 307, "correlation": 1861 } }, { "ph": "s", "id": 1861, "pid": 1728719, "tid": 1728719, "ts": 6950381110611.701, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381112415.985, "dur": 14.047 , "args": { "External id": 446,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1867, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1867, "pid": 0, "tid": 7, "ts": 6950381112415.985, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110622.241, "dur": 1.713 , "args": { "External id": 446,"cbid": 652, "correlation": 1867 } }, { "ph": "s", "id": 1867, "pid": 1728719, "tid": 1728719, "ts": 6950381110622.241, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381112430.352, "dur": 13.248 , "args": { "External id": 447,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1873, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1873, "pid": 0, "tid": 7, "ts": 6950381112430.352, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110632.360, "dur": 1.463 , "args": { "External id": 447,"cbid": 652, "correlation": 1873 } }, { "ph": "s", "id": 1873, "pid": 1728719, "tid": 1728719, "ts": 6950381110632.360, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110645.154, "dur": 0.320 , "args": { "External id": 448,"cbid": 200, "correlation": 1892 } }, { "ph": "f", "id": 1892, "pid": 1728719, "tid": 1728719, "ts": 6950381110645.154, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110645.594, "dur": 0.131 , "args": { "External id": 448,"cbid": 468, "correlation": 1893 } }, { "ph": "f", "id": 1893, "pid": 1728719, "tid": 1728719, "ts": 6950381110645.594, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110645.925, "dur": 0.020 , "args": { "External id": 448,"cbid": 1, "correlation": 1895 } }, { "ph": "f", "id": 1895, "pid": 1728719, "tid": 1728719, "ts": 6950381110645.925, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110646.015, "dur": 0.100 , "args": { "External id": 448,"cbid": 468, "correlation": 1896 } }, { "ph": "f", "id": 1896, "pid": 1728719, "tid": 1728719, "ts": 6950381110646.015, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110646.276, "dur": 0.030 , "args": { "External id": 448,"cbid": 1, "correlation": 1898 } }, { "ph": "f", "id": 1898, "pid": 1728719, "tid": 1728719, "ts": 6950381110646.276, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110646.386, "dur": 0.100 , "args": { "External id": 448,"cbid": 468, "correlation": 1899 } }, { "ph": "f", "id": 1899, "pid": 1728719, "tid": 1728719, "ts": 6950381110646.386, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110646.666, "dur": 0.020 , "args": { "External id": 448,"cbid": 1, "correlation": 1901 } }, { "ph": "f", "id": 1901, "pid": 1728719, "tid": 1728719, "ts": 6950381110646.666, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110646.797, "dur": 0.440 , "args": { "External id": 448,"cbid": 273, "correlation": 1902 } }, { "ph": "f", "id": 1902, "pid": 1728719, "tid": 1728719, "ts": 6950381110646.797, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381112443.920, "dur": 683.385 , "args": { "External id": 448,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1905, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1905, "pid": 0, "tid": 7, "ts": 6950381112443.920, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110647.598, "dur": 1.804 , "args": { "External id": 448,"cbid": 211, "correlation": 1905 } }, { "ph": "s", "id": 1905, "pid": 1728719, "tid": 1728719, "ts": 6950381110647.598, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110660.202, "dur": 0.100 , "args": { "External id": 460,"cbid": 200, "correlation": 1929 } }, { "ph": "f", "id": 1929, "pid": 1728719, "tid": 1728719, "ts": 6950381110660.202, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110660.402, "dur": 0.120 , "args": { "External id": 460,"cbid": 468, "correlation": 1930 } }, { "ph": "f", "id": 1930, "pid": 1728719, "tid": 1728719, "ts": 6950381110660.402, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110660.723, "dur": 0.020 , "args": { "External id": 460,"cbid": 1, "correlation": 1932 } }, { "ph": "f", "id": 1932, "pid": 1728719, "tid": 1728719, "ts": 6950381110660.723, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110660.813, "dur": 0.100 , "args": { "External id": 460,"cbid": 468, "correlation": 1933 } }, { "ph": "f", "id": 1933, "pid": 1728719, "tid": 1728719, "ts": 6950381110660.813, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110661.053, "dur": 0.020 , "args": { "External id": 460,"cbid": 1, "correlation": 1935 } }, { "ph": "f", "id": 1935, "pid": 1728719, "tid": 1728719, "ts": 6950381110661.053, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110661.183, "dur": 0.101 , "args": { "External id": 460,"cbid": 468, "correlation": 1936 } }, { "ph": "f", "id": 1936, "pid": 1728719, "tid": 1728719, "ts": 6950381110661.183, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110661.434, "dur": 0.020 , "args": { "External id": 460,"cbid": 1, "correlation": 1938 } }, { "ph": "f", "id": 1938, "pid": 1728719, "tid": 1728719, "ts": 6950381110661.434, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110661.534, "dur": 0.391 , "args": { "External id": 460,"cbid": 273, "correlation": 1939 } }, { "ph": "f", "id": 1939, "pid": 1728719, "tid": 1728719, "ts": 6950381110661.534, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381113127.593, "dur": 691.768 , "args": { "External id": 460,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1942, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1942, "pid": 0, "tid": 7, "ts": 6950381113127.593, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110662.265, "dur": 1.714 , "args": { "External id": 460,"cbid": 211, "correlation": 1942 } }, { "ph": "s", "id": 1942, "pid": 1728719, "tid": 1728719, "ts": 6950381110662.265, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381113819.681, "dur": 154.911 , "args": { "External id": 472,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1955, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1955, "pid": 0, "tid": 7, "ts": 6950381113819.681, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110677.303, "dur": 2.325 , "args": { "External id": 472,"cbid": 307, "correlation": 1955 } }, { "ph": "s", "id": 1955, "pid": 1728719, "tid": 1728719, "ts": 6950381110677.303, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381113974.848, "dur": 57.343 , "args": { "External id": 473,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1961, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1961, "pid": 0, "tid": 7, "ts": 6950381113974.848, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110691.310, "dur": 1.953 , "args": { "External id": 473,"cbid": 652, "correlation": 1961 } }, { "ph": "s", "id": 1961, "pid": 1728719, "tid": 1728719, "ts": 6950381110691.310, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110704.364, "dur": 0.100 , "args": { "External id": 474,"cbid": 200, "correlation": 1980 } }, { "ph": "f", "id": 1980, "pid": 1728719, "tid": 1728719, "ts": 6950381110704.364, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110704.614, "dur": 0.131 , "args": { "External id": 474,"cbid": 468, "correlation": 1981 } }, { "ph": "f", "id": 1981, "pid": 1728719, "tid": 1728719, "ts": 6950381110704.614, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110704.935, "dur": 0.030 , "args": { "External id": 474,"cbid": 1, "correlation": 1983 } }, { "ph": "f", "id": 1983, "pid": 1728719, "tid": 1728719, "ts": 6950381110704.935, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110705.025, "dur": 0.100 , "args": { "External id": 474,"cbid": 468, "correlation": 1984 } }, { "ph": "f", "id": 1984, "pid": 1728719, "tid": 1728719, "ts": 6950381110705.025, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110705.266, "dur": 0.020 , "args": { "External id": 474,"cbid": 1, "correlation": 1986 } }, { "ph": "f", "id": 1986, "pid": 1728719, "tid": 1728719, "ts": 6950381110705.266, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110705.376, "dur": 0.100 , "args": { "External id": 474,"cbid": 468, "correlation": 1987 } }, { "ph": "f", "id": 1987, "pid": 1728719, "tid": 1728719, "ts": 6950381110705.376, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110705.626, "dur": 0.020 , "args": { "External id": 474,"cbid": 1, "correlation": 1989 } }, { "ph": "f", "id": 1989, "pid": 1728719, "tid": 1728719, "ts": 6950381110705.626, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110705.746, "dur": 0.391 , "args": { "External id": 474,"cbid": 273, "correlation": 1990 } }, { "ph": "f", "id": 1990, "pid": 1728719, "tid": 1728719, "ts": 6950381110705.746, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381114032.479, "dur": 701.049 , "args": { "External id": 474,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 1993, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 1993, "pid": 0, "tid": 7, "ts": 6950381114032.479, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110706.498, "dur": 1.673 , "args": { "External id": 474,"cbid": 211, "correlation": 1993 } }, { "ph": "s", "id": 1993, "pid": 1728719, "tid": 1728719, "ts": 6950381110706.498, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381114733.816, "dur": 43.935 , "args": { "External id": 486,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2004, "registers per thread": 31, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2004, "pid": 0, "tid": 7, "ts": 6950381114733.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110722.417, "dur": 1.954 , "args": { "External id": 486,"cbid": 307, "correlation": 2004 } }, { "ph": "s", "id": 2004, "pid": 1728719, "tid": 1728719, "ts": 6950381110722.417, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381114778.039, "dur": 21.536 , "args": { "External id": 498,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2016, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2016, "pid": 0, "tid": 7, "ts": 6950381114778.039, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110862.438, "dur": 3.086 , "args": { "External id": 498,"cbid": 307, "correlation": 2016 } }, { "ph": "s", "id": 2016, "pid": 1728719, "tid": 1728719, "ts": 6950381110862.438, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381114799.895, "dur": 12.736 , "args": { "External id": 499,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2030, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2030, "pid": 0, "tid": 7, "ts": 6950381114799.895, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110882.976, "dur": 1.914 , "args": { "External id": 499,"cbid": 652, "correlation": 2030 } }, { "ph": "s", "id": 2030, "pid": 1728719, "tid": 1728719, "ts": 6950381110882.976, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381114812.919, "dur": 12.287 , "args": { "External id": 500,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2044, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2044, "pid": 0, "tid": 7, "ts": 6950381114812.919, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110896.471, "dur": 1.603 , "args": { "External id": 500,"cbid": 652, "correlation": 2044 } }, { "ph": "s", "id": 2044, "pid": 1728719, "tid": 1728719, "ts": 6950381110896.471, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381114825.495, "dur": 10.527 , "args": { "External id": 501,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2058, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2058, "pid": 0, "tid": 7, "ts": 6950381114825.495, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381110908.474, "dur": 1.493 , "args": { "External id": 501,"cbid": 652, "correlation": 2058 } }, { "ph": "s", "id": 2058, "pid": 1728719, "tid": 1728719, "ts": 6950381110908.474, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110923.151, "dur": 0.231 , "args": { "External id": 502,"cbid": 200, "correlation": 2073 } }, { "ph": "f", "id": 2073, "pid": 1728719, "tid": 1728719, "ts": 6950381110923.151, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110923.562, "dur": 0.220 , "args": { "External id": 502,"cbid": 468, "correlation": 2074 } }, { "ph": "f", "id": 2074, "pid": 1728719, "tid": 1728719, "ts": 6950381110923.562, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110924.123, "dur": 0.030 , "args": { "External id": 502,"cbid": 1, "correlation": 2076 } }, { "ph": "f", "id": 2076, "pid": 1728719, "tid": 1728719, "ts": 6950381110924.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110924.233, "dur": 0.100 , "args": { "External id": 502,"cbid": 468, "correlation": 2077 } }, { "ph": "f", "id": 2077, "pid": 1728719, "tid": 1728719, "ts": 6950381110924.233, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110924.504, "dur": 0.020 , "args": { "External id": 502,"cbid": 1, "correlation": 2079 } }, { "ph": "f", "id": 2079, "pid": 1728719, "tid": 1728719, "ts": 6950381110924.504, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110924.624, "dur": 0.100 , "args": { "External id": 502,"cbid": 468, "correlation": 2080 } }, { "ph": "f", "id": 2080, "pid": 1728719, "tid": 1728719, "ts": 6950381110924.624, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110924.894, "dur": 0.020 , "args": { "External id": 502,"cbid": 1, "correlation": 2082 } }, { "ph": "f", "id": 2082, "pid": 1728719, "tid": 1728719, "ts": 6950381110924.894, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110925.075, "dur": 0.841 , "args": { "External id": 502,"cbid": 273, "correlation": 2083 } }, { "ph": "f", "id": 2083, "pid": 1728719, "tid": 1728719, "ts": 6950381110925.075, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381114836.310, "dur": 220.286 , "args": { "External id": 502,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2086, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2086, "pid": 0, "tid": 7, "ts": 6950381114836.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110926.437, "dur": 1.904 , "args": { "External id": 502,"cbid": 211, "correlation": 2086 } }, { "ph": "s", "id": 2086, "pid": 1728719, "tid": 1728719, "ts": 6950381110926.437, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381115056.884, "dur": 8.768 , "args": { "External id": 514,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2099, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2099, "pid": 0, "tid": 7, "ts": 6950381115056.884, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110943.018, "dur": 1.934 , "args": { "External id": 514,"cbid": 307, "correlation": 2099 } }, { "ph": "s", "id": 2099, "pid": 1728719, "tid": 1728719, "ts": 6950381110943.018, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381115065.940, "dur": 36.095 , "args": { "External id": 519,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2122, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2122, "pid": 0, "tid": 7, "ts": 6950381115065.940, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110970.429, "dur": 1.914 , "args": { "External id": 519,"cbid": 307, "correlation": 2122 } }, { "ph": "s", "id": 2122, "pid": 1728719, "tid": 1728719, "ts": 6950381110970.429, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110981.981, "dur": 0.120 , "args": { "External id": 520,"cbid": 200, "correlation": 2137 } }, { "ph": "f", "id": 2137, "pid": 1728719, "tid": 1728719, "ts": 6950381110981.981, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110982.201, "dur": 0.150 , "args": { "External id": 520,"cbid": 468, "correlation": 2138 } }, { "ph": "f", "id": 2138, "pid": 1728719, "tid": 1728719, "ts": 6950381110982.201, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110982.612, "dur": 0.030 , "args": { "External id": 520,"cbid": 1, "correlation": 2140 } }, { "ph": "f", "id": 2140, "pid": 1728719, "tid": 1728719, "ts": 6950381110982.612, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110982.702, "dur": 0.110 , "args": { "External id": 520,"cbid": 468, "correlation": 2141 } }, { "ph": "f", "id": 2141, "pid": 1728719, "tid": 1728719, "ts": 6950381110982.702, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110982.972, "dur": 0.020 , "args": { "External id": 520,"cbid": 1, "correlation": 2143 } }, { "ph": "f", "id": 2143, "pid": 1728719, "tid": 1728719, "ts": 6950381110982.972, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110983.093, "dur": 0.090 , "args": { "External id": 520,"cbid": 468, "correlation": 2144 } }, { "ph": "f", "id": 2144, "pid": 1728719, "tid": 1728719, "ts": 6950381110983.093, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381110983.333, "dur": 0.020 , "args": { "External id": 520,"cbid": 1, "correlation": 2146 } }, { "ph": "f", "id": 2146, "pid": 1728719, "tid": 1728719, "ts": 6950381110983.333, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381110983.453, "dur": 0.451 , "args": { "External id": 520,"cbid": 273, "correlation": 2147 } }, { "ph": "f", "id": 2147, "pid": 1728719, "tid": 1728719, "ts": 6950381110983.453, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381115102.323, "dur": 233.694 , "args": { "External id": 520,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2150, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2150, "pid": 0, "tid": 7, "ts": 6950381115102.323, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110984.295, "dur": 1.773 , "args": { "External id": 520,"cbid": 211, "correlation": 2150 } }, { "ph": "s", "id": 2150, "pid": 1728719, "tid": 1728719, "ts": 6950381110984.295, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381115336.337, "dur": 8.896 , "args": { "External id": 532,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2159, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2159, "pid": 0, "tid": 7, "ts": 6950381115336.337, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381110992.821, "dur": 1.613 , "args": { "External id": 532,"cbid": 307, "correlation": 2159 } }, { "ph": "s", "id": 2159, "pid": 1728719, "tid": 1728719, "ts": 6950381110992.821, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381115345.521, "dur": 26.463 , "args": { "External id": 537,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2182, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2182, "pid": 0, "tid": 7, "ts": 6950381115345.521, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111010.955, "dur": 1.693 , "args": { "External id": 537,"cbid": 307, "correlation": 2182 } }, { "ph": "s", "id": 2182, "pid": 1728719, "tid": 1728719, "ts": 6950381111010.955, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111021.274, "dur": 0.090 , "args": { "External id": 538,"cbid": 200, "correlation": 2197 } }, { "ph": "f", "id": 2197, "pid": 1728719, "tid": 1728719, "ts": 6950381111021.274, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111021.444, "dur": 0.150 , "args": { "External id": 538,"cbid": 468, "correlation": 2198 } }, { "ph": "f", "id": 2198, "pid": 1728719, "tid": 1728719, "ts": 6950381111021.444, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111021.795, "dur": 0.030 , "args": { "External id": 538,"cbid": 1, "correlation": 2200 } }, { "ph": "f", "id": 2200, "pid": 1728719, "tid": 1728719, "ts": 6950381111021.795, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111021.885, "dur": 0.170 , "args": { "External id": 538,"cbid": 468, "correlation": 2201 } }, { "ph": "f", "id": 2201, "pid": 1728719, "tid": 1728719, "ts": 6950381111021.885, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111022.205, "dur": 0.021 , "args": { "External id": 538,"cbid": 1, "correlation": 2203 } }, { "ph": "f", "id": 2203, "pid": 1728719, "tid": 1728719, "ts": 6950381111022.205, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111022.316, "dur": 0.090 , "args": { "External id": 538,"cbid": 468, "correlation": 2204 } }, { "ph": "f", "id": 2204, "pid": 1728719, "tid": 1728719, "ts": 6950381111022.316, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111022.556, "dur": 0.020 , "args": { "External id": 538,"cbid": 1, "correlation": 2206 } }, { "ph": "f", "id": 2206, "pid": 1728719, "tid": 1728719, "ts": 6950381111022.556, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111022.676, "dur": 0.371 , "args": { "External id": 538,"cbid": 273, "correlation": 2207 } }, { "ph": "f", "id": 2207, "pid": 1728719, "tid": 1728719, "ts": 6950381111022.676, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381115372.272, "dur": 235.646 , "args": { "External id": 538,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2210, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2210, "pid": 0, "tid": 7, "ts": 6950381115372.272, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111023.408, "dur": 1.693 , "args": { "External id": 538,"cbid": 211, "correlation": 2210 } }, { "ph": "s", "id": 2210, "pid": 1728719, "tid": 1728719, "ts": 6950381111023.408, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381115608.206, "dur": 60.703 , "args": { "External id": 550,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2224, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2224, "pid": 0, "tid": 7, "ts": 6950381115608.206, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111037.233, "dur": 1.644 , "args": { "External id": 550,"cbid": 307, "correlation": 2224 } }, { "ph": "s", "id": 2224, "pid": 1728719, "tid": 1728719, "ts": 6950381111037.233, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381115669.197, "dur": 63.135 , "args": { "External id": 551,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2234, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2234, "pid": 0, "tid": 7, "ts": 6950381115669.197, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111053.213, "dur": 1.824 , "args": { "External id": 551,"cbid": 307, "correlation": 2234 } }, { "ph": "s", "id": 2234, "pid": 1728719, "tid": 1728719, "ts": 6950381111053.213, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381115732.620, "dur": 39.776 , "args": { "External id": 552,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2243, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2243, "pid": 0, "tid": 7, "ts": 6950381115732.620, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111068.742, "dur": 1.954 , "args": { "External id": 552,"cbid": 307, "correlation": 2243 } }, { "ph": "s", "id": 2243, "pid": 1728719, "tid": 1728719, "ts": 6950381111068.742, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111087.707, "dur": 0.161 , "args": { "External id": 560,"cbid": 200, "correlation": 2278 } }, { "ph": "f", "id": 2278, "pid": 1728719, "tid": 1728719, "ts": 6950381111087.707, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111087.948, "dur": 0.080 , "args": { "External id": 560,"cbid": 200, "correlation": 2279 } }, { "ph": "f", "id": 2279, "pid": 1728719, "tid": 1728719, "ts": 6950381111087.948, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111088.128, "dur": 0.511 , "args": { "External id": 560,"cbid": 273, "correlation": 2280 } }, { "ph": "f", "id": 2280, "pid": 1728719, "tid": 1728719, "ts": 6950381111088.128, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381115772.652, "dur": 851.831 , "args": { "External id": 560,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2282, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2282, "pid": 0, "tid": 7, "ts": 6950381115772.652, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111089.020, "dur": 1.984 , "args": { "External id": 560,"cbid": 211, "correlation": 2282 } }, { "ph": "s", "id": 2282, "pid": 1728719, "tid": 1728719, "ts": 6950381111089.020, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381116624.739, "dur": 9.440 , "args": { "External id": 569,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2302, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2302, "pid": 0, "tid": 7, "ts": 6950381116624.739, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111108.396, "dur": 1.633 , "args": { "External id": 569,"cbid": 652, "correlation": 2302 } }, { "ph": "s", "id": 2302, "pid": 1728719, "tid": 1728719, "ts": 6950381111108.396, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111119.677, "dur": 0.120 , "args": { "External id": 570,"cbid": 200, "correlation": 2317 } }, { "ph": "f", "id": 2317, "pid": 1728719, "tid": 1728719, "ts": 6950381111119.677, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111119.897, "dur": 0.151 , "args": { "External id": 570,"cbid": 468, "correlation": 2318 } }, { "ph": "f", "id": 2318, "pid": 1728719, "tid": 1728719, "ts": 6950381111119.897, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111120.298, "dur": 0.030 , "args": { "External id": 570,"cbid": 1, "correlation": 2320 } }, { "ph": "f", "id": 2320, "pid": 1728719, "tid": 1728719, "ts": 6950381111120.298, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111120.398, "dur": 0.181 , "args": { "External id": 570,"cbid": 468, "correlation": 2321 } }, { "ph": "f", "id": 2321, "pid": 1728719, "tid": 1728719, "ts": 6950381111120.398, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111120.729, "dur": 0.020 , "args": { "External id": 570,"cbid": 1, "correlation": 2323 } }, { "ph": "f", "id": 2323, "pid": 1728719, "tid": 1728719, "ts": 6950381111120.729, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111120.849, "dur": 0.090 , "args": { "External id": 570,"cbid": 468, "correlation": 2324 } }, { "ph": "f", "id": 2324, "pid": 1728719, "tid": 1728719, "ts": 6950381111120.849, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111121.110, "dur": 0.020 , "args": { "External id": 570,"cbid": 1, "correlation": 2326 } }, { "ph": "f", "id": 2326, "pid": 1728719, "tid": 1728719, "ts": 6950381111121.110, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111121.230, "dur": 0.411 , "args": { "External id": 570,"cbid": 273, "correlation": 2327 } }, { "ph": "f", "id": 2327, "pid": 1728719, "tid": 1728719, "ts": 6950381111121.230, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381116634.435, "dur": 211.581 , "args": { "External id": 570,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2330, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2330, "pid": 0, "tid": 7, "ts": 6950381116634.435, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111122.021, "dur": 1.764 , "args": { "External id": 570,"cbid": 211, "correlation": 2330 } }, { "ph": "s", "id": 2330, "pid": 1728719, "tid": 1728719, "ts": 6950381111122.021, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381116846.304, "dur": 46.752 , "args": { "External id": 582,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2342, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2342, "pid": 0, "tid": 7, "ts": 6950381116846.304, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111137.129, "dur": 2.816 , "args": { "External id": 582,"cbid": 307, "correlation": 2342 } }, { "ph": "s", "id": 2342, "pid": 1728719, "tid": 1728719, "ts": 6950381111137.129, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381116893.344, "dur": 13.472 , "args": { "External id": 583,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2348, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2348, "pid": 0, "tid": 7, "ts": 6950381116893.344, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111147.268, "dur": 1.563 , "args": { "External id": 583,"cbid": 652, "correlation": 2348 } }, { "ph": "s", "id": 2348, "pid": 1728719, "tid": 1728719, "ts": 6950381111147.268, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381116907.136, "dur": 13.056 , "args": { "External id": 584,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2354, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2354, "pid": 0, "tid": 7, "ts": 6950381116907.136, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111156.796, "dur": 1.613 , "args": { "External id": 584,"cbid": 652, "correlation": 2354 } }, { "ph": "s", "id": 2354, "pid": 1728719, "tid": 1728719, "ts": 6950381111156.796, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111169.309, "dur": 0.100 , "args": { "External id": 585,"cbid": 200, "correlation": 2373 } }, { "ph": "f", "id": 2373, "pid": 1728719, "tid": 1728719, "ts": 6950381111169.309, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111169.520, "dur": 0.130 , "args": { "External id": 585,"cbid": 468, "correlation": 2374 } }, { "ph": "f", "id": 2374, "pid": 1728719, "tid": 1728719, "ts": 6950381111169.520, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111169.910, "dur": 0.020 , "args": { "External id": 585,"cbid": 1, "correlation": 2376 } }, { "ph": "f", "id": 2376, "pid": 1728719, "tid": 1728719, "ts": 6950381111169.910, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111170.081, "dur": 0.100 , "args": { "External id": 585,"cbid": 468, "correlation": 2377 } }, { "ph": "f", "id": 2377, "pid": 1728719, "tid": 1728719, "ts": 6950381111170.081, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111170.331, "dur": 0.020 , "args": { "External id": 585,"cbid": 1, "correlation": 2379 } }, { "ph": "f", "id": 2379, "pid": 1728719, "tid": 1728719, "ts": 6950381111170.331, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111170.441, "dur": 0.101 , "args": { "External id": 585,"cbid": 468, "correlation": 2380 } }, { "ph": "f", "id": 2380, "pid": 1728719, "tid": 1728719, "ts": 6950381111170.441, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111170.692, "dur": 0.020 , "args": { "External id": 585,"cbid": 1, "correlation": 2382 } }, { "ph": "f", "id": 2382, "pid": 1728719, "tid": 1728719, "ts": 6950381111170.692, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111170.802, "dur": 0.411 , "args": { "External id": 585,"cbid": 273, "correlation": 2383 } }, { "ph": "f", "id": 2383, "pid": 1728719, "tid": 1728719, "ts": 6950381111170.802, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381116920.512, "dur": 679.064 , "args": { "External id": 585,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2386, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2386, "pid": 0, "tid": 7, "ts": 6950381116920.512, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111171.563, "dur": 1.714 , "args": { "External id": 585,"cbid": 211, "correlation": 2386 } }, { "ph": "s", "id": 2386, "pid": 1728719, "tid": 1728719, "ts": 6950381111171.563, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111183.536, "dur": 0.090 , "args": { "External id": 597,"cbid": 200, "correlation": 2410 } }, { "ph": "f", "id": 2410, "pid": 1728719, "tid": 1728719, "ts": 6950381111183.536, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111183.706, "dur": 0.120 , "args": { "External id": 597,"cbid": 468, "correlation": 2411 } }, { "ph": "f", "id": 2411, "pid": 1728719, "tid": 1728719, "ts": 6950381111183.706, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111184.007, "dur": 0.020 , "args": { "External id": 597,"cbid": 1, "correlation": 2413 } }, { "ph": "f", "id": 2413, "pid": 1728719, "tid": 1728719, "ts": 6950381111184.007, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111184.097, "dur": 0.100 , "args": { "External id": 597,"cbid": 468, "correlation": 2414 } }, { "ph": "f", "id": 2414, "pid": 1728719, "tid": 1728719, "ts": 6950381111184.097, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111184.337, "dur": 0.020 , "args": { "External id": 597,"cbid": 1, "correlation": 2416 } }, { "ph": "f", "id": 2416, "pid": 1728719, "tid": 1728719, "ts": 6950381111184.337, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111184.447, "dur": 0.111 , "args": { "External id": 597,"cbid": 468, "correlation": 2417 } }, { "ph": "f", "id": 2417, "pid": 1728719, "tid": 1728719, "ts": 6950381111184.447, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111184.708, "dur": 0.020 , "args": { "External id": 597,"cbid": 1, "correlation": 2419 } }, { "ph": "f", "id": 2419, "pid": 1728719, "tid": 1728719, "ts": 6950381111184.708, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111184.808, "dur": 0.331 , "args": { "External id": 597,"cbid": 273, "correlation": 2420 } }, { "ph": "f", "id": 2420, "pid": 1728719, "tid": 1728719, "ts": 6950381111184.808, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381117599.864, "dur": 691.481 , "args": { "External id": 597,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2423, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2423, "pid": 0, "tid": 7, "ts": 6950381117599.864, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111185.519, "dur": 1.774 , "args": { "External id": 597,"cbid": 211, "correlation": 2423 } }, { "ph": "s", "id": 2423, "pid": 1728719, "tid": 1728719, "ts": 6950381111185.519, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381118291.633, "dur": 154.718 , "args": { "External id": 609,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2436, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2436, "pid": 0, "tid": 7, "ts": 6950381118291.633, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111198.423, "dur": 1.954 , "args": { "External id": 609,"cbid": 307, "correlation": 2436 } }, { "ph": "s", "id": 2436, "pid": 1728719, "tid": 1728719, "ts": 6950381111198.423, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381118446.639, "dur": 59.071 , "args": { "External id": 610,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2442, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2442, "pid": 0, "tid": 7, "ts": 6950381118446.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111211.478, "dur": 1.613 , "args": { "External id": 610,"cbid": 652, "correlation": 2442 } }, { "ph": "s", "id": 2442, "pid": 1728719, "tid": 1728719, "ts": 6950381111211.478, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111223.089, "dur": 0.091 , "args": { "External id": 611,"cbid": 200, "correlation": 2461 } }, { "ph": "f", "id": 2461, "pid": 1728719, "tid": 1728719, "ts": 6950381111223.089, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111223.260, "dur": 0.130 , "args": { "External id": 611,"cbid": 468, "correlation": 2462 } }, { "ph": "f", "id": 2462, "pid": 1728719, "tid": 1728719, "ts": 6950381111223.260, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111223.570, "dur": 0.030 , "args": { "External id": 611,"cbid": 1, "correlation": 2464 } }, { "ph": "f", "id": 2464, "pid": 1728719, "tid": 1728719, "ts": 6950381111223.570, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111223.660, "dur": 0.101 , "args": { "External id": 611,"cbid": 468, "correlation": 2465 } }, { "ph": "f", "id": 2465, "pid": 1728719, "tid": 1728719, "ts": 6950381111223.660, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111223.911, "dur": 0.020 , "args": { "External id": 611,"cbid": 1, "correlation": 2467 } }, { "ph": "f", "id": 2467, "pid": 1728719, "tid": 1728719, "ts": 6950381111223.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111224.031, "dur": 0.100 , "args": { "External id": 611,"cbid": 468, "correlation": 2468 } }, { "ph": "f", "id": 2468, "pid": 1728719, "tid": 1728719, "ts": 6950381111224.031, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111224.292, "dur": 0.020 , "args": { "External id": 611,"cbid": 1, "correlation": 2470 } }, { "ph": "f", "id": 2470, "pid": 1728719, "tid": 1728719, "ts": 6950381111224.292, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111224.402, "dur": 0.361 , "args": { "External id": 611,"cbid": 273, "correlation": 2471 } }, { "ph": "f", "id": 2471, "pid": 1728719, "tid": 1728719, "ts": 6950381111224.402, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381118505.998, "dur": 689.433 , "args": { "External id": 611,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2474, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2474, "pid": 0, "tid": 7, "ts": 6950381118505.998, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111225.133, "dur": 1.673 , "args": { "External id": 611,"cbid": 211, "correlation": 2474 } }, { "ph": "s", "id": 2474, "pid": 1728719, "tid": 1728719, "ts": 6950381111225.133, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381119195.751, "dur": 40.192 , "args": { "External id": 623,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2485, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2485, "pid": 0, "tid": 7, "ts": 6950381119195.751, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111240.973, "dur": 3.146 , "args": { "External id": 623,"cbid": 307, "correlation": 2485 } }, { "ph": "s", "id": 2485, "pid": 1728719, "tid": 1728719, "ts": 6950381111240.973, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381119236.199, "dur": 21.119 , "args": { "External id": 635,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2497, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2497, "pid": 0, "tid": 7, "ts": 6950381119236.199, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111342.402, "dur": 2.765 , "args": { "External id": 635,"cbid": 307, "correlation": 2497 } }, { "ph": "s", "id": 2497, "pid": 1728719, "tid": 1728719, "ts": 6950381111342.402, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381119257.638, "dur": 12.928 , "args": { "External id": 636,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2511, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2511, "pid": 0, "tid": 7, "ts": 6950381119257.638, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111358.301, "dur": 1.773 , "args": { "External id": 636,"cbid": 652, "correlation": 2511 } }, { "ph": "s", "id": 2511, "pid": 1728719, "tid": 1728719, "ts": 6950381111358.301, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381119270.854, "dur": 12.416 , "args": { "External id": 637,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2525, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2525, "pid": 0, "tid": 7, "ts": 6950381119270.854, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111370.915, "dur": 1.593 , "args": { "External id": 637,"cbid": 652, "correlation": 2525 } }, { "ph": "s", "id": 2525, "pid": 1728719, "tid": 1728719, "ts": 6950381111370.915, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381119283.590, "dur": 10.976 , "args": { "External id": 638,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2539, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2539, "pid": 0, "tid": 7, "ts": 6950381119283.590, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111382.025, "dur": 1.653 , "args": { "External id": 638,"cbid": 652, "correlation": 2539 } }, { "ph": "s", "id": 2539, "pid": 1728719, "tid": 1728719, "ts": 6950381111382.025, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111394.028, "dur": 0.150 , "args": { "External id": 639,"cbid": 200, "correlation": 2554 } }, { "ph": "f", "id": 2554, "pid": 1728719, "tid": 1728719, "ts": 6950381111394.028, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111394.348, "dur": 0.181 , "args": { "External id": 639,"cbid": 468, "correlation": 2555 } }, { "ph": "f", "id": 2555, "pid": 1728719, "tid": 1728719, "ts": 6950381111394.348, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111394.889, "dur": 0.030 , "args": { "External id": 639,"cbid": 1, "correlation": 2557 } }, { "ph": "f", "id": 2557, "pid": 1728719, "tid": 1728719, "ts": 6950381111394.889, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111395.009, "dur": 0.101 , "args": { "External id": 639,"cbid": 468, "correlation": 2558 } }, { "ph": "f", "id": 2558, "pid": 1728719, "tid": 1728719, "ts": 6950381111395.009, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111395.260, "dur": 0.020 , "args": { "External id": 639,"cbid": 1, "correlation": 2560 } }, { "ph": "f", "id": 2560, "pid": 1728719, "tid": 1728719, "ts": 6950381111395.260, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111395.370, "dur": 0.100 , "args": { "External id": 639,"cbid": 468, "correlation": 2561 } }, { "ph": "f", "id": 2561, "pid": 1728719, "tid": 1728719, "ts": 6950381111395.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111395.631, "dur": 0.020 , "args": { "External id": 639,"cbid": 1, "correlation": 2563 } }, { "ph": "f", "id": 2563, "pid": 1728719, "tid": 1728719, "ts": 6950381111395.631, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111395.771, "dur": 0.661 , "args": { "External id": 639,"cbid": 273, "correlation": 2564 } }, { "ph": "f", "id": 2564, "pid": 1728719, "tid": 1728719, "ts": 6950381111395.771, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381119294.854, "dur": 220.510 , "args": { "External id": 639,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2567, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2567, "pid": 0, "tid": 7, "ts": 6950381119294.854, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111396.883, "dur": 2.024 , "args": { "External id": 639,"cbid": 211, "correlation": 2567 } }, { "ph": "s", "id": 2567, "pid": 1728719, "tid": 1728719, "ts": 6950381111396.883, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381119515.652, "dur": 8.863 , "args": { "External id": 651,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2580, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2580, "pid": 0, "tid": 7, "ts": 6950381119515.652, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111408.715, "dur": 1.743 , "args": { "External id": 651,"cbid": 307, "correlation": 2580 } }, { "ph": "s", "id": 2580, "pid": 1728719, "tid": 1728719, "ts": 6950381111408.715, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381119524.803, "dur": 36.960 , "args": { "External id": 656,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2603, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2603, "pid": 0, "tid": 7, "ts": 6950381119524.803, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111430.055, "dur": 1.713 , "args": { "External id": 656,"cbid": 307, "correlation": 2603 } }, { "ph": "s", "id": 2603, "pid": 1728719, "tid": 1728719, "ts": 6950381111430.055, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111440.414, "dur": 0.100 , "args": { "External id": 657,"cbid": 200, "correlation": 2618 } }, { "ph": "f", "id": 2618, "pid": 1728719, "tid": 1728719, "ts": 6950381111440.414, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111440.614, "dur": 0.141 , "args": { "External id": 657,"cbid": 468, "correlation": 2619 } }, { "ph": "f", "id": 2619, "pid": 1728719, "tid": 1728719, "ts": 6950381111440.614, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111440.995, "dur": 0.030 , "args": { "External id": 657,"cbid": 1, "correlation": 2621 } }, { "ph": "f", "id": 2621, "pid": 1728719, "tid": 1728719, "ts": 6950381111440.995, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111441.095, "dur": 0.100 , "args": { "External id": 657,"cbid": 468, "correlation": 2622 } }, { "ph": "f", "id": 2622, "pid": 1728719, "tid": 1728719, "ts": 6950381111441.095, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111441.346, "dur": 0.020 , "args": { "External id": 657,"cbid": 1, "correlation": 2624 } }, { "ph": "f", "id": 2624, "pid": 1728719, "tid": 1728719, "ts": 6950381111441.346, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111441.466, "dur": 0.100 , "args": { "External id": 657,"cbid": 468, "correlation": 2625 } }, { "ph": "f", "id": 2625, "pid": 1728719, "tid": 1728719, "ts": 6950381111441.466, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111441.716, "dur": 0.020 , "args": { "External id": 657,"cbid": 1, "correlation": 2627 } }, { "ph": "f", "id": 2627, "pid": 1728719, "tid": 1728719, "ts": 6950381111441.716, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111441.837, "dur": 0.430 , "args": { "External id": 657,"cbid": 273, "correlation": 2628 } }, { "ph": "f", "id": 2628, "pid": 1728719, "tid": 1728719, "ts": 6950381111441.837, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381119562.051, "dur": 235.101 , "args": { "External id": 657,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2631, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2631, "pid": 0, "tid": 7, "ts": 6950381119562.051, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111442.668, "dur": 1.743 , "args": { "External id": 657,"cbid": 211, "correlation": 2631 } }, { "ph": "s", "id": 2631, "pid": 1728719, "tid": 1728719, "ts": 6950381111442.668, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381119797.473, "dur": 8.831 , "args": { "External id": 669,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2640, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2640, "pid": 0, "tid": 7, "ts": 6950381119797.473, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111450.593, "dur": 1.583 , "args": { "External id": 669,"cbid": 307, "correlation": 2640 } }, { "ph": "s", "id": 2640, "pid": 1728719, "tid": 1728719, "ts": 6950381111450.593, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381119806.592, "dur": 27.264 , "args": { "External id": 674,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2663, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2663, "pid": 0, "tid": 7, "ts": 6950381119806.592, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111467.795, "dur": 1.633 , "args": { "External id": 674,"cbid": 307, "correlation": 2663 } }, { "ph": "s", "id": 2663, "pid": 1728719, "tid": 1728719, "ts": 6950381111467.795, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111478.084, "dur": 0.090 , "args": { "External id": 675,"cbid": 200, "correlation": 2678 } }, { "ph": "f", "id": 2678, "pid": 1728719, "tid": 1728719, "ts": 6950381111478.084, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111478.284, "dur": 0.131 , "args": { "External id": 675,"cbid": 468, "correlation": 2679 } }, { "ph": "f", "id": 2679, "pid": 1728719, "tid": 1728719, "ts": 6950381111478.284, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111478.605, "dur": 0.040 , "args": { "External id": 675,"cbid": 1, "correlation": 2681 } }, { "ph": "f", "id": 2681, "pid": 1728719, "tid": 1728719, "ts": 6950381111478.605, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111478.715, "dur": 0.090 , "args": { "External id": 675,"cbid": 468, "correlation": 2682 } }, { "ph": "f", "id": 2682, "pid": 1728719, "tid": 1728719, "ts": 6950381111478.715, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111478.946, "dur": 0.030 , "args": { "External id": 675,"cbid": 1, "correlation": 2684 } }, { "ph": "f", "id": 2684, "pid": 1728719, "tid": 1728719, "ts": 6950381111478.946, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111479.066, "dur": 0.090 , "args": { "External id": 675,"cbid": 468, "correlation": 2685 } }, { "ph": "f", "id": 2685, "pid": 1728719, "tid": 1728719, "ts": 6950381111479.066, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111479.306, "dur": 0.020 , "args": { "External id": 675,"cbid": 1, "correlation": 2687 } }, { "ph": "f", "id": 2687, "pid": 1728719, "tid": 1728719, "ts": 6950381111479.306, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111479.407, "dur": 0.380 , "args": { "External id": 675,"cbid": 273, "correlation": 2688 } }, { "ph": "f", "id": 2688, "pid": 1728719, "tid": 1728719, "ts": 6950381111479.407, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381119834.144, "dur": 236.798 , "args": { "External id": 675,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2691, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2691, "pid": 0, "tid": 7, "ts": 6950381119834.144, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111480.148, "dur": 1.743 , "args": { "External id": 675,"cbid": 211, "correlation": 2691 } }, { "ph": "s", "id": 2691, "pid": 1728719, "tid": 1728719, "ts": 6950381111480.148, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381120071.230, "dur": 58.367 , "args": { "External id": 687,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2705, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2705, "pid": 0, "tid": 7, "ts": 6950381120071.230, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111489.936, "dur": 1.593 , "args": { "External id": 687,"cbid": 307, "correlation": 2705 } }, { "ph": "s", "id": 2705, "pid": 1728719, "tid": 1728719, "ts": 6950381111489.936, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381120129.885, "dur": 66.079 , "args": { "External id": 688,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2715, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2715, "pid": 0, "tid": 7, "ts": 6950381120129.885, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111500.856, "dur": 1.834 , "args": { "External id": 688,"cbid": 307, "correlation": 2715 } }, { "ph": "s", "id": 2715, "pid": 1728719, "tid": 1728719, "ts": 6950381111500.856, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381120196.220, "dur": 37.856 , "args": { "External id": 689,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2724, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2724, "pid": 0, "tid": 7, "ts": 6950381120196.220, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111511.727, "dur": 1.663 , "args": { "External id": 689,"cbid": 307, "correlation": 2724 } }, { "ph": "s", "id": 2724, "pid": 1728719, "tid": 1728719, "ts": 6950381111511.727, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111529.390, "dur": 0.120 , "args": { "External id": 697,"cbid": 200, "correlation": 2759 } }, { "ph": "f", "id": 2759, "pid": 1728719, "tid": 1728719, "ts": 6950381111529.390, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111529.600, "dur": 0.070 , "args": { "External id": 697,"cbid": 200, "correlation": 2760 } }, { "ph": "f", "id": 2760, "pid": 1728719, "tid": 1728719, "ts": 6950381111529.600, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111529.770, "dur": 0.571 , "args": { "External id": 697,"cbid": 273, "correlation": 2761 } }, { "ph": "f", "id": 2761, "pid": 1728719, "tid": 1728719, "ts": 6950381111529.770, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381120234.332, "dur": 850.839 , "args": { "External id": 697,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2763, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2763, "pid": 0, "tid": 7, "ts": 6950381120234.332, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111530.672, "dur": 1.903 , "args": { "External id": 697,"cbid": 211, "correlation": 2763 } }, { "ph": "s", "id": 2763, "pid": 1728719, "tid": 1728719, "ts": 6950381111530.672, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381121085.459, "dur": 9.664 , "args": { "External id": 706,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2783, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2783, "pid": 0, "tid": 7, "ts": 6950381121085.459, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111552.623, "dur": 1.643 , "args": { "External id": 706,"cbid": 652, "correlation": 2783 } }, { "ph": "s", "id": 2783, "pid": 1728719, "tid": 1728719, "ts": 6950381111552.623, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111563.974, "dur": 0.140 , "args": { "External id": 707,"cbid": 200, "correlation": 2798 } }, { "ph": "f", "id": 2798, "pid": 1728719, "tid": 1728719, "ts": 6950381111563.974, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111564.224, "dur": 0.241 , "args": { "External id": 707,"cbid": 468, "correlation": 2799 } }, { "ph": "f", "id": 2799, "pid": 1728719, "tid": 1728719, "ts": 6950381111564.224, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111564.735, "dur": 0.040 , "args": { "External id": 707,"cbid": 1, "correlation": 2801 } }, { "ph": "f", "id": 2801, "pid": 1728719, "tid": 1728719, "ts": 6950381111564.735, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111564.846, "dur": 0.100 , "args": { "External id": 707,"cbid": 468, "correlation": 2802 } }, { "ph": "f", "id": 2802, "pid": 1728719, "tid": 1728719, "ts": 6950381111564.846, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111565.086, "dur": 0.030 , "args": { "External id": 707,"cbid": 1, "correlation": 2804 } }, { "ph": "f", "id": 2804, "pid": 1728719, "tid": 1728719, "ts": 6950381111565.086, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111565.206, "dur": 0.100 , "args": { "External id": 707,"cbid": 468, "correlation": 2805 } }, { "ph": "f", "id": 2805, "pid": 1728719, "tid": 1728719, "ts": 6950381111565.206, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111565.477, "dur": 0.040 , "args": { "External id": 707,"cbid": 1, "correlation": 2807 } }, { "ph": "f", "id": 2807, "pid": 1728719, "tid": 1728719, "ts": 6950381111565.477, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111565.597, "dur": 0.441 , "args": { "External id": 707,"cbid": 273, "correlation": 2808 } }, { "ph": "f", "id": 2808, "pid": 1728719, "tid": 1728719, "ts": 6950381111565.597, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381121095.379, "dur": 212.477 , "args": { "External id": 707,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2811, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2811, "pid": 0, "tid": 7, "ts": 6950381121095.379, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111566.449, "dur": 1.753 , "args": { "External id": 707,"cbid": 211, "correlation": 2811 } }, { "ph": "s", "id": 2811, "pid": 1728719, "tid": 1728719, "ts": 6950381111566.449, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381121308.144, "dur": 44.672 , "args": { "External id": 719,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2823, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2823, "pid": 0, "tid": 7, "ts": 6950381121308.144, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111577.359, "dur": 1.713 , "args": { "External id": 719,"cbid": 307, "correlation": 2823 } }, { "ph": "s", "id": 2823, "pid": 1728719, "tid": 1728719, "ts": 6950381111577.359, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381121353.136, "dur": 13.344 , "args": { "External id": 720,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2829, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2829, "pid": 0, "tid": 7, "ts": 6950381121353.136, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111585.905, "dur": 1.523 , "args": { "External id": 720,"cbid": 652, "correlation": 2829 } }, { "ph": "s", "id": 2829, "pid": 1728719, "tid": 1728719, "ts": 6950381111585.905, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381121366.800, "dur": 13.087 , "args": { "External id": 721,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2835, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2835, "pid": 0, "tid": 7, "ts": 6950381121366.800, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111595.062, "dur": 1.483 , "args": { "External id": 721,"cbid": 652, "correlation": 2835 } }, { "ph": "s", "id": 2835, "pid": 1728719, "tid": 1728719, "ts": 6950381111595.062, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111606.894, "dur": 0.100 , "args": { "External id": 722,"cbid": 200, "correlation": 2854 } }, { "ph": "f", "id": 2854, "pid": 1728719, "tid": 1728719, "ts": 6950381111606.894, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111607.114, "dur": 0.120 , "args": { "External id": 722,"cbid": 468, "correlation": 2855 } }, { "ph": "f", "id": 2855, "pid": 1728719, "tid": 1728719, "ts": 6950381111607.114, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111607.425, "dur": 0.030 , "args": { "External id": 722,"cbid": 1, "correlation": 2857 } }, { "ph": "f", "id": 2857, "pid": 1728719, "tid": 1728719, "ts": 6950381111607.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111607.595, "dur": 0.100 , "args": { "External id": 722,"cbid": 468, "correlation": 2858 } }, { "ph": "f", "id": 2858, "pid": 1728719, "tid": 1728719, "ts": 6950381111607.595, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111607.846, "dur": 0.020 , "args": { "External id": 722,"cbid": 1, "correlation": 2860 } }, { "ph": "f", "id": 2860, "pid": 1728719, "tid": 1728719, "ts": 6950381111607.846, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111607.956, "dur": 0.100 , "args": { "External id": 722,"cbid": 468, "correlation": 2861 } }, { "ph": "f", "id": 2861, "pid": 1728719, "tid": 1728719, "ts": 6950381111607.956, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111608.206, "dur": 0.020 , "args": { "External id": 722,"cbid": 1, "correlation": 2863 } }, { "ph": "f", "id": 2863, "pid": 1728719, "tid": 1728719, "ts": 6950381111608.206, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111608.316, "dur": 0.381 , "args": { "External id": 722,"cbid": 273, "correlation": 2864 } }, { "ph": "f", "id": 2864, "pid": 1728719, "tid": 1728719, "ts": 6950381111608.316, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381121380.207, "dur": 675.033 , "args": { "External id": 722,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2867, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2867, "pid": 0, "tid": 7, "ts": 6950381121380.207, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111609.038, "dur": 1.753 , "args": { "External id": 722,"cbid": 211, "correlation": 2867 } }, { "ph": "s", "id": 2867, "pid": 1728719, "tid": 1728719, "ts": 6950381111609.038, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111621.241, "dur": 0.140 , "args": { "External id": 734,"cbid": 200, "correlation": 2891 } }, { "ph": "f", "id": 2891, "pid": 1728719, "tid": 1728719, "ts": 6950381111621.241, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111621.481, "dur": 0.130 , "args": { "External id": 734,"cbid": 468, "correlation": 2892 } }, { "ph": "f", "id": 2892, "pid": 1728719, "tid": 1728719, "ts": 6950381111621.481, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111621.812, "dur": 0.020 , "args": { "External id": 734,"cbid": 1, "correlation": 2894 } }, { "ph": "f", "id": 2894, "pid": 1728719, "tid": 1728719, "ts": 6950381111621.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111621.902, "dur": 0.100 , "args": { "External id": 734,"cbid": 468, "correlation": 2895 } }, { "ph": "f", "id": 2895, "pid": 1728719, "tid": 1728719, "ts": 6950381111621.902, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111622.152, "dur": 0.020 , "args": { "External id": 734,"cbid": 1, "correlation": 2897 } }, { "ph": "f", "id": 2897, "pid": 1728719, "tid": 1728719, "ts": 6950381111622.152, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111622.262, "dur": 0.091 , "args": { "External id": 734,"cbid": 468, "correlation": 2898 } }, { "ph": "f", "id": 2898, "pid": 1728719, "tid": 1728719, "ts": 6950381111622.262, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111622.503, "dur": 0.020 , "args": { "External id": 734,"cbid": 1, "correlation": 2900 } }, { "ph": "f", "id": 2900, "pid": 1728719, "tid": 1728719, "ts": 6950381111622.503, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111622.603, "dur": 0.331 , "args": { "External id": 734,"cbid": 273, "correlation": 2901 } }, { "ph": "f", "id": 2901, "pid": 1728719, "tid": 1728719, "ts": 6950381111622.603, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381122055.560, "dur": 697.784 , "args": { "External id": 734,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2904, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2904, "pid": 0, "tid": 7, "ts": 6950381122055.560, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111623.274, "dur": 1.794 , "args": { "External id": 734,"cbid": 211, "correlation": 2904 } }, { "ph": "s", "id": 2904, "pid": 1728719, "tid": 1728719, "ts": 6950381111623.274, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381122753.664, "dur": 154.271 , "args": { "External id": 746,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2917, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2917, "pid": 0, "tid": 7, "ts": 6950381122753.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111633.203, "dur": 1.623 , "args": { "External id": 746,"cbid": 307, "correlation": 2917 } }, { "ph": "s", "id": 2917, "pid": 1728719, "tid": 1728719, "ts": 6950381111633.203, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381122908.191, "dur": 58.271 , "args": { "External id": 747,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2923, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2923, "pid": 0, "tid": 7, "ts": 6950381122908.191, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111641.929, "dur": 1.884 , "args": { "External id": 747,"cbid": 652, "correlation": 2923 } }, { "ph": "s", "id": 2923, "pid": 1728719, "tid": 1728719, "ts": 6950381111641.929, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111653.581, "dur": 0.090 , "args": { "External id": 748,"cbid": 200, "correlation": 2942 } }, { "ph": "f", "id": 2942, "pid": 1728719, "tid": 1728719, "ts": 6950381111653.581, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111653.771, "dur": 0.120 , "args": { "External id": 748,"cbid": 468, "correlation": 2943 } }, { "ph": "f", "id": 2943, "pid": 1728719, "tid": 1728719, "ts": 6950381111653.771, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111654.112, "dur": 0.020 , "args": { "External id": 748,"cbid": 1, "correlation": 2945 } }, { "ph": "f", "id": 2945, "pid": 1728719, "tid": 1728719, "ts": 6950381111654.112, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111654.202, "dur": 0.100 , "args": { "External id": 748,"cbid": 468, "correlation": 2946 } }, { "ph": "f", "id": 2946, "pid": 1728719, "tid": 1728719, "ts": 6950381111654.202, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111654.462, "dur": 0.020 , "args": { "External id": 748,"cbid": 1, "correlation": 2948 } }, { "ph": "f", "id": 2948, "pid": 1728719, "tid": 1728719, "ts": 6950381111654.462, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111654.573, "dur": 0.100 , "args": { "External id": 748,"cbid": 468, "correlation": 2949 } }, { "ph": "f", "id": 2949, "pid": 1728719, "tid": 1728719, "ts": 6950381111654.573, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111654.853, "dur": 0.020 , "args": { "External id": 748,"cbid": 1, "correlation": 2951 } }, { "ph": "f", "id": 2951, "pid": 1728719, "tid": 1728719, "ts": 6950381111654.853, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111654.953, "dur": 0.371 , "args": { "External id": 748,"cbid": 273, "correlation": 2952 } }, { "ph": "f", "id": 2952, "pid": 1728719, "tid": 1728719, "ts": 6950381111654.953, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381122966.750, "dur": 692.537 , "args": { "External id": 748,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2955, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2955, "pid": 0, "tid": 7, "ts": 6950381122966.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111655.695, "dur": 1.703 , "args": { "External id": 748,"cbid": 211, "correlation": 2955 } }, { "ph": "s", "id": 2955, "pid": 1728719, "tid": 1728719, "ts": 6950381111655.695, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381123659.575, "dur": 40.351 , "args": { "External id": 760,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2966, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2966, "pid": 0, "tid": 7, "ts": 6950381123659.575, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111666.605, "dur": 1.693 , "args": { "External id": 760,"cbid": 307, "correlation": 2966 } }, { "ph": "s", "id": 2966, "pid": 1728719, "tid": 1728719, "ts": 6950381111666.605, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381123700.182, "dur": 21.504 , "args": { "External id": 772,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2978, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2978, "pid": 0, "tid": 7, "ts": 6950381123700.182, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111756.031, "dur": 2.495 , "args": { "External id": 772,"cbid": 307, "correlation": 2978 } }, { "ph": "s", "id": 2978, "pid": 1728719, "tid": 1728719, "ts": 6950381111756.031, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381123722.006, "dur": 12.544 , "args": { "External id": 773,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2992, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 2992, "pid": 0, "tid": 7, "ts": 6950381123722.006, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111770.959, "dur": 1.703 , "args": { "External id": 773,"cbid": 652, "correlation": 2992 } }, { "ph": "s", "id": 2992, "pid": 1728719, "tid": 1728719, "ts": 6950381111770.959, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381123734.838, "dur": 12.224 , "args": { "External id": 774,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3006, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3006, "pid": 0, "tid": 7, "ts": 6950381123734.838, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111783.042, "dur": 1.573 , "args": { "External id": 774,"cbid": 652, "correlation": 3006 } }, { "ph": "s", "id": 3006, "pid": 1728719, "tid": 1728719, "ts": 6950381111783.042, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381123747.350, "dur": 10.912 , "args": { "External id": 775,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3020, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3020, "pid": 0, "tid": 7, "ts": 6950381123747.350, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111794.383, "dur": 1.513 , "args": { "External id": 775,"cbid": 652, "correlation": 3020 } }, { "ph": "s", "id": 3020, "pid": 1728719, "tid": 1728719, "ts": 6950381111794.383, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111806.014, "dur": 0.181 , "args": { "External id": 776,"cbid": 200, "correlation": 3035 } }, { "ph": "f", "id": 3035, "pid": 1728719, "tid": 1728719, "ts": 6950381111806.014, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111806.355, "dur": 0.241 , "args": { "External id": 776,"cbid": 468, "correlation": 3036 } }, { "ph": "f", "id": 3036, "pid": 1728719, "tid": 1728719, "ts": 6950381111806.355, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111806.926, "dur": 0.030 , "args": { "External id": 776,"cbid": 1, "correlation": 3038 } }, { "ph": "f", "id": 3038, "pid": 1728719, "tid": 1728719, "ts": 6950381111806.926, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111807.026, "dur": 0.101 , "args": { "External id": 776,"cbid": 468, "correlation": 3039 } }, { "ph": "f", "id": 3039, "pid": 1728719, "tid": 1728719, "ts": 6950381111807.026, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111807.277, "dur": 0.020 , "args": { "External id": 776,"cbid": 1, "correlation": 3041 } }, { "ph": "f", "id": 3041, "pid": 1728719, "tid": 1728719, "ts": 6950381111807.277, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111807.407, "dur": 0.090 , "args": { "External id": 776,"cbid": 468, "correlation": 3042 } }, { "ph": "f", "id": 3042, "pid": 1728719, "tid": 1728719, "ts": 6950381111807.407, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111807.668, "dur": 0.020 , "args": { "External id": 776,"cbid": 1, "correlation": 3044 } }, { "ph": "f", "id": 3044, "pid": 1728719, "tid": 1728719, "ts": 6950381111807.668, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111807.818, "dur": 0.601 , "args": { "External id": 776,"cbid": 273, "correlation": 3045 } }, { "ph": "f", "id": 3045, "pid": 1728719, "tid": 1728719, "ts": 6950381111807.818, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381123758.550, "dur": 216.157 , "args": { "External id": 776,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3048, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3048, "pid": 0, "tid": 7, "ts": 6950381123758.550, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111808.850, "dur": 1.943 , "args": { "External id": 776,"cbid": 211, "correlation": 3048 } }, { "ph": "s", "id": 3048, "pid": 1728719, "tid": 1728719, "ts": 6950381111808.850, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381123974.995, "dur": 8.768 , "args": { "External id": 788,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3061, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3061, "pid": 0, "tid": 7, "ts": 6950381123974.995, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111819.980, "dur": 1.734 , "args": { "External id": 788,"cbid": 307, "correlation": 3061 } }, { "ph": "s", "id": 3061, "pid": 1728719, "tid": 1728719, "ts": 6950381111819.980, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381123984.051, "dur": 39.904 , "args": { "External id": 793,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3084, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3084, "pid": 0, "tid": 7, "ts": 6950381123984.051, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111839.857, "dur": 1.734 , "args": { "External id": 793,"cbid": 307, "correlation": 3084 } }, { "ph": "s", "id": 3084, "pid": 1728719, "tid": 1728719, "ts": 6950381111839.857, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111850.167, "dur": 0.110 , "args": { "External id": 794,"cbid": 200, "correlation": 3099 } }, { "ph": "f", "id": 3099, "pid": 1728719, "tid": 1728719, "ts": 6950381111850.167, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111850.367, "dur": 0.140 , "args": { "External id": 794,"cbid": 468, "correlation": 3100 } }, { "ph": "f", "id": 3100, "pid": 1728719, "tid": 1728719, "ts": 6950381111850.367, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111850.738, "dur": 0.030 , "args": { "External id": 794,"cbid": 1, "correlation": 3102 } }, { "ph": "f", "id": 3102, "pid": 1728719, "tid": 1728719, "ts": 6950381111850.738, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111850.838, "dur": 0.090 , "args": { "External id": 794,"cbid": 468, "correlation": 3103 } }, { "ph": "f", "id": 3103, "pid": 1728719, "tid": 1728719, "ts": 6950381111850.838, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111851.078, "dur": 0.020 , "args": { "External id": 794,"cbid": 1, "correlation": 3105 } }, { "ph": "f", "id": 3105, "pid": 1728719, "tid": 1728719, "ts": 6950381111851.078, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111851.189, "dur": 0.100 , "args": { "External id": 794,"cbid": 468, "correlation": 3106 } }, { "ph": "f", "id": 3106, "pid": 1728719, "tid": 1728719, "ts": 6950381111851.189, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111851.449, "dur": 0.020 , "args": { "External id": 794,"cbid": 1, "correlation": 3108 } }, { "ph": "f", "id": 3108, "pid": 1728719, "tid": 1728719, "ts": 6950381111851.449, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111851.569, "dur": 0.401 , "args": { "External id": 794,"cbid": 273, "correlation": 3109 } }, { "ph": "f", "id": 3109, "pid": 1728719, "tid": 1728719, "ts": 6950381111851.569, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381124024.243, "dur": 228.157 , "args": { "External id": 794,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3112, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3112, "pid": 0, "tid": 7, "ts": 6950381124024.243, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111852.351, "dur": 1.833 , "args": { "External id": 794,"cbid": 211, "correlation": 3112 } }, { "ph": "s", "id": 3112, "pid": 1728719, "tid": 1728719, "ts": 6950381111852.351, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381124252.688, "dur": 8.928 , "args": { "External id": 806,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3121, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3121, "pid": 0, "tid": 7, "ts": 6950381124252.688, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111860.165, "dur": 1.613 , "args": { "External id": 806,"cbid": 307, "correlation": 3121 } }, { "ph": "s", "id": 3121, "pid": 1728719, "tid": 1728719, "ts": 6950381111860.165, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381124261.904, "dur": 26.112 , "args": { "External id": 811,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3144, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3144, "pid": 0, "tid": 7, "ts": 6950381124261.904, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111877.768, "dur": 1.583 , "args": { "External id": 811,"cbid": 307, "correlation": 3144 } }, { "ph": "s", "id": 3144, "pid": 1728719, "tid": 1728719, "ts": 6950381111877.768, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111888.037, "dur": 0.090 , "args": { "External id": 812,"cbid": 200, "correlation": 3159 } }, { "ph": "f", "id": 3159, "pid": 1728719, "tid": 1728719, "ts": 6950381111888.037, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111888.228, "dur": 0.130 , "args": { "External id": 812,"cbid": 468, "correlation": 3160 } }, { "ph": "f", "id": 3160, "pid": 1728719, "tid": 1728719, "ts": 6950381111888.228, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111888.568, "dur": 0.030 , "args": { "External id": 812,"cbid": 1, "correlation": 3162 } }, { "ph": "f", "id": 3162, "pid": 1728719, "tid": 1728719, "ts": 6950381111888.568, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111888.658, "dur": 0.101 , "args": { "External id": 812,"cbid": 468, "correlation": 3163 } }, { "ph": "f", "id": 3163, "pid": 1728719, "tid": 1728719, "ts": 6950381111888.658, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111888.919, "dur": 0.020 , "args": { "External id": 812,"cbid": 1, "correlation": 3165 } }, { "ph": "f", "id": 3165, "pid": 1728719, "tid": 1728719, "ts": 6950381111888.919, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111889.029, "dur": 0.100 , "args": { "External id": 812,"cbid": 468, "correlation": 3166 } }, { "ph": "f", "id": 3166, "pid": 1728719, "tid": 1728719, "ts": 6950381111889.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111889.320, "dur": 0.020 , "args": { "External id": 812,"cbid": 1, "correlation": 3168 } }, { "ph": "f", "id": 3168, "pid": 1728719, "tid": 1728719, "ts": 6950381111889.320, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111889.440, "dur": 0.360 , "args": { "External id": 812,"cbid": 273, "correlation": 3169 } }, { "ph": "f", "id": 3169, "pid": 1728719, "tid": 1728719, "ts": 6950381111889.440, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381124288.304, "dur": 231.997 , "args": { "External id": 812,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3172, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3172, "pid": 0, "tid": 7, "ts": 6950381124288.304, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111890.171, "dur": 1.733 , "args": { "External id": 812,"cbid": 211, "correlation": 3172 } }, { "ph": "s", "id": 3172, "pid": 1728719, "tid": 1728719, "ts": 6950381111890.171, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381124520.589, "dur": 60.512 , "args": { "External id": 824,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3186, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3186, "pid": 0, "tid": 7, "ts": 6950381124520.589, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111899.929, "dur": 1.503 , "args": { "External id": 824,"cbid": 307, "correlation": 3186 } }, { "ph": "s", "id": 3186, "pid": 1728719, "tid": 1728719, "ts": 6950381111899.929, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381124581.389, "dur": 63.199 , "args": { "External id": 825,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3196, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3196, "pid": 0, "tid": 7, "ts": 6950381124581.389, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111910.168, "dur": 1.744 , "args": { "External id": 825,"cbid": 307, "correlation": 3196 } }, { "ph": "s", "id": 3196, "pid": 1728719, "tid": 1728719, "ts": 6950381111910.168, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381124644.876, "dur": 39.743 , "args": { "External id": 826,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3205, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3205, "pid": 0, "tid": 7, "ts": 6950381124644.876, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111920.197, "dur": 1.623 , "args": { "External id": 826,"cbid": 307, "correlation": 3205 } }, { "ph": "s", "id": 3205, "pid": 1728719, "tid": 1728719, "ts": 6950381111920.197, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111936.267, "dur": 0.120 , "args": { "External id": 834,"cbid": 200, "correlation": 3240 } }, { "ph": "f", "id": 3240, "pid": 1728719, "tid": 1728719, "ts": 6950381111936.267, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111936.457, "dur": 0.161 , "args": { "External id": 834,"cbid": 200, "correlation": 3241 } }, { "ph": "f", "id": 3241, "pid": 1728719, "tid": 1728719, "ts": 6950381111936.457, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111936.718, "dur": 0.471 , "args": { "External id": 834,"cbid": 273, "correlation": 3242 } }, { "ph": "f", "id": 3242, "pid": 1728719, "tid": 1728719, "ts": 6950381111936.718, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381124684.876, "dur": 837.334 , "args": { "External id": 834,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3244, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3244, "pid": 0, "tid": 7, "ts": 6950381124684.876, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111937.519, "dur": 1.884 , "args": { "External id": 834,"cbid": 211, "correlation": 3244 } }, { "ph": "s", "id": 3244, "pid": 1728719, "tid": 1728719, "ts": 6950381111937.519, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381125522.498, "dur": 9.312 , "args": { "External id": 843,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3264, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3264, "pid": 0, "tid": 7, "ts": 6950381125522.498, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111954.831, "dur": 1.553 , "args": { "External id": 843,"cbid": 652, "correlation": 3264 } }, { "ph": "s", "id": 3264, "pid": 1728719, "tid": 1728719, "ts": 6950381111954.831, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111965.632, "dur": 0.110 , "args": { "External id": 844,"cbid": 200, "correlation": 3279 } }, { "ph": "f", "id": 3279, "pid": 1728719, "tid": 1728719, "ts": 6950381111965.632, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111965.852, "dur": 0.150 , "args": { "External id": 844,"cbid": 468, "correlation": 3280 } }, { "ph": "f", "id": 3280, "pid": 1728719, "tid": 1728719, "ts": 6950381111965.852, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111966.303, "dur": 0.030 , "args": { "External id": 844,"cbid": 1, "correlation": 3282 } }, { "ph": "f", "id": 3282, "pid": 1728719, "tid": 1728719, "ts": 6950381111966.303, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111966.403, "dur": 0.100 , "args": { "External id": 844,"cbid": 468, "correlation": 3283 } }, { "ph": "f", "id": 3283, "pid": 1728719, "tid": 1728719, "ts": 6950381111966.403, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111966.654, "dur": 0.020 , "args": { "External id": 844,"cbid": 1, "correlation": 3285 } }, { "ph": "f", "id": 3285, "pid": 1728719, "tid": 1728719, "ts": 6950381111966.654, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111966.764, "dur": 0.100 , "args": { "External id": 844,"cbid": 468, "correlation": 3286 } }, { "ph": "f", "id": 3286, "pid": 1728719, "tid": 1728719, "ts": 6950381111966.764, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381111967.014, "dur": 0.020 , "args": { "External id": 844,"cbid": 1, "correlation": 3288 } }, { "ph": "f", "id": 3288, "pid": 1728719, "tid": 1728719, "ts": 6950381111967.014, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381111967.124, "dur": 0.431 , "args": { "External id": 844,"cbid": 273, "correlation": 3289 } }, { "ph": "f", "id": 3289, "pid": 1728719, "tid": 1728719, "ts": 6950381111967.124, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381125532.098, "dur": 209.662 , "args": { "External id": 844,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3292, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3292, "pid": 0, "tid": 7, "ts": 6950381125532.098, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111967.956, "dur": 1.733 , "args": { "External id": 844,"cbid": 211, "correlation": 3292 } }, { "ph": "s", "id": 3292, "pid": 1728719, "tid": 1728719, "ts": 6950381111967.956, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381125742.048, "dur": 47.648 , "args": { "External id": 856,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3304, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3304, "pid": 0, "tid": 7, "ts": 6950381125742.048, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381111978.175, "dur": 1.653 , "args": { "External id": 856,"cbid": 307, "correlation": 3304 } }, { "ph": "s", "id": 3304, "pid": 1728719, "tid": 1728719, "ts": 6950381111978.175, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381125789.984, "dur": 13.247 , "args": { "External id": 857,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3310, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3310, "pid": 0, "tid": 7, "ts": 6950381125789.984, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111986.040, "dur": 1.683 , "args": { "External id": 857,"cbid": 652, "correlation": 3310 } }, { "ph": "s", "id": 3310, "pid": 1728719, "tid": 1728719, "ts": 6950381111986.040, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381125803.551, "dur": 12.576 , "args": { "External id": 858,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3316, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3316, "pid": 0, "tid": 7, "ts": 6950381125803.551, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381111994.686, "dur": 1.422 , "args": { "External id": 858,"cbid": 652, "correlation": 3316 } }, { "ph": "s", "id": 3316, "pid": 1728719, "tid": 1728719, "ts": 6950381111994.686, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112006.187, "dur": 0.090 , "args": { "External id": 859,"cbid": 200, "correlation": 3335 } }, { "ph": "f", "id": 3335, "pid": 1728719, "tid": 1728719, "ts": 6950381112006.187, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112006.387, "dur": 0.121 , "args": { "External id": 859,"cbid": 468, "correlation": 3336 } }, { "ph": "f", "id": 3336, "pid": 1728719, "tid": 1728719, "ts": 6950381112006.387, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112006.728, "dur": 0.020 , "args": { "External id": 859,"cbid": 1, "correlation": 3338 } }, { "ph": "f", "id": 3338, "pid": 1728719, "tid": 1728719, "ts": 6950381112006.728, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112006.818, "dur": 0.100 , "args": { "External id": 859,"cbid": 468, "correlation": 3339 } }, { "ph": "f", "id": 3339, "pid": 1728719, "tid": 1728719, "ts": 6950381112006.818, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112007.059, "dur": 0.020 , "args": { "External id": 859,"cbid": 1, "correlation": 3341 } }, { "ph": "f", "id": 3341, "pid": 1728719, "tid": 1728719, "ts": 6950381112007.059, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112007.169, "dur": 0.100 , "args": { "External id": 859,"cbid": 468, "correlation": 3342 } }, { "ph": "f", "id": 3342, "pid": 1728719, "tid": 1728719, "ts": 6950381112007.169, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112007.429, "dur": 0.020 , "args": { "External id": 859,"cbid": 1, "correlation": 3344 } }, { "ph": "f", "id": 3344, "pid": 1728719, "tid": 1728719, "ts": 6950381112007.429, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112007.530, "dur": 0.370 , "args": { "External id": 859,"cbid": 273, "correlation": 3345 } }, { "ph": "f", "id": 3345, "pid": 1728719, "tid": 1728719, "ts": 6950381112007.530, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381125816.415, "dur": 659.961 , "args": { "External id": 859,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3348, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3348, "pid": 0, "tid": 7, "ts": 6950381125816.415, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112008.261, "dur": 1.673 , "args": { "External id": 859,"cbid": 211, "correlation": 3348 } }, { "ph": "s", "id": 3348, "pid": 1728719, "tid": 1728719, "ts": 6950381112008.261, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112020.073, "dur": 0.090 , "args": { "External id": 871,"cbid": 200, "correlation": 3372 } }, { "ph": "f", "id": 3372, "pid": 1728719, "tid": 1728719, "ts": 6950381112020.073, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112020.263, "dur": 0.111 , "args": { "External id": 871,"cbid": 468, "correlation": 3373 } }, { "ph": "f", "id": 3373, "pid": 1728719, "tid": 1728719, "ts": 6950381112020.263, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112020.574, "dur": 0.020 , "args": { "External id": 871,"cbid": 1, "correlation": 3375 } }, { "ph": "f", "id": 3375, "pid": 1728719, "tid": 1728719, "ts": 6950381112020.574, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112020.664, "dur": 0.090 , "args": { "External id": 871,"cbid": 468, "correlation": 3376 } }, { "ph": "f", "id": 3376, "pid": 1728719, "tid": 1728719, "ts": 6950381112020.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112020.904, "dur": 0.021 , "args": { "External id": 871,"cbid": 1, "correlation": 3378 } }, { "ph": "f", "id": 3378, "pid": 1728719, "tid": 1728719, "ts": 6950381112020.904, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112021.015, "dur": 0.090 , "args": { "External id": 871,"cbid": 468, "correlation": 3379 } }, { "ph": "f", "id": 3379, "pid": 1728719, "tid": 1728719, "ts": 6950381112021.015, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112021.265, "dur": 0.020 , "args": { "External id": 871,"cbid": 1, "correlation": 3381 } }, { "ph": "f", "id": 3381, "pid": 1728719, "tid": 1728719, "ts": 6950381112021.265, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112021.365, "dur": 0.331 , "args": { "External id": 871,"cbid": 273, "correlation": 3382 } }, { "ph": "f", "id": 3382, "pid": 1728719, "tid": 1728719, "ts": 6950381112021.365, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381126476.664, "dur": 675.481 , "args": { "External id": 871,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3385, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3385, "pid": 0, "tid": 7, "ts": 6950381126476.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112022.037, "dur": 1.673 , "args": { "External id": 871,"cbid": 211, "correlation": 3385 } }, { "ph": "s", "id": 3385, "pid": 1728719, "tid": 1728719, "ts": 6950381112022.037, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381127152.433, "dur": 152.190 , "args": { "External id": 883,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3398, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3398, "pid": 0, "tid": 7, "ts": 6950381127152.433, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112031.484, "dur": 1.713 , "args": { "External id": 883,"cbid": 307, "correlation": 3398 } }, { "ph": "s", "id": 3398, "pid": 1728719, "tid": 1728719, "ts": 6950381112031.484, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381127304.911, "dur": 59.488 , "args": { "External id": 884,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3404, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3404, "pid": 0, "tid": 7, "ts": 6950381127304.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112040.230, "dur": 1.664 , "args": { "External id": 884,"cbid": 652, "correlation": 3404 } }, { "ph": "s", "id": 3404, "pid": 1728719, "tid": 1728719, "ts": 6950381112040.230, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112051.842, "dur": 0.090 , "args": { "External id": 885,"cbid": 200, "correlation": 3423 } }, { "ph": "f", "id": 3423, "pid": 1728719, "tid": 1728719, "ts": 6950381112051.842, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112052.032, "dur": 0.121 , "args": { "External id": 885,"cbid": 468, "correlation": 3424 } }, { "ph": "f", "id": 3424, "pid": 1728719, "tid": 1728719, "ts": 6950381112052.032, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112052.413, "dur": 0.030 , "args": { "External id": 885,"cbid": 1, "correlation": 3426 } }, { "ph": "f", "id": 3426, "pid": 1728719, "tid": 1728719, "ts": 6950381112052.413, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112052.503, "dur": 0.100 , "args": { "External id": 885,"cbid": 468, "correlation": 3427 } }, { "ph": "f", "id": 3427, "pid": 1728719, "tid": 1728719, "ts": 6950381112052.503, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112052.904, "dur": 0.020 , "args": { "External id": 885,"cbid": 1, "correlation": 3429 } }, { "ph": "f", "id": 3429, "pid": 1728719, "tid": 1728719, "ts": 6950381112052.904, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112053.004, "dur": 0.100 , "args": { "External id": 885,"cbid": 468, "correlation": 3430 } }, { "ph": "f", "id": 3430, "pid": 1728719, "tid": 1728719, "ts": 6950381112053.004, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112053.255, "dur": 0.020 , "args": { "External id": 885,"cbid": 1, "correlation": 3432 } }, { "ph": "f", "id": 3432, "pid": 1728719, "tid": 1728719, "ts": 6950381112053.255, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112053.365, "dur": 0.341 , "args": { "External id": 885,"cbid": 273, "correlation": 3433 } }, { "ph": "f", "id": 3433, "pid": 1728719, "tid": 1728719, "ts": 6950381112053.365, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381127364.687, "dur": 682.904 , "args": { "External id": 885,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3436, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3436, "pid": 0, "tid": 7, "ts": 6950381127364.687, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112054.076, "dur": 1.693 , "args": { "External id": 885,"cbid": 211, "correlation": 3436 } }, { "ph": "s", "id": 3436, "pid": 1728719, "tid": 1728719, "ts": 6950381112054.076, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381128047.847, "dur": 40.224 , "args": { "External id": 897,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3447, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3447, "pid": 0, "tid": 7, "ts": 6950381128047.847, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112064.486, "dur": 1.753 , "args": { "External id": 897,"cbid": 307, "correlation": 3447 } }, { "ph": "s", "id": 3447, "pid": 1728719, "tid": 1728719, "ts": 6950381112064.486, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381128088.359, "dur": 20.607 , "args": { "External id": 909,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3459, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3459, "pid": 0, "tid": 7, "ts": 6950381128088.359, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112151.227, "dur": 2.605 , "args": { "External id": 909,"cbid": 307, "correlation": 3459 } }, { "ph": "s", "id": 3459, "pid": 1728719, "tid": 1728719, "ts": 6950381112151.227, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381128109.254, "dur": 12.449 , "args": { "External id": 910,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3473, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3473, "pid": 0, "tid": 7, "ts": 6950381128109.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112165.934, "dur": 1.864 , "args": { "External id": 910,"cbid": 652, "correlation": 3473 } }, { "ph": "s", "id": 3473, "pid": 1728719, "tid": 1728719, "ts": 6950381112165.934, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381128122.022, "dur": 11.904 , "args": { "External id": 911,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3487, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3487, "pid": 0, "tid": 7, "ts": 6950381128122.022, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112177.837, "dur": 1.452 , "args": { "External id": 911,"cbid": 652, "correlation": 3487 } }, { "ph": "s", "id": 3487, "pid": 1728719, "tid": 1728719, "ts": 6950381112177.837, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381128134.214, "dur": 10.944 , "args": { "External id": 912,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3501, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3501, "pid": 0, "tid": 7, "ts": 6950381128134.214, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112188.256, "dur": 1.483 , "args": { "External id": 912,"cbid": 652, "correlation": 3501 } }, { "ph": "s", "id": 3501, "pid": 1728719, "tid": 1728719, "ts": 6950381112188.256, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112199.797, "dur": 0.161 , "args": { "External id": 913,"cbid": 200, "correlation": 3516 } }, { "ph": "f", "id": 3516, "pid": 1728719, "tid": 1728719, "ts": 6950381112199.797, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112200.138, "dur": 0.200 , "args": { "External id": 913,"cbid": 468, "correlation": 3517 } }, { "ph": "f", "id": 3517, "pid": 1728719, "tid": 1728719, "ts": 6950381112200.138, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112200.669, "dur": 0.040 , "args": { "External id": 913,"cbid": 1, "correlation": 3519 } }, { "ph": "f", "id": 3519, "pid": 1728719, "tid": 1728719, "ts": 6950381112200.669, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112200.799, "dur": 0.101 , "args": { "External id": 913,"cbid": 468, "correlation": 3520 } }, { "ph": "f", "id": 3520, "pid": 1728719, "tid": 1728719, "ts": 6950381112200.799, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112201.050, "dur": 0.020 , "args": { "External id": 913,"cbid": 1, "correlation": 3522 } }, { "ph": "f", "id": 3522, "pid": 1728719, "tid": 1728719, "ts": 6950381112201.050, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112201.180, "dur": 0.090 , "args": { "External id": 913,"cbid": 468, "correlation": 3523 } }, { "ph": "f", "id": 3523, "pid": 1728719, "tid": 1728719, "ts": 6950381112201.180, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112201.441, "dur": 0.020 , "args": { "External id": 913,"cbid": 1, "correlation": 3525 } }, { "ph": "f", "id": 3525, "pid": 1728719, "tid": 1728719, "ts": 6950381112201.441, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112201.591, "dur": 0.691 , "args": { "External id": 913,"cbid": 273, "correlation": 3526 } }, { "ph": "f", "id": 3526, "pid": 1728719, "tid": 1728719, "ts": 6950381112201.591, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381128145.446, "dur": 215.870 , "args": { "External id": 913,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3529, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3529, "pid": 0, "tid": 7, "ts": 6950381128145.446, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112202.753, "dur": 1.924 , "args": { "External id": 913,"cbid": 211, "correlation": 3529 } }, { "ph": "s", "id": 3529, "pid": 1728719, "tid": 1728719, "ts": 6950381112202.753, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381128361.604, "dur": 8.736 , "args": { "External id": 925,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3542, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3542, "pid": 0, "tid": 7, "ts": 6950381128361.604, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112214.304, "dur": 1.674 , "args": { "External id": 925,"cbid": 307, "correlation": 3542 } }, { "ph": "s", "id": 3542, "pid": 1728719, "tid": 1728719, "ts": 6950381112214.304, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381128370.628, "dur": 36.575 , "args": { "External id": 930,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3565, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3565, "pid": 0, "tid": 7, "ts": 6950381128370.628, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112235.013, "dur": 1.703 , "args": { "External id": 930,"cbid": 307, "correlation": 3565 } }, { "ph": "s", "id": 3565, "pid": 1728719, "tid": 1728719, "ts": 6950381112235.013, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112245.753, "dur": 0.100 , "args": { "External id": 931,"cbid": 200, "correlation": 3580 } }, { "ph": "f", "id": 3580, "pid": 1728719, "tid": 1728719, "ts": 6950381112245.753, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112245.963, "dur": 0.151 , "args": { "External id": 931,"cbid": 468, "correlation": 3581 } }, { "ph": "f", "id": 3581, "pid": 1728719, "tid": 1728719, "ts": 6950381112245.963, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112246.334, "dur": 0.020 , "args": { "External id": 931,"cbid": 1, "correlation": 3583 } }, { "ph": "f", "id": 3583, "pid": 1728719, "tid": 1728719, "ts": 6950381112246.334, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112246.424, "dur": 0.100 , "args": { "External id": 931,"cbid": 468, "correlation": 3584 } }, { "ph": "f", "id": 3584, "pid": 1728719, "tid": 1728719, "ts": 6950381112246.424, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112246.685, "dur": 0.020 , "args": { "External id": 931,"cbid": 1, "correlation": 3586 } }, { "ph": "f", "id": 3586, "pid": 1728719, "tid": 1728719, "ts": 6950381112246.685, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112246.795, "dur": 0.090 , "args": { "External id": 931,"cbid": 468, "correlation": 3587 } }, { "ph": "f", "id": 3587, "pid": 1728719, "tid": 1728719, "ts": 6950381112246.795, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112247.086, "dur": 0.020 , "args": { "External id": 931,"cbid": 1, "correlation": 3589 } }, { "ph": "f", "id": 3589, "pid": 1728719, "tid": 1728719, "ts": 6950381112247.086, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112247.206, "dur": 0.390 , "args": { "External id": 931,"cbid": 273, "correlation": 3590 } }, { "ph": "f", "id": 3590, "pid": 1728719, "tid": 1728719, "ts": 6950381112247.206, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381128407.491, "dur": 229.726 , "args": { "External id": 931,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3593, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3593, "pid": 0, "tid": 7, "ts": 6950381128407.491, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112247.997, "dur": 1.753 , "args": { "External id": 931,"cbid": 211, "correlation": 3593 } }, { "ph": "s", "id": 3593, "pid": 1728719, "tid": 1728719, "ts": 6950381112247.997, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381128637.537, "dur": 8.768 , "args": { "External id": 943,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3602, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3602, "pid": 0, "tid": 7, "ts": 6950381128637.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112255.972, "dur": 1.593 , "args": { "External id": 943,"cbid": 307, "correlation": 3602 } }, { "ph": "s", "id": 3602, "pid": 1728719, "tid": 1728719, "ts": 6950381112255.972, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381128646.593, "dur": 26.719 , "args": { "External id": 948,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3625, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3625, "pid": 0, "tid": 7, "ts": 6950381128646.593, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112272.182, "dur": 1.543 , "args": { "External id": 948,"cbid": 307, "correlation": 3625 } }, { "ph": "s", "id": 3625, "pid": 1728719, "tid": 1728719, "ts": 6950381112272.182, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112282.191, "dur": 0.100 , "args": { "External id": 949,"cbid": 200, "correlation": 3640 } }, { "ph": "f", "id": 3640, "pid": 1728719, "tid": 1728719, "ts": 6950381112282.191, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112282.391, "dur": 0.120 , "args": { "External id": 949,"cbid": 468, "correlation": 3641 } }, { "ph": "f", "id": 3641, "pid": 1728719, "tid": 1728719, "ts": 6950381112282.391, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112282.712, "dur": 0.030 , "args": { "External id": 949,"cbid": 1, "correlation": 3643 } }, { "ph": "f", "id": 3643, "pid": 1728719, "tid": 1728719, "ts": 6950381112282.712, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112282.812, "dur": 0.090 , "args": { "External id": 949,"cbid": 468, "correlation": 3644 } }, { "ph": "f", "id": 3644, "pid": 1728719, "tid": 1728719, "ts": 6950381112282.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112283.062, "dur": 0.021 , "args": { "External id": 949,"cbid": 1, "correlation": 3646 } }, { "ph": "f", "id": 3646, "pid": 1728719, "tid": 1728719, "ts": 6950381112283.062, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112283.173, "dur": 0.100 , "args": { "External id": 949,"cbid": 468, "correlation": 3647 } }, { "ph": "f", "id": 3647, "pid": 1728719, "tid": 1728719, "ts": 6950381112283.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112283.423, "dur": 0.020 , "args": { "External id": 949,"cbid": 1, "correlation": 3649 } }, { "ph": "f", "id": 3649, "pid": 1728719, "tid": 1728719, "ts": 6950381112283.423, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112283.533, "dur": 0.371 , "args": { "External id": 949,"cbid": 273, "correlation": 3650 } }, { "ph": "f", "id": 3650, "pid": 1728719, "tid": 1728719, "ts": 6950381112283.533, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381128673.600, "dur": 231.614 , "args": { "External id": 949,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3653, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3653, "pid": 0, "tid": 7, "ts": 6950381128673.600, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112284.285, "dur": 1.683 , "args": { "External id": 949,"cbid": 211, "correlation": 3653 } }, { "ph": "s", "id": 3653, "pid": 1728719, "tid": 1728719, "ts": 6950381112284.285, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381128905.534, "dur": 58.911 , "args": { "External id": 961,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3667, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3667, "pid": 0, "tid": 7, "ts": 6950381128905.534, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112293.792, "dur": 1.563 , "args": { "External id": 961,"cbid": 307, "correlation": 3667 } }, { "ph": "s", "id": 3667, "pid": 1728719, "tid": 1728719, "ts": 6950381112293.792, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381128964.701, "dur": 66.047 , "args": { "External id": 962,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3677, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3677, "pid": 0, "tid": 7, "ts": 6950381128964.701, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112304.172, "dur": 1.663 , "args": { "External id": 962,"cbid": 307, "correlation": 3677 } }, { "ph": "s", "id": 3677, "pid": 1728719, "tid": 1728719, "ts": 6950381112304.172, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381129031.036, "dur": 39.552 , "args": { "External id": 963,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3686, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3686, "pid": 0, "tid": 7, "ts": 6950381129031.036, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112313.760, "dur": 1.663 , "args": { "External id": 963,"cbid": 307, "correlation": 3686 } }, { "ph": "s", "id": 3686, "pid": 1728719, "tid": 1728719, "ts": 6950381112313.760, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112329.329, "dur": 0.110 , "args": { "External id": 971,"cbid": 200, "correlation": 3721 } }, { "ph": "f", "id": 3721, "pid": 1728719, "tid": 1728719, "ts": 6950381112329.329, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112329.509, "dur": 0.080 , "args": { "External id": 971,"cbid": 200, "correlation": 3722 } }, { "ph": "f", "id": 3722, "pid": 1728719, "tid": 1728719, "ts": 6950381112329.509, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112329.679, "dur": 0.461 , "args": { "External id": 971,"cbid": 273, "correlation": 3723 } }, { "ph": "f", "id": 3723, "pid": 1728719, "tid": 1728719, "ts": 6950381112329.679, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381129070.876, "dur": 841.207 , "args": { "External id": 971,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3725, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3725, "pid": 0, "tid": 7, "ts": 6950381129070.876, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112330.541, "dur": 1.853 , "args": { "External id": 971,"cbid": 211, "correlation": 3725 } }, { "ph": "s", "id": 3725, "pid": 1728719, "tid": 1728719, "ts": 6950381112330.541, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381129912.371, "dur": 9.632 , "args": { "External id": 980,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3745, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3745, "pid": 0, "tid": 7, "ts": 6950381129912.371, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112347.392, "dur": 1.613 , "args": { "External id": 980,"cbid": 652, "correlation": 3745 } }, { "ph": "s", "id": 3745, "pid": 1728719, "tid": 1728719, "ts": 6950381112347.392, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112358.152, "dur": 0.110 , "args": { "External id": 981,"cbid": 200, "correlation": 3760 } }, { "ph": "f", "id": 3760, "pid": 1728719, "tid": 1728719, "ts": 6950381112358.152, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112358.363, "dur": 0.160 , "args": { "External id": 981,"cbid": 468, "correlation": 3761 } }, { "ph": "f", "id": 3761, "pid": 1728719, "tid": 1728719, "ts": 6950381112358.363, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112358.793, "dur": 0.031 , "args": { "External id": 981,"cbid": 1, "correlation": 3763 } }, { "ph": "f", "id": 3763, "pid": 1728719, "tid": 1728719, "ts": 6950381112358.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112358.894, "dur": 0.090 , "args": { "External id": 981,"cbid": 468, "correlation": 3764 } }, { "ph": "f", "id": 3764, "pid": 1728719, "tid": 1728719, "ts": 6950381112358.894, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112359.144, "dur": 0.020 , "args": { "External id": 981,"cbid": 1, "correlation": 3766 } }, { "ph": "f", "id": 3766, "pid": 1728719, "tid": 1728719, "ts": 6950381112359.144, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112359.254, "dur": 0.090 , "args": { "External id": 981,"cbid": 468, "correlation": 3767 } }, { "ph": "f", "id": 3767, "pid": 1728719, "tid": 1728719, "ts": 6950381112359.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112359.515, "dur": 0.020 , "args": { "External id": 981,"cbid": 1, "correlation": 3769 } }, { "ph": "f", "id": 3769, "pid": 1728719, "tid": 1728719, "ts": 6950381112359.515, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112359.615, "dur": 0.391 , "args": { "External id": 981,"cbid": 273, "correlation": 3770 } }, { "ph": "f", "id": 3770, "pid": 1728719, "tid": 1728719, "ts": 6950381112359.615, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381129922.259, "dur": 207.325 , "args": { "External id": 981,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3773, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3773, "pid": 0, "tid": 7, "ts": 6950381129922.259, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112360.376, "dur": 1.794 , "args": { "External id": 981,"cbid": 211, "correlation": 3773 } }, { "ph": "s", "id": 3773, "pid": 1728719, "tid": 1728719, "ts": 6950381112360.376, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381130129.872, "dur": 46.528 , "args": { "External id": 993,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3785, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3785, "pid": 0, "tid": 7, "ts": 6950381130129.872, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112370.445, "dur": 1.783 , "args": { "External id": 993,"cbid": 307, "correlation": 3785 } }, { "ph": "s", "id": 3785, "pid": 1728719, "tid": 1728719, "ts": 6950381112370.445, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381130176.688, "dur": 13.056 , "args": { "External id": 994,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3791, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3791, "pid": 0, "tid": 7, "ts": 6950381130176.688, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112378.420, "dur": 1.453 , "args": { "External id": 994,"cbid": 652, "correlation": 3791 } }, { "ph": "s", "id": 3791, "pid": 1728719, "tid": 1728719, "ts": 6950381112378.420, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381130190.064, "dur": 12.896 , "args": { "External id": 995,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3797, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3797, "pid": 0, "tid": 7, "ts": 6950381130190.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112386.685, "dur": 1.493 , "args": { "External id": 995,"cbid": 652, "correlation": 3797 } }, { "ph": "s", "id": 3797, "pid": 1728719, "tid": 1728719, "ts": 6950381112386.685, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112398.217, "dur": 0.090 , "args": { "External id": 996,"cbid": 200, "correlation": 3816 } }, { "ph": "f", "id": 3816, "pid": 1728719, "tid": 1728719, "ts": 6950381112398.217, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112398.417, "dur": 0.120 , "args": { "External id": 996,"cbid": 468, "correlation": 3817 } }, { "ph": "f", "id": 3817, "pid": 1728719, "tid": 1728719, "ts": 6950381112398.417, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112398.748, "dur": 0.020 , "args": { "External id": 996,"cbid": 1, "correlation": 3819 } }, { "ph": "f", "id": 3819, "pid": 1728719, "tid": 1728719, "ts": 6950381112398.748, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112398.838, "dur": 0.100 , "args": { "External id": 996,"cbid": 468, "correlation": 3820 } }, { "ph": "f", "id": 3820, "pid": 1728719, "tid": 1728719, "ts": 6950381112398.838, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112399.088, "dur": 0.020 , "args": { "External id": 996,"cbid": 1, "correlation": 3822 } }, { "ph": "f", "id": 3822, "pid": 1728719, "tid": 1728719, "ts": 6950381112399.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112399.199, "dur": 0.100 , "args": { "External id": 996,"cbid": 468, "correlation": 3823 } }, { "ph": "f", "id": 3823, "pid": 1728719, "tid": 1728719, "ts": 6950381112399.199, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112399.469, "dur": 0.020 , "args": { "External id": 996,"cbid": 1, "correlation": 3825 } }, { "ph": "f", "id": 3825, "pid": 1728719, "tid": 1728719, "ts": 6950381112399.469, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112399.569, "dur": 0.361 , "args": { "External id": 996,"cbid": 273, "correlation": 3826 } }, { "ph": "f", "id": 3826, "pid": 1728719, "tid": 1728719, "ts": 6950381112399.569, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381130203.280, "dur": 661.625 , "args": { "External id": 996,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3829, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3829, "pid": 0, "tid": 7, "ts": 6950381130203.280, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112400.281, "dur": 1.683 , "args": { "External id": 996,"cbid": 211, "correlation": 3829 } }, { "ph": "s", "id": 3829, "pid": 1728719, "tid": 1728719, "ts": 6950381112400.281, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112412.013, "dur": 0.090 , "args": { "External id": 1008,"cbid": 200, "correlation": 3853 } }, { "ph": "f", "id": 3853, "pid": 1728719, "tid": 1728719, "ts": 6950381112412.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112412.213, "dur": 0.110 , "args": { "External id": 1008,"cbid": 468, "correlation": 3854 } }, { "ph": "f", "id": 3854, "pid": 1728719, "tid": 1728719, "ts": 6950381112412.213, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112412.513, "dur": 0.030 , "args": { "External id": 1008,"cbid": 1, "correlation": 3856 } }, { "ph": "f", "id": 3856, "pid": 1728719, "tid": 1728719, "ts": 6950381112412.513, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112412.604, "dur": 0.100 , "args": { "External id": 1008,"cbid": 468, "correlation": 3857 } }, { "ph": "f", "id": 3857, "pid": 1728719, "tid": 1728719, "ts": 6950381112412.604, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112412.844, "dur": 0.020 , "args": { "External id": 1008,"cbid": 1, "correlation": 3859 } }, { "ph": "f", "id": 3859, "pid": 1728719, "tid": 1728719, "ts": 6950381112412.844, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112412.954, "dur": 0.100 , "args": { "External id": 1008,"cbid": 468, "correlation": 3860 } }, { "ph": "f", "id": 3860, "pid": 1728719, "tid": 1728719, "ts": 6950381112412.954, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112413.195, "dur": 0.020 , "args": { "External id": 1008,"cbid": 1, "correlation": 3862 } }, { "ph": "f", "id": 3862, "pid": 1728719, "tid": 1728719, "ts": 6950381112413.195, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112413.305, "dur": 0.321 , "args": { "External id": 1008,"cbid": 273, "correlation": 3863 } }, { "ph": "f", "id": 3863, "pid": 1728719, "tid": 1728719, "ts": 6950381112413.305, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381130865.193, "dur": 675.032 , "args": { "External id": 1008,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3866, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3866, "pid": 0, "tid": 7, "ts": 6950381130865.193, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112413.976, "dur": 1.663 , "args": { "External id": 1008,"cbid": 211, "correlation": 3866 } }, { "ph": "s", "id": 3866, "pid": 1728719, "tid": 1728719, "ts": 6950381112413.976, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381131540.513, "dur": 153.375 , "args": { "External id": 1020,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3879, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3879, "pid": 0, "tid": 7, "ts": 6950381131540.513, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112423.223, "dur": 1.663 , "args": { "External id": 1020,"cbid": 307, "correlation": 3879 } }, { "ph": "s", "id": 3879, "pid": 1728719, "tid": 1728719, "ts": 6950381112423.223, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381131694.144, "dur": 58.719 , "args": { "External id": 1021,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3885, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3885, "pid": 0, "tid": 7, "ts": 6950381131694.144, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112431.609, "dur": 1.713 , "args": { "External id": 1021,"cbid": 652, "correlation": 3885 } }, { "ph": "s", "id": 3885, "pid": 1728719, "tid": 1728719, "ts": 6950381112431.609, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112445.936, "dur": 0.090 , "args": { "External id": 1022,"cbid": 200, "correlation": 3904 } }, { "ph": "f", "id": 3904, "pid": 1728719, "tid": 1728719, "ts": 6950381112445.936, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112446.136, "dur": 0.140 , "args": { "External id": 1022,"cbid": 468, "correlation": 3905 } }, { "ph": "f", "id": 3905, "pid": 1728719, "tid": 1728719, "ts": 6950381112446.136, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112446.497, "dur": 0.020 , "args": { "External id": 1022,"cbid": 1, "correlation": 3907 } }, { "ph": "f", "id": 3907, "pid": 1728719, "tid": 1728719, "ts": 6950381112446.497, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112446.587, "dur": 0.100 , "args": { "External id": 1022,"cbid": 468, "correlation": 3908 } }, { "ph": "f", "id": 3908, "pid": 1728719, "tid": 1728719, "ts": 6950381112446.587, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112446.837, "dur": 0.020 , "args": { "External id": 1022,"cbid": 1, "correlation": 3910 } }, { "ph": "f", "id": 3910, "pid": 1728719, "tid": 1728719, "ts": 6950381112446.837, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112446.948, "dur": 0.090 , "args": { "External id": 1022,"cbid": 468, "correlation": 3911 } }, { "ph": "f", "id": 3911, "pid": 1728719, "tid": 1728719, "ts": 6950381112446.948, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112447.208, "dur": 0.030 , "args": { "External id": 1022,"cbid": 1, "correlation": 3913 } }, { "ph": "f", "id": 3913, "pid": 1728719, "tid": 1728719, "ts": 6950381112447.208, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112447.328, "dur": 0.431 , "args": { "External id": 1022,"cbid": 273, "correlation": 3914 } }, { "ph": "f", "id": 3914, "pid": 1728719, "tid": 1728719, "ts": 6950381112447.328, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381131753.151, "dur": 679.929 , "args": { "External id": 1022,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3917, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3917, "pid": 0, "tid": 7, "ts": 6950381131753.151, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112448.120, "dur": 1.773 , "args": { "External id": 1022,"cbid": 211, "correlation": 3917 } }, { "ph": "s", "id": 3917, "pid": 1728719, "tid": 1728719, "ts": 6950381112448.120, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381132433.368, "dur": 40.287 , "args": { "External id": 1034,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3928, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3928, "pid": 0, "tid": 7, "ts": 6950381132433.368, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112458.469, "dur": 1.653 , "args": { "External id": 1034,"cbid": 307, "correlation": 3928 } }, { "ph": "s", "id": 3928, "pid": 1728719, "tid": 1728719, "ts": 6950381112458.469, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381132473.943, "dur": 20.512 , "args": { "External id": 1046,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3940, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3940, "pid": 0, "tid": 7, "ts": 6950381132473.943, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112548.426, "dur": 2.776 , "args": { "External id": 1046,"cbid": 307, "correlation": 3940 } }, { "ph": "s", "id": 3940, "pid": 1728719, "tid": 1728719, "ts": 6950381112548.426, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381132494.743, "dur": 12.480 , "args": { "External id": 1047,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3954, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3954, "pid": 0, "tid": 7, "ts": 6950381132494.743, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112563.645, "dur": 1.743 , "args": { "External id": 1047,"cbid": 652, "correlation": 3954 } }, { "ph": "s", "id": 3954, "pid": 1728719, "tid": 1728719, "ts": 6950381112563.645, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381132507.543, "dur": 12.192 , "args": { "External id": 1048,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3968, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3968, "pid": 0, "tid": 7, "ts": 6950381132507.543, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112575.547, "dur": 1.573 , "args": { "External id": 1048,"cbid": 652, "correlation": 3968 } }, { "ph": "s", "id": 3968, "pid": 1728719, "tid": 1728719, "ts": 6950381112575.547, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381132520.023, "dur": 10.912 , "args": { "External id": 1049,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 3982, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 3982, "pid": 0, "tid": 7, "ts": 6950381132520.023, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112586.217, "dur": 1.623 , "args": { "External id": 1049,"cbid": 652, "correlation": 3982 } }, { "ph": "s", "id": 3982, "pid": 1728719, "tid": 1728719, "ts": 6950381112586.217, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112598.319, "dur": 0.201 , "args": { "External id": 1050,"cbid": 200, "correlation": 3997 } }, { "ph": "f", "id": 3997, "pid": 1728719, "tid": 1728719, "ts": 6950381112598.319, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112598.710, "dur": 0.210 , "args": { "External id": 1050,"cbid": 468, "correlation": 3998 } }, { "ph": "f", "id": 3998, "pid": 1728719, "tid": 1728719, "ts": 6950381112598.710, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112599.351, "dur": 0.030 , "args": { "External id": 1050,"cbid": 1, "correlation": 4000 } }, { "ph": "f", "id": 4000, "pid": 1728719, "tid": 1728719, "ts": 6950381112599.351, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112599.542, "dur": 0.110 , "args": { "External id": 1050,"cbid": 468, "correlation": 4001 } }, { "ph": "f", "id": 4001, "pid": 1728719, "tid": 1728719, "ts": 6950381112599.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112599.812, "dur": 0.030 , "args": { "External id": 1050,"cbid": 1, "correlation": 4003 } }, { "ph": "f", "id": 4003, "pid": 1728719, "tid": 1728719, "ts": 6950381112599.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112599.952, "dur": 0.091 , "args": { "External id": 1050,"cbid": 468, "correlation": 4004 } }, { "ph": "f", "id": 4004, "pid": 1728719, "tid": 1728719, "ts": 6950381112599.952, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112600.223, "dur": 0.020 , "args": { "External id": 1050,"cbid": 1, "correlation": 4006 } }, { "ph": "f", "id": 4006, "pid": 1728719, "tid": 1728719, "ts": 6950381112600.223, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112600.373, "dur": 0.651 , "args": { "External id": 1050,"cbid": 273, "correlation": 4007 } }, { "ph": "f", "id": 4007, "pid": 1728719, "tid": 1728719, "ts": 6950381112600.373, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381132531.223, "dur": 215.901 , "args": { "External id": 1050,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4010, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4010, "pid": 0, "tid": 7, "ts": 6950381132531.223, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112601.505, "dur": 1.924 , "args": { "External id": 1050,"cbid": 211, "correlation": 4010 } }, { "ph": "s", "id": 4010, "pid": 1728719, "tid": 1728719, "ts": 6950381112601.505, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381132747.412, "dur": 8.832 , "args": { "External id": 1062,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4023, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4023, "pid": 0, "tid": 7, "ts": 6950381132747.412, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112613.217, "dur": 1.753 , "args": { "External id": 1062,"cbid": 307, "correlation": 4023 } }, { "ph": "s", "id": 4023, "pid": 1728719, "tid": 1728719, "ts": 6950381112613.217, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381132756.532, "dur": 37.504 , "args": { "External id": 1067,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4046, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4046, "pid": 0, "tid": 7, "ts": 6950381132756.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112633.455, "dur": 1.713 , "args": { "External id": 1067,"cbid": 307, "correlation": 4046 } }, { "ph": "s", "id": 4046, "pid": 1728719, "tid": 1728719, "ts": 6950381112633.455, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112643.564, "dur": 0.110 , "args": { "External id": 1068,"cbid": 200, "correlation": 4061 } }, { "ph": "f", "id": 4061, "pid": 1728719, "tid": 1728719, "ts": 6950381112643.564, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112643.774, "dur": 0.130 , "args": { "External id": 1068,"cbid": 468, "correlation": 4062 } }, { "ph": "f", "id": 4062, "pid": 1728719, "tid": 1728719, "ts": 6950381112643.774, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112644.125, "dur": 0.030 , "args": { "External id": 1068,"cbid": 1, "correlation": 4064 } }, { "ph": "f", "id": 4064, "pid": 1728719, "tid": 1728719, "ts": 6950381112644.125, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112644.235, "dur": 0.100 , "args": { "External id": 1068,"cbid": 468, "correlation": 4065 } }, { "ph": "f", "id": 4065, "pid": 1728719, "tid": 1728719, "ts": 6950381112644.235, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112644.485, "dur": 0.020 , "args": { "External id": 1068,"cbid": 1, "correlation": 4067 } }, { "ph": "f", "id": 4067, "pid": 1728719, "tid": 1728719, "ts": 6950381112644.485, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112644.595, "dur": 0.101 , "args": { "External id": 1068,"cbid": 468, "correlation": 4068 } }, { "ph": "f", "id": 4068, "pid": 1728719, "tid": 1728719, "ts": 6950381112644.595, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112644.866, "dur": 0.020 , "args": { "External id": 1068,"cbid": 1, "correlation": 4070 } }, { "ph": "f", "id": 4070, "pid": 1728719, "tid": 1728719, "ts": 6950381112644.866, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112644.986, "dur": 0.481 , "args": { "External id": 1068,"cbid": 273, "correlation": 4071 } }, { "ph": "f", "id": 4071, "pid": 1728719, "tid": 1728719, "ts": 6950381112644.986, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381132794.324, "dur": 227.133 , "args": { "External id": 1068,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4074, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4074, "pid": 0, "tid": 7, "ts": 6950381132794.324, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112645.838, "dur": 1.663 , "args": { "External id": 1068,"cbid": 211, "correlation": 4074 } }, { "ph": "s", "id": 4074, "pid": 1728719, "tid": 1728719, "ts": 6950381112645.838, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381133021.777, "dur": 8.832 , "args": { "External id": 1080,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4083, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4083, "pid": 0, "tid": 7, "ts": 6950381133021.777, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112653.462, "dur": 1.493 , "args": { "External id": 1080,"cbid": 307, "correlation": 4083 } }, { "ph": "s", "id": 4083, "pid": 1728719, "tid": 1728719, "ts": 6950381112653.462, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381133030.897, "dur": 25.728 , "args": { "External id": 1085,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4106, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4106, "pid": 0, "tid": 7, "ts": 6950381133030.897, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112670.163, "dur": 1.493 , "args": { "External id": 1085,"cbid": 307, "correlation": 4106 } }, { "ph": "s", "id": 4106, "pid": 1728719, "tid": 1728719, "ts": 6950381112670.163, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112680.182, "dur": 0.090 , "args": { "External id": 1086,"cbid": 200, "correlation": 4121 } }, { "ph": "f", "id": 4121, "pid": 1728719, "tid": 1728719, "ts": 6950381112680.182, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112680.362, "dur": 0.120 , "args": { "External id": 1086,"cbid": 468, "correlation": 4122 } }, { "ph": "f", "id": 4122, "pid": 1728719, "tid": 1728719, "ts": 6950381112680.362, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112680.713, "dur": 0.020 , "args": { "External id": 1086,"cbid": 1, "correlation": 4124 } }, { "ph": "f", "id": 4124, "pid": 1728719, "tid": 1728719, "ts": 6950381112680.713, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112680.803, "dur": 0.100 , "args": { "External id": 1086,"cbid": 468, "correlation": 4125 } }, { "ph": "f", "id": 4125, "pid": 1728719, "tid": 1728719, "ts": 6950381112680.803, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112681.043, "dur": 0.020 , "args": { "External id": 1086,"cbid": 1, "correlation": 4127 } }, { "ph": "f", "id": 4127, "pid": 1728719, "tid": 1728719, "ts": 6950381112681.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112681.153, "dur": 0.091 , "args": { "External id": 1086,"cbid": 468, "correlation": 4128 } }, { "ph": "f", "id": 4128, "pid": 1728719, "tid": 1728719, "ts": 6950381112681.153, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112681.424, "dur": 0.020 , "args": { "External id": 1086,"cbid": 1, "correlation": 4130 } }, { "ph": "f", "id": 4130, "pid": 1728719, "tid": 1728719, "ts": 6950381112681.424, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112681.534, "dur": 0.361 , "args": { "External id": 1086,"cbid": 273, "correlation": 4131 } }, { "ph": "f", "id": 4131, "pid": 1728719, "tid": 1728719, "ts": 6950381112681.534, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381133056.881, "dur": 233.789 , "args": { "External id": 1086,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4134, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4134, "pid": 0, "tid": 7, "ts": 6950381133056.881, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112682.256, "dur": 1.713 , "args": { "External id": 1086,"cbid": 211, "correlation": 4134 } }, { "ph": "s", "id": 4134, "pid": 1728719, "tid": 1728719, "ts": 6950381112682.256, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381133290.958, "dur": 60.448 , "args": { "External id": 1098,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4148, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4148, "pid": 0, "tid": 7, "ts": 6950381133290.958, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112692.334, "dur": 1.613 , "args": { "External id": 1098,"cbid": 307, "correlation": 4148 } }, { "ph": "s", "id": 4148, "pid": 1728719, "tid": 1728719, "ts": 6950381112692.334, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381133351.694, "dur": 62.815 , "args": { "External id": 1099,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4158, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4158, "pid": 0, "tid": 7, "ts": 6950381133351.694, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112702.734, "dur": 1.683 , "args": { "External id": 1099,"cbid": 307, "correlation": 4158 } }, { "ph": "s", "id": 4158, "pid": 1728719, "tid": 1728719, "ts": 6950381112702.734, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381133414.797, "dur": 39.744 , "args": { "External id": 1100,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4167, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4167, "pid": 0, "tid": 7, "ts": 6950381133414.797, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112712.612, "dur": 1.603 , "args": { "External id": 1100,"cbid": 307, "correlation": 4167 } }, { "ph": "s", "id": 4167, "pid": 1728719, "tid": 1728719, "ts": 6950381112712.612, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112728.952, "dur": 0.121 , "args": { "External id": 1108,"cbid": 200, "correlation": 4202 } }, { "ph": "f", "id": 4202, "pid": 1728719, "tid": 1728719, "ts": 6950381112728.952, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112729.223, "dur": 0.070 , "args": { "External id": 1108,"cbid": 200, "correlation": 4203 } }, { "ph": "f", "id": 4203, "pid": 1728719, "tid": 1728719, "ts": 6950381112729.223, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112729.383, "dur": 0.471 , "args": { "External id": 1108,"cbid": 273, "correlation": 4204 } }, { "ph": "f", "id": 4204, "pid": 1728719, "tid": 1728719, "ts": 6950381112729.383, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381133454.829, "dur": 842.006 , "args": { "External id": 1108,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4206, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4206, "pid": 0, "tid": 7, "ts": 6950381133454.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112730.245, "dur": 1.933 , "args": { "External id": 1108,"cbid": 211, "correlation": 4206 } }, { "ph": "s", "id": 4206, "pid": 1728719, "tid": 1728719, "ts": 6950381112730.245, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381134297.123, "dur": 9.280 , "args": { "External id": 1117,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4226, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4226, "pid": 0, "tid": 7, "ts": 6950381134297.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112747.387, "dur": 1.563 , "args": { "External id": 1117,"cbid": 652, "correlation": 4226 } }, { "ph": "s", "id": 4226, "pid": 1728719, "tid": 1728719, "ts": 6950381112747.387, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112758.247, "dur": 0.100 , "args": { "External id": 1118,"cbid": 200, "correlation": 4241 } }, { "ph": "f", "id": 4241, "pid": 1728719, "tid": 1728719, "ts": 6950381112758.247, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112758.457, "dur": 0.211 , "args": { "External id": 1118,"cbid": 468, "correlation": 4242 } }, { "ph": "f", "id": 4242, "pid": 1728719, "tid": 1728719, "ts": 6950381112758.457, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112758.918, "dur": 0.030 , "args": { "External id": 1118,"cbid": 1, "correlation": 4244 } }, { "ph": "f", "id": 4244, "pid": 1728719, "tid": 1728719, "ts": 6950381112758.918, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112759.008, "dur": 0.271 , "args": { "External id": 1118,"cbid": 468, "correlation": 4245 } }, { "ph": "f", "id": 4245, "pid": 1728719, "tid": 1728719, "ts": 6950381112759.008, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112759.459, "dur": 0.020 , "args": { "External id": 1118,"cbid": 1, "correlation": 4247 } }, { "ph": "f", "id": 4247, "pid": 1728719, "tid": 1728719, "ts": 6950381112759.459, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112759.569, "dur": 0.101 , "args": { "External id": 1118,"cbid": 468, "correlation": 4248 } }, { "ph": "f", "id": 4248, "pid": 1728719, "tid": 1728719, "ts": 6950381112759.569, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112759.820, "dur": 0.020 , "args": { "External id": 1118,"cbid": 1, "correlation": 4250 } }, { "ph": "f", "id": 4250, "pid": 1728719, "tid": 1728719, "ts": 6950381112759.820, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112759.920, "dur": 0.421 , "args": { "External id": 1118,"cbid": 273, "correlation": 4251 } }, { "ph": "f", "id": 4251, "pid": 1728719, "tid": 1728719, "ts": 6950381112759.920, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381134306.659, "dur": 209.022 , "args": { "External id": 1118,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4254, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4254, "pid": 0, "tid": 7, "ts": 6950381134306.659, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112760.742, "dur": 1.743 , "args": { "External id": 1118,"cbid": 211, "correlation": 4254 } }, { "ph": "s", "id": 4254, "pid": 1728719, "tid": 1728719, "ts": 6950381112760.742, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381134515.969, "dur": 47.424 , "args": { "External id": 1130,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4266, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4266, "pid": 0, "tid": 7, "ts": 6950381134515.969, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112770.550, "dur": 1.703 , "args": { "External id": 1130,"cbid": 307, "correlation": 4266 } }, { "ph": "s", "id": 4266, "pid": 1728719, "tid": 1728719, "ts": 6950381112770.550, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381134563.713, "dur": 13.055 , "args": { "External id": 1131,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4272, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4272, "pid": 0, "tid": 7, "ts": 6950381134563.713, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112778.605, "dur": 1.453 , "args": { "External id": 1131,"cbid": 652, "correlation": 4272 } }, { "ph": "s", "id": 4272, "pid": 1728719, "tid": 1728719, "ts": 6950381112778.605, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381134577.089, "dur": 12.639 , "args": { "External id": 1132,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4278, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4278, "pid": 0, "tid": 7, "ts": 6950381134577.089, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112786.820, "dur": 1.463 , "args": { "External id": 1132,"cbid": 652, "correlation": 4278 } }, { "ph": "s", "id": 4278, "pid": 1728719, "tid": 1728719, "ts": 6950381112786.820, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112798.231, "dur": 0.101 , "args": { "External id": 1133,"cbid": 200, "correlation": 4297 } }, { "ph": "f", "id": 4297, "pid": 1728719, "tid": 1728719, "ts": 6950381112798.231, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112798.482, "dur": 0.120 , "args": { "External id": 1133,"cbid": 468, "correlation": 4298 } }, { "ph": "f", "id": 4298, "pid": 1728719, "tid": 1728719, "ts": 6950381112798.482, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112798.813, "dur": 0.020 , "args": { "External id": 1133,"cbid": 1, "correlation": 4300 } }, { "ph": "f", "id": 4300, "pid": 1728719, "tid": 1728719, "ts": 6950381112798.813, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112798.923, "dur": 0.100 , "args": { "External id": 1133,"cbid": 468, "correlation": 4301 } }, { "ph": "f", "id": 4301, "pid": 1728719, "tid": 1728719, "ts": 6950381112798.923, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112799.163, "dur": 0.030 , "args": { "External id": 1133,"cbid": 1, "correlation": 4303 } }, { "ph": "f", "id": 4303, "pid": 1728719, "tid": 1728719, "ts": 6950381112799.163, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112799.273, "dur": 0.101 , "args": { "External id": 1133,"cbid": 468, "correlation": 4304 } }, { "ph": "f", "id": 4304, "pid": 1728719, "tid": 1728719, "ts": 6950381112799.273, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112799.524, "dur": 0.020 , "args": { "External id": 1133,"cbid": 1, "correlation": 4306 } }, { "ph": "f", "id": 4306, "pid": 1728719, "tid": 1728719, "ts": 6950381112799.524, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112799.624, "dur": 0.371 , "args": { "External id": 1133,"cbid": 273, "correlation": 4307 } }, { "ph": "f", "id": 4307, "pid": 1728719, "tid": 1728719, "ts": 6950381112799.624, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381134590.016, "dur": 658.873 , "args": { "External id": 1133,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4310, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4310, "pid": 0, "tid": 7, "ts": 6950381134590.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112800.355, "dur": 1.704 , "args": { "External id": 1133,"cbid": 211, "correlation": 4310 } }, { "ph": "s", "id": 4310, "pid": 1728719, "tid": 1728719, "ts": 6950381112800.355, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112812.358, "dur": 0.080 , "args": { "External id": 1145,"cbid": 200, "correlation": 4334 } }, { "ph": "f", "id": 4334, "pid": 1728719, "tid": 1728719, "ts": 6950381112812.358, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112812.528, "dur": 0.110 , "args": { "External id": 1145,"cbid": 468, "correlation": 4335 } }, { "ph": "f", "id": 4335, "pid": 1728719, "tid": 1728719, "ts": 6950381112812.528, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112812.839, "dur": 0.020 , "args": { "External id": 1145,"cbid": 1, "correlation": 4337 } }, { "ph": "f", "id": 4337, "pid": 1728719, "tid": 1728719, "ts": 6950381112812.839, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112812.949, "dur": 0.100 , "args": { "External id": 1145,"cbid": 468, "correlation": 4338 } }, { "ph": "f", "id": 4338, "pid": 1728719, "tid": 1728719, "ts": 6950381112812.949, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112813.219, "dur": 0.020 , "args": { "External id": 1145,"cbid": 1, "correlation": 4340 } }, { "ph": "f", "id": 4340, "pid": 1728719, "tid": 1728719, "ts": 6950381112813.219, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112813.330, "dur": 0.100 , "args": { "External id": 1145,"cbid": 468, "correlation": 4341 } }, { "ph": "f", "id": 4341, "pid": 1728719, "tid": 1728719, "ts": 6950381112813.330, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112813.600, "dur": 0.020 , "args": { "External id": 1145,"cbid": 1, "correlation": 4343 } }, { "ph": "f", "id": 4343, "pid": 1728719, "tid": 1728719, "ts": 6950381112813.600, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112813.700, "dur": 0.321 , "args": { "External id": 1145,"cbid": 273, "correlation": 4344 } }, { "ph": "f", "id": 4344, "pid": 1728719, "tid": 1728719, "ts": 6950381112813.700, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381135249.177, "dur": 674.073 , "args": { "External id": 1145,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4347, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4347, "pid": 0, "tid": 7, "ts": 6950381135249.177, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112814.371, "dur": 1.634 , "args": { "External id": 1145,"cbid": 211, "correlation": 4347 } }, { "ph": "s", "id": 4347, "pid": 1728719, "tid": 1728719, "ts": 6950381112814.371, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381135923.570, "dur": 153.630 , "args": { "External id": 1157,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4360, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4360, "pid": 0, "tid": 7, "ts": 6950381135923.570, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112823.999, "dur": 1.643 , "args": { "External id": 1157,"cbid": 307, "correlation": 4360 } }, { "ph": "s", "id": 4360, "pid": 1728719, "tid": 1728719, "ts": 6950381112823.999, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381136077.456, "dur": 59.999 , "args": { "External id": 1158,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4366, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4366, "pid": 0, "tid": 7, "ts": 6950381136077.456, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112832.445, "dur": 1.844 , "args": { "External id": 1158,"cbid": 652, "correlation": 4366 } }, { "ph": "s", "id": 4366, "pid": 1728719, "tid": 1728719, "ts": 6950381112832.445, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112844.017, "dur": 0.090 , "args": { "External id": 1159,"cbid": 200, "correlation": 4385 } }, { "ph": "f", "id": 4385, "pid": 1728719, "tid": 1728719, "ts": 6950381112844.017, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112844.197, "dur": 0.110 , "args": { "External id": 1159,"cbid": 468, "correlation": 4386 } }, { "ph": "f", "id": 4386, "pid": 1728719, "tid": 1728719, "ts": 6950381112844.197, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112844.498, "dur": 0.030 , "args": { "External id": 1159,"cbid": 1, "correlation": 4388 } }, { "ph": "f", "id": 4388, "pid": 1728719, "tid": 1728719, "ts": 6950381112844.498, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112844.658, "dur": 0.100 , "args": { "External id": 1159,"cbid": 468, "correlation": 4389 } }, { "ph": "f", "id": 4389, "pid": 1728719, "tid": 1728719, "ts": 6950381112844.658, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112844.908, "dur": 0.020 , "args": { "External id": 1159,"cbid": 1, "correlation": 4391 } }, { "ph": "f", "id": 4391, "pid": 1728719, "tid": 1728719, "ts": 6950381112844.908, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112845.019, "dur": 0.100 , "args": { "External id": 1159,"cbid": 468, "correlation": 4392 } }, { "ph": "f", "id": 4392, "pid": 1728719, "tid": 1728719, "ts": 6950381112845.019, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112845.269, "dur": 0.020 , "args": { "External id": 1159,"cbid": 1, "correlation": 4394 } }, { "ph": "f", "id": 4394, "pid": 1728719, "tid": 1728719, "ts": 6950381112845.269, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112845.379, "dur": 0.351 , "args": { "External id": 1159,"cbid": 273, "correlation": 4395 } }, { "ph": "f", "id": 4395, "pid": 1728719, "tid": 1728719, "ts": 6950381112845.379, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381136137.743, "dur": 681.625 , "args": { "External id": 1159,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4398, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4398, "pid": 0, "tid": 7, "ts": 6950381136137.743, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112846.081, "dur": 1.713 , "args": { "External id": 1159,"cbid": 211, "correlation": 4398 } }, { "ph": "s", "id": 4398, "pid": 1728719, "tid": 1728719, "ts": 6950381112846.081, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381136819.656, "dur": 40.224 , "args": { "External id": 1171,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4409, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4409, "pid": 0, "tid": 7, "ts": 6950381136819.656, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112855.959, "dur": 1.643 , "args": { "External id": 1171,"cbid": 307, "correlation": 4409 } }, { "ph": "s", "id": 4409, "pid": 1728719, "tid": 1728719, "ts": 6950381112855.959, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381136860.136, "dur": 20.831 , "args": { "External id": 1183,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4421, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4421, "pid": 0, "tid": 7, "ts": 6950381136860.136, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112937.681, "dur": 2.415 , "args": { "External id": 1183,"cbid": 307, "correlation": 4421 } }, { "ph": "s", "id": 4421, "pid": 1728719, "tid": 1728719, "ts": 6950381112937.681, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381136881.287, "dur": 12.544 , "args": { "External id": 1184,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4435, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4435, "pid": 0, "tid": 7, "ts": 6950381136881.287, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112951.697, "dur": 1.693 , "args": { "External id": 1184,"cbid": 652, "correlation": 4435 } }, { "ph": "s", "id": 4435, "pid": 1728719, "tid": 1728719, "ts": 6950381112951.697, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381136894.151, "dur": 12.128 , "args": { "External id": 1185,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4449, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4449, "pid": 0, "tid": 7, "ts": 6950381136894.151, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112963.269, "dur": 1.613 , "args": { "External id": 1185,"cbid": 652, "correlation": 4449 } }, { "ph": "s", "id": 4449, "pid": 1728719, "tid": 1728719, "ts": 6950381112963.269, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381136906.567, "dur": 11.040 , "args": { "External id": 1186,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4463, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4463, "pid": 0, "tid": 7, "ts": 6950381136906.567, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381112973.858, "dur": 1.613 , "args": { "External id": 1186,"cbid": 652, "correlation": 4463 } }, { "ph": "s", "id": 4463, "pid": 1728719, "tid": 1728719, "ts": 6950381112973.858, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112985.310, "dur": 0.160 , "args": { "External id": 1187,"cbid": 200, "correlation": 4478 } }, { "ph": "f", "id": 4478, "pid": 1728719, "tid": 1728719, "ts": 6950381112985.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112985.640, "dur": 0.221 , "args": { "External id": 1187,"cbid": 468, "correlation": 4479 } }, { "ph": "f", "id": 4479, "pid": 1728719, "tid": 1728719, "ts": 6950381112985.640, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112986.181, "dur": 0.030 , "args": { "External id": 1187,"cbid": 1, "correlation": 4481 } }, { "ph": "f", "id": 4481, "pid": 1728719, "tid": 1728719, "ts": 6950381112986.181, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112986.281, "dur": 0.121 , "args": { "External id": 1187,"cbid": 468, "correlation": 4482 } }, { "ph": "f", "id": 4482, "pid": 1728719, "tid": 1728719, "ts": 6950381112986.281, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112986.552, "dur": 0.020 , "args": { "External id": 1187,"cbid": 1, "correlation": 4484 } }, { "ph": "f", "id": 4484, "pid": 1728719, "tid": 1728719, "ts": 6950381112986.552, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112986.682, "dur": 0.100 , "args": { "External id": 1187,"cbid": 468, "correlation": 4485 } }, { "ph": "f", "id": 4485, "pid": 1728719, "tid": 1728719, "ts": 6950381112986.682, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381112986.933, "dur": 0.020 , "args": { "External id": 1187,"cbid": 1, "correlation": 4487 } }, { "ph": "f", "id": 4487, "pid": 1728719, "tid": 1728719, "ts": 6950381112986.933, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381112987.083, "dur": 0.721 , "args": { "External id": 1187,"cbid": 273, "correlation": 4488 } }, { "ph": "f", "id": 4488, "pid": 1728719, "tid": 1728719, "ts": 6950381112987.083, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381136917.863, "dur": 215.582 , "args": { "External id": 1187,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4491, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4491, "pid": 0, "tid": 7, "ts": 6950381136917.863, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112988.295, "dur": 1.844 , "args": { "External id": 1187,"cbid": 211, "correlation": 4491 } }, { "ph": "s", "id": 4491, "pid": 1728719, "tid": 1728719, "ts": 6950381112988.295, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381137133.733, "dur": 8.864 , "args": { "External id": 1199,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4504, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4504, "pid": 0, "tid": 7, "ts": 6950381137133.733, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381112998.955, "dur": 1.783 , "args": { "External id": 1199,"cbid": 307, "correlation": 4504 } }, { "ph": "s", "id": 4504, "pid": 1728719, "tid": 1728719, "ts": 6950381112998.955, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381137142.852, "dur": 37.952 , "args": { "External id": 1204,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4527, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4527, "pid": 0, "tid": 7, "ts": 6950381137142.852, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113018.401, "dur": 1.683 , "args": { "External id": 1204,"cbid": 307, "correlation": 4527 } }, { "ph": "s", "id": 4527, "pid": 1728719, "tid": 1728719, "ts": 6950381113018.401, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113029.091, "dur": 0.120 , "args": { "External id": 1205,"cbid": 200, "correlation": 4542 } }, { "ph": "f", "id": 4542, "pid": 1728719, "tid": 1728719, "ts": 6950381113029.091, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113029.302, "dur": 0.130 , "args": { "External id": 1205,"cbid": 468, "correlation": 4543 } }, { "ph": "f", "id": 4543, "pid": 1728719, "tid": 1728719, "ts": 6950381113029.302, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113029.642, "dur": 0.030 , "args": { "External id": 1205,"cbid": 1, "correlation": 4545 } }, { "ph": "f", "id": 4545, "pid": 1728719, "tid": 1728719, "ts": 6950381113029.642, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113029.742, "dur": 0.101 , "args": { "External id": 1205,"cbid": 468, "correlation": 4546 } }, { "ph": "f", "id": 4546, "pid": 1728719, "tid": 1728719, "ts": 6950381113029.742, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113030.013, "dur": 0.020 , "args": { "External id": 1205,"cbid": 1, "correlation": 4548 } }, { "ph": "f", "id": 4548, "pid": 1728719, "tid": 1728719, "ts": 6950381113030.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113030.123, "dur": 0.090 , "args": { "External id": 1205,"cbid": 468, "correlation": 4549 } }, { "ph": "f", "id": 4549, "pid": 1728719, "tid": 1728719, "ts": 6950381113030.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113030.394, "dur": 0.020 , "args": { "External id": 1205,"cbid": 1, "correlation": 4551 } }, { "ph": "f", "id": 4551, "pid": 1728719, "tid": 1728719, "ts": 6950381113030.394, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113030.514, "dur": 0.421 , "args": { "External id": 1205,"cbid": 273, "correlation": 4552 } }, { "ph": "f", "id": 4552, "pid": 1728719, "tid": 1728719, "ts": 6950381113030.514, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381137181.092, "dur": 227.806 , "args": { "External id": 1205,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4555, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4555, "pid": 0, "tid": 7, "ts": 6950381137181.092, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113031.325, "dur": 1.673 , "args": { "External id": 1205,"cbid": 211, "correlation": 4555 } }, { "ph": "s", "id": 4555, "pid": 1728719, "tid": 1728719, "ts": 6950381113031.325, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381137409.186, "dur": 8.831 , "args": { "External id": 1217,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4564, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4564, "pid": 0, "tid": 7, "ts": 6950381137409.186, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113038.799, "dur": 1.583 , "args": { "External id": 1217,"cbid": 307, "correlation": 4564 } }, { "ph": "s", "id": 4564, "pid": 1728719, "tid": 1728719, "ts": 6950381113038.799, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381137418.306, "dur": 26.943 , "args": { "External id": 1222,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4587, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4587, "pid": 0, "tid": 7, "ts": 6950381137418.306, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113055.160, "dur": 1.563 , "args": { "External id": 1222,"cbid": 307, "correlation": 4587 } }, { "ph": "s", "id": 4587, "pid": 1728719, "tid": 1728719, "ts": 6950381113055.160, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113065.178, "dur": 0.091 , "args": { "External id": 1223,"cbid": 200, "correlation": 4602 } }, { "ph": "f", "id": 4602, "pid": 1728719, "tid": 1728719, "ts": 6950381113065.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113065.349, "dur": 0.130 , "args": { "External id": 1223,"cbid": 468, "correlation": 4603 } }, { "ph": "f", "id": 4603, "pid": 1728719, "tid": 1728719, "ts": 6950381113065.349, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113065.689, "dur": 0.020 , "args": { "External id": 1223,"cbid": 1, "correlation": 4605 } }, { "ph": "f", "id": 4605, "pid": 1728719, "tid": 1728719, "ts": 6950381113065.689, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113065.779, "dur": 0.091 , "args": { "External id": 1223,"cbid": 468, "correlation": 4606 } }, { "ph": "f", "id": 4606, "pid": 1728719, "tid": 1728719, "ts": 6950381113065.779, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113066.020, "dur": 0.020 , "args": { "External id": 1223,"cbid": 1, "correlation": 4608 } }, { "ph": "f", "id": 4608, "pid": 1728719, "tid": 1728719, "ts": 6950381113066.020, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113066.130, "dur": 0.100 , "args": { "External id": 1223,"cbid": 468, "correlation": 4609 } }, { "ph": "f", "id": 4609, "pid": 1728719, "tid": 1728719, "ts": 6950381113066.130, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113066.381, "dur": 0.020 , "args": { "External id": 1223,"cbid": 1, "correlation": 4611 } }, { "ph": "f", "id": 4611, "pid": 1728719, "tid": 1728719, "ts": 6950381113066.381, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113066.491, "dur": 0.391 , "args": { "External id": 1223,"cbid": 273, "correlation": 4612 } }, { "ph": "f", "id": 4612, "pid": 1728719, "tid": 1728719, "ts": 6950381113066.491, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381137445.505, "dur": 231.326 , "args": { "External id": 1223,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4615, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4615, "pid": 0, "tid": 7, "ts": 6950381137445.505, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113067.232, "dur": 1.753 , "args": { "External id": 1223,"cbid": 211, "correlation": 4615 } }, { "ph": "s", "id": 4615, "pid": 1728719, "tid": 1728719, "ts": 6950381113067.232, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381137677.119, "dur": 59.167 , "args": { "External id": 1235,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4629, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4629, "pid": 0, "tid": 7, "ts": 6950381137677.119, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113076.720, "dur": 1.533 , "args": { "External id": 1235,"cbid": 307, "correlation": 4629 } }, { "ph": "s", "id": 4629, "pid": 1728719, "tid": 1728719, "ts": 6950381113076.720, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381137736.574, "dur": 65.983 , "args": { "External id": 1236,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4639, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4639, "pid": 0, "tid": 7, "ts": 6950381137736.574, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113086.568, "dur": 1.613 , "args": { "External id": 1236,"cbid": 307, "correlation": 4639 } }, { "ph": "s", "id": 4639, "pid": 1728719, "tid": 1728719, "ts": 6950381113086.568, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381137802.813, "dur": 39.840 , "args": { "External id": 1237,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4648, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4648, "pid": 0, "tid": 7, "ts": 6950381137802.813, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113095.725, "dur": 1.613 , "args": { "External id": 1237,"cbid": 307, "correlation": 4648 } }, { "ph": "s", "id": 4648, "pid": 1728719, "tid": 1728719, "ts": 6950381113095.725, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113111.264, "dur": 0.130 , "args": { "External id": 1245,"cbid": 200, "correlation": 4683 } }, { "ph": "f", "id": 4683, "pid": 1728719, "tid": 1728719, "ts": 6950381113111.264, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113111.454, "dur": 0.081 , "args": { "External id": 1245,"cbid": 200, "correlation": 4684 } }, { "ph": "f", "id": 4684, "pid": 1728719, "tid": 1728719, "ts": 6950381113111.454, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113111.625, "dur": 0.491 , "args": { "External id": 1245,"cbid": 273, "correlation": 4685 } }, { "ph": "f", "id": 4685, "pid": 1728719, "tid": 1728719, "ts": 6950381113111.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381137842.909, "dur": 844.087 , "args": { "External id": 1245,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4687, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4687, "pid": 0, "tid": 7, "ts": 6950381137842.909, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113112.436, "dur": 1.804 , "args": { "External id": 1245,"cbid": 211, "correlation": 4687 } }, { "ph": "s", "id": 4687, "pid": 1728719, "tid": 1728719, "ts": 6950381113112.436, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381138687.252, "dur": 9.376 , "args": { "External id": 1254,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4707, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4707, "pid": 0, "tid": 7, "ts": 6950381138687.252, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113128.947, "dur": 1.773 , "args": { "External id": 1254,"cbid": 652, "correlation": 4707 } }, { "ph": "s", "id": 4707, "pid": 1728719, "tid": 1728719, "ts": 6950381113128.947, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113139.937, "dur": 0.111 , "args": { "External id": 1255,"cbid": 200, "correlation": 4722 } }, { "ph": "f", "id": 4722, "pid": 1728719, "tid": 1728719, "ts": 6950381113139.937, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113140.158, "dur": 0.170 , "args": { "External id": 1255,"cbid": 468, "correlation": 4723 } }, { "ph": "f", "id": 4723, "pid": 1728719, "tid": 1728719, "ts": 6950381113140.158, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113140.569, "dur": 0.030 , "args": { "External id": 1255,"cbid": 1, "correlation": 4725 } }, { "ph": "f", "id": 4725, "pid": 1728719, "tid": 1728719, "ts": 6950381113140.569, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113140.669, "dur": 0.170 , "args": { "External id": 1255,"cbid": 468, "correlation": 4726 } }, { "ph": "f", "id": 4726, "pid": 1728719, "tid": 1728719, "ts": 6950381113140.669, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113140.989, "dur": 0.020 , "args": { "External id": 1255,"cbid": 1, "correlation": 4728 } }, { "ph": "f", "id": 4728, "pid": 1728719, "tid": 1728719, "ts": 6950381113140.989, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113141.110, "dur": 0.090 , "args": { "External id": 1255,"cbid": 468, "correlation": 4729 } }, { "ph": "f", "id": 4729, "pid": 1728719, "tid": 1728719, "ts": 6950381113141.110, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113141.410, "dur": 0.020 , "args": { "External id": 1255,"cbid": 1, "correlation": 4731 } }, { "ph": "f", "id": 4731, "pid": 1728719, "tid": 1728719, "ts": 6950381113141.410, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113141.520, "dur": 0.401 , "args": { "External id": 1255,"cbid": 273, "correlation": 4732 } }, { "ph": "f", "id": 4732, "pid": 1728719, "tid": 1728719, "ts": 6950381113141.520, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381138696.884, "dur": 209.918 , "args": { "External id": 1255,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4735, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4735, "pid": 0, "tid": 7, "ts": 6950381138696.884, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113142.312, "dur": 1.683 , "args": { "External id": 1255,"cbid": 211, "correlation": 4735 } }, { "ph": "s", "id": 4735, "pid": 1728719, "tid": 1728719, "ts": 6950381113142.312, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381138907.089, "dur": 46.016 , "args": { "External id": 1267,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4747, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4747, "pid": 0, "tid": 7, "ts": 6950381138907.089, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113152.311, "dur": 1.693 , "args": { "External id": 1267,"cbid": 307, "correlation": 4747 } }, { "ph": "s", "id": 4747, "pid": 1728719, "tid": 1728719, "ts": 6950381113152.311, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381138953.425, "dur": 13.056 , "args": { "External id": 1268,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4753, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4753, "pid": 0, "tid": 7, "ts": 6950381138953.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113160.566, "dur": 1.453 , "args": { "External id": 1268,"cbid": 652, "correlation": 4753 } }, { "ph": "s", "id": 4753, "pid": 1728719, "tid": 1728719, "ts": 6950381113160.566, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381138966.769, "dur": 12.864 , "args": { "External id": 1269,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4759, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4759, "pid": 0, "tid": 7, "ts": 6950381138966.769, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113168.881, "dur": 1.613 , "args": { "External id": 1269,"cbid": 652, "correlation": 4759 } }, { "ph": "s", "id": 4759, "pid": 1728719, "tid": 1728719, "ts": 6950381113168.881, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113180.553, "dur": 0.100 , "args": { "External id": 1270,"cbid": 200, "correlation": 4778 } }, { "ph": "f", "id": 4778, "pid": 1728719, "tid": 1728719, "ts": 6950381113180.553, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113180.753, "dur": 0.121 , "args": { "External id": 1270,"cbid": 468, "correlation": 4779 } }, { "ph": "f", "id": 4779, "pid": 1728719, "tid": 1728719, "ts": 6950381113180.753, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113181.084, "dur": 0.020 , "args": { "External id": 1270,"cbid": 1, "correlation": 4781 } }, { "ph": "f", "id": 4781, "pid": 1728719, "tid": 1728719, "ts": 6950381113181.084, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113181.174, "dur": 0.100 , "args": { "External id": 1270,"cbid": 468, "correlation": 4782 } }, { "ph": "f", "id": 4782, "pid": 1728719, "tid": 1728719, "ts": 6950381113181.174, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113181.425, "dur": 0.020 , "args": { "External id": 1270,"cbid": 1, "correlation": 4784 } }, { "ph": "f", "id": 4784, "pid": 1728719, "tid": 1728719, "ts": 6950381113181.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113181.545, "dur": 0.100 , "args": { "External id": 1270,"cbid": 468, "correlation": 4785 } }, { "ph": "f", "id": 4785, "pid": 1728719, "tid": 1728719, "ts": 6950381113181.545, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113181.795, "dur": 0.020 , "args": { "External id": 1270,"cbid": 1, "correlation": 4787 } }, { "ph": "f", "id": 4787, "pid": 1728719, "tid": 1728719, "ts": 6950381113181.795, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113181.916, "dur": 0.360 , "args": { "External id": 1270,"cbid": 273, "correlation": 4788 } }, { "ph": "f", "id": 4788, "pid": 1728719, "tid": 1728719, "ts": 6950381113181.916, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381138979.953, "dur": 664.793 , "args": { "External id": 1270,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4791, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4791, "pid": 0, "tid": 7, "ts": 6950381138979.953, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113182.627, "dur": 1.653 , "args": { "External id": 1270,"cbid": 211, "correlation": 4791 } }, { "ph": "s", "id": 4791, "pid": 1728719, "tid": 1728719, "ts": 6950381113182.627, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113194.299, "dur": 0.090 , "args": { "External id": 1282,"cbid": 200, "correlation": 4815 } }, { "ph": "f", "id": 4815, "pid": 1728719, "tid": 1728719, "ts": 6950381113194.299, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113194.479, "dur": 0.110 , "args": { "External id": 1282,"cbid": 468, "correlation": 4816 } }, { "ph": "f", "id": 4816, "pid": 1728719, "tid": 1728719, "ts": 6950381113194.479, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113194.790, "dur": 0.020 , "args": { "External id": 1282,"cbid": 1, "correlation": 4818 } }, { "ph": "f", "id": 4818, "pid": 1728719, "tid": 1728719, "ts": 6950381113194.790, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113194.880, "dur": 0.090 , "args": { "External id": 1282,"cbid": 468, "correlation": 4819 } }, { "ph": "f", "id": 4819, "pid": 1728719, "tid": 1728719, "ts": 6950381113194.880, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113195.130, "dur": 0.030 , "args": { "External id": 1282,"cbid": 1, "correlation": 4821 } }, { "ph": "f", "id": 4821, "pid": 1728719, "tid": 1728719, "ts": 6950381113195.130, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113195.240, "dur": 0.101 , "args": { "External id": 1282,"cbid": 468, "correlation": 4822 } }, { "ph": "f", "id": 4822, "pid": 1728719, "tid": 1728719, "ts": 6950381113195.240, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113195.531, "dur": 0.020 , "args": { "External id": 1282,"cbid": 1, "correlation": 4824 } }, { "ph": "f", "id": 4824, "pid": 1728719, "tid": 1728719, "ts": 6950381113195.531, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113195.631, "dur": 0.341 , "args": { "External id": 1282,"cbid": 273, "correlation": 4825 } }, { "ph": "f", "id": 4825, "pid": 1728719, "tid": 1728719, "ts": 6950381113195.631, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381139645.034, "dur": 674.040 , "args": { "External id": 1282,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4828, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4828, "pid": 0, "tid": 7, "ts": 6950381139645.034, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113196.312, "dur": 1.633 , "args": { "External id": 1282,"cbid": 211, "correlation": 4828 } }, { "ph": "s", "id": 4828, "pid": 1728719, "tid": 1728719, "ts": 6950381113196.312, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381140319.362, "dur": 156.255 , "args": { "External id": 1294,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4841, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4841, "pid": 0, "tid": 7, "ts": 6950381140319.362, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113205.459, "dur": 1.704 , "args": { "External id": 1294,"cbid": 307, "correlation": 4841 } }, { "ph": "s", "id": 4841, "pid": 1728719, "tid": 1728719, "ts": 6950381113205.459, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381140475.872, "dur": 58.112 , "args": { "External id": 1295,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4847, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4847, "pid": 0, "tid": 7, "ts": 6950381140475.872, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113213.765, "dur": 1.753 , "args": { "External id": 1295,"cbid": 652, "correlation": 4847 } }, { "ph": "s", "id": 4847, "pid": 1728719, "tid": 1728719, "ts": 6950381113213.765, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113225.076, "dur": 0.080 , "args": { "External id": 1296,"cbid": 200, "correlation": 4866 } }, { "ph": "f", "id": 4866, "pid": 1728719, "tid": 1728719, "ts": 6950381113225.076, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113225.236, "dur": 0.110 , "args": { "External id": 1296,"cbid": 468, "correlation": 4867 } }, { "ph": "f", "id": 4867, "pid": 1728719, "tid": 1728719, "ts": 6950381113225.236, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113225.587, "dur": 0.020 , "args": { "External id": 1296,"cbid": 1, "correlation": 4869 } }, { "ph": "f", "id": 4869, "pid": 1728719, "tid": 1728719, "ts": 6950381113225.587, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113225.677, "dur": 0.150 , "args": { "External id": 1296,"cbid": 468, "correlation": 4870 } }, { "ph": "f", "id": 4870, "pid": 1728719, "tid": 1728719, "ts": 6950381113225.677, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113225.978, "dur": 0.020 , "args": { "External id": 1296,"cbid": 1, "correlation": 4872 } }, { "ph": "f", "id": 4872, "pid": 1728719, "tid": 1728719, "ts": 6950381113225.978, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113226.088, "dur": 0.090 , "args": { "External id": 1296,"cbid": 468, "correlation": 4873 } }, { "ph": "f", "id": 4873, "pid": 1728719, "tid": 1728719, "ts": 6950381113226.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113226.328, "dur": 0.020 , "args": { "External id": 1296,"cbid": 1, "correlation": 4875 } }, { "ph": "f", "id": 4875, "pid": 1728719, "tid": 1728719, "ts": 6950381113226.328, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113226.449, "dur": 0.350 , "args": { "External id": 1296,"cbid": 273, "correlation": 4876 } }, { "ph": "f", "id": 4876, "pid": 1728719, "tid": 1728719, "ts": 6950381113226.449, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381140534.272, "dur": 680.089 , "args": { "External id": 1296,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4879, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4879, "pid": 0, "tid": 7, "ts": 6950381140534.272, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113227.150, "dur": 1.653 , "args": { "External id": 1296,"cbid": 211, "correlation": 4879 } }, { "ph": "s", "id": 4879, "pid": 1728719, "tid": 1728719, "ts": 6950381113227.150, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381141214.649, "dur": 40.415 , "args": { "External id": 1308,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4890, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4890, "pid": 0, "tid": 7, "ts": 6950381141214.649, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113237.189, "dur": 1.683 , "args": { "External id": 1308,"cbid": 307, "correlation": 4890 } }, { "ph": "s", "id": 4890, "pid": 1728719, "tid": 1728719, "ts": 6950381113237.189, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381141255.320, "dur": 20.768 , "args": { "External id": 1320,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4902, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4902, "pid": 0, "tid": 7, "ts": 6950381141255.320, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113318.169, "dur": 2.685 , "args": { "External id": 1320,"cbid": 307, "correlation": 4902 } }, { "ph": "s", "id": 4902, "pid": 1728719, "tid": 1728719, "ts": 6950381113318.169, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381141276.408, "dur": 12.448 , "args": { "External id": 1321,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4916, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4916, "pid": 0, "tid": 7, "ts": 6950381141276.408, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113332.586, "dur": 1.633 , "args": { "External id": 1321,"cbid": 652, "correlation": 4916 } }, { "ph": "s", "id": 4916, "pid": 1728719, "tid": 1728719, "ts": 6950381113332.586, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381141289.176, "dur": 12.320 , "args": { "External id": 1322,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4930, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4930, "pid": 0, "tid": 7, "ts": 6950381141289.176, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113344.318, "dur": 1.643 , "args": { "External id": 1322,"cbid": 652, "correlation": 4930 } }, { "ph": "s", "id": 4930, "pid": 1728719, "tid": 1728719, "ts": 6950381113344.318, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381141301.784, "dur": 10.816 , "args": { "External id": 1323,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4944, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4944, "pid": 0, "tid": 7, "ts": 6950381141301.784, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113354.948, "dur": 1.563 , "args": { "External id": 1323,"cbid": 652, "correlation": 4944 } }, { "ph": "s", "id": 4944, "pid": 1728719, "tid": 1728719, "ts": 6950381113354.948, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113366.249, "dur": 0.180 , "args": { "External id": 1324,"cbid": 200, "correlation": 4959 } }, { "ph": "f", "id": 4959, "pid": 1728719, "tid": 1728719, "ts": 6950381113366.249, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113366.589, "dur": 0.271 , "args": { "External id": 1324,"cbid": 468, "correlation": 4960 } }, { "ph": "f", "id": 4960, "pid": 1728719, "tid": 1728719, "ts": 6950381113366.589, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113367.180, "dur": 0.031 , "args": { "External id": 1324,"cbid": 1, "correlation": 4962 } }, { "ph": "f", "id": 4962, "pid": 1728719, "tid": 1728719, "ts": 6950381113367.180, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113367.341, "dur": 0.100 , "args": { "External id": 1324,"cbid": 468, "correlation": 4963 } }, { "ph": "f", "id": 4963, "pid": 1728719, "tid": 1728719, "ts": 6950381113367.341, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113367.591, "dur": 0.020 , "args": { "External id": 1324,"cbid": 1, "correlation": 4965 } }, { "ph": "f", "id": 4965, "pid": 1728719, "tid": 1728719, "ts": 6950381113367.591, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113367.732, "dur": 0.100 , "args": { "External id": 1324,"cbid": 468, "correlation": 4966 } }, { "ph": "f", "id": 4966, "pid": 1728719, "tid": 1728719, "ts": 6950381113367.732, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113368.022, "dur": 0.020 , "args": { "External id": 1324,"cbid": 1, "correlation": 4968 } }, { "ph": "f", "id": 4968, "pid": 1728719, "tid": 1728719, "ts": 6950381113368.022, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113368.162, "dur": 0.611 , "args": { "External id": 1324,"cbid": 273, "correlation": 4969 } }, { "ph": "f", "id": 4969, "pid": 1728719, "tid": 1728719, "ts": 6950381113368.162, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381141312.888, "dur": 216.669 , "args": { "External id": 1324,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4972, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4972, "pid": 0, "tid": 7, "ts": 6950381141312.888, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113369.254, "dur": 1.744 , "args": { "External id": 1324,"cbid": 211, "correlation": 4972 } }, { "ph": "s", "id": 4972, "pid": 1728719, "tid": 1728719, "ts": 6950381113369.254, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381141529.877, "dur": 8.832 , "args": { "External id": 1336,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 4985, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 4985, "pid": 0, "tid": 7, "ts": 6950381141529.877, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113379.644, "dur": 1.683 , "args": { "External id": 1336,"cbid": 307, "correlation": 4985 } }, { "ph": "s", "id": 4985, "pid": 1728719, "tid": 1728719, "ts": 6950381113379.644, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381141538.997, "dur": 36.544 , "args": { "External id": 1341,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5008, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5008, "pid": 0, "tid": 7, "ts": 6950381141538.997, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113398.479, "dur": 1.663 , "args": { "External id": 1341,"cbid": 307, "correlation": 5008 } }, { "ph": "s", "id": 5008, "pid": 1728719, "tid": 1728719, "ts": 6950381113398.479, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113408.808, "dur": 0.100 , "args": { "External id": 1342,"cbid": 200, "correlation": 5023 } }, { "ph": "f", "id": 5023, "pid": 1728719, "tid": 1728719, "ts": 6950381113408.808, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113408.988, "dur": 0.141 , "args": { "External id": 1342,"cbid": 468, "correlation": 5024 } }, { "ph": "f", "id": 5024, "pid": 1728719, "tid": 1728719, "ts": 6950381113408.988, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113409.359, "dur": 0.040 , "args": { "External id": 1342,"cbid": 1, "correlation": 5026 } }, { "ph": "f", "id": 5026, "pid": 1728719, "tid": 1728719, "ts": 6950381113409.359, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113409.469, "dur": 0.130 , "args": { "External id": 1342,"cbid": 468, "correlation": 5027 } }, { "ph": "f", "id": 5027, "pid": 1728719, "tid": 1728719, "ts": 6950381113409.469, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113409.750, "dur": 0.030 , "args": { "External id": 1342,"cbid": 1, "correlation": 5029 } }, { "ph": "f", "id": 5029, "pid": 1728719, "tid": 1728719, "ts": 6950381113409.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113409.880, "dur": 0.100 , "args": { "External id": 1342,"cbid": 468, "correlation": 5030 } }, { "ph": "f", "id": 5030, "pid": 1728719, "tid": 1728719, "ts": 6950381113409.880, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113410.140, "dur": 0.020 , "args": { "External id": 1342,"cbid": 1, "correlation": 5032 } }, { "ph": "f", "id": 5032, "pid": 1728719, "tid": 1728719, "ts": 6950381113410.140, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113410.261, "dur": 0.441 , "args": { "External id": 1342,"cbid": 273, "correlation": 5033 } }, { "ph": "f", "id": 5033, "pid": 1728719, "tid": 1728719, "ts": 6950381113410.261, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381141575.829, "dur": 227.005 , "args": { "External id": 1342,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5036, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5036, "pid": 0, "tid": 7, "ts": 6950381141575.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113411.102, "dur": 1.774 , "args": { "External id": 1342,"cbid": 211, "correlation": 5036 } }, { "ph": "s", "id": 5036, "pid": 1728719, "tid": 1728719, "ts": 6950381113411.102, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381141803.122, "dur": 8.864 , "args": { "External id": 1354,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5045, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5045, "pid": 0, "tid": 7, "ts": 6950381141803.122, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113418.696, "dur": 1.593 , "args": { "External id": 1354,"cbid": 307, "correlation": 5045 } }, { "ph": "s", "id": 5045, "pid": 1728719, "tid": 1728719, "ts": 6950381113418.696, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381141812.274, "dur": 25.792 , "args": { "External id": 1359,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5068, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5068, "pid": 0, "tid": 7, "ts": 6950381141812.274, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113435.788, "dur": 1.633 , "args": { "External id": 1359,"cbid": 307, "correlation": 5068 } }, { "ph": "s", "id": 5068, "pid": 1728719, "tid": 1728719, "ts": 6950381113435.788, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113446.288, "dur": 0.080 , "args": { "External id": 1360,"cbid": 200, "correlation": 5083 } }, { "ph": "f", "id": 5083, "pid": 1728719, "tid": 1728719, "ts": 6950381113446.288, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113446.468, "dur": 0.130 , "args": { "External id": 1360,"cbid": 468, "correlation": 5084 } }, { "ph": "f", "id": 5084, "pid": 1728719, "tid": 1728719, "ts": 6950381113446.468, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113446.809, "dur": 0.030 , "args": { "External id": 1360,"cbid": 1, "correlation": 5086 } }, { "ph": "f", "id": 5086, "pid": 1728719, "tid": 1728719, "ts": 6950381113446.809, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113446.909, "dur": 0.100 , "args": { "External id": 1360,"cbid": 468, "correlation": 5087 } }, { "ph": "f", "id": 5087, "pid": 1728719, "tid": 1728719, "ts": 6950381113446.909, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113447.159, "dur": 0.020 , "args": { "External id": 1360,"cbid": 1, "correlation": 5089 } }, { "ph": "f", "id": 5089, "pid": 1728719, "tid": 1728719, "ts": 6950381113447.159, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113447.270, "dur": 0.100 , "args": { "External id": 1360,"cbid": 468, "correlation": 5090 } }, { "ph": "f", "id": 5090, "pid": 1728719, "tid": 1728719, "ts": 6950381113447.270, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113447.530, "dur": 0.020 , "args": { "External id": 1360,"cbid": 1, "correlation": 5092 } }, { "ph": "f", "id": 5092, "pid": 1728719, "tid": 1728719, "ts": 6950381113447.530, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113447.630, "dur": 0.371 , "args": { "External id": 1360,"cbid": 273, "correlation": 5093 } }, { "ph": "f", "id": 5093, "pid": 1728719, "tid": 1728719, "ts": 6950381113447.630, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381141838.354, "dur": 232.637 , "args": { "External id": 1360,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5096, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5096, "pid": 0, "tid": 7, "ts": 6950381141838.354, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113448.372, "dur": 1.653 , "args": { "External id": 1360,"cbid": 211, "correlation": 5096 } }, { "ph": "s", "id": 5096, "pid": 1728719, "tid": 1728719, "ts": 6950381113448.372, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381142071.279, "dur": 60.096 , "args": { "External id": 1372,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5110, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5110, "pid": 0, "tid": 7, "ts": 6950381142071.279, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113457.899, "dur": 1.543 , "args": { "External id": 1372,"cbid": 307, "correlation": 5110 } }, { "ph": "s", "id": 5110, "pid": 1728719, "tid": 1728719, "ts": 6950381113457.899, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381142131.631, "dur": 62.015 , "args": { "External id": 1373,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5120, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5120, "pid": 0, "tid": 7, "ts": 6950381142131.631, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113468.229, "dur": 1.663 , "args": { "External id": 1373,"cbid": 307, "correlation": 5120 } }, { "ph": "s", "id": 5120, "pid": 1728719, "tid": 1728719, "ts": 6950381113468.229, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381142193.902, "dur": 41.471 , "args": { "External id": 1374,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5129, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5129, "pid": 0, "tid": 7, "ts": 6950381142193.902, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113478.418, "dur": 1.583 , "args": { "External id": 1374,"cbid": 307, "correlation": 5129 } }, { "ph": "s", "id": 5129, "pid": 1728719, "tid": 1728719, "ts": 6950381113478.418, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113493.626, "dur": 0.100 , "args": { "External id": 1382,"cbid": 200, "correlation": 5164 } }, { "ph": "f", "id": 5164, "pid": 1728719, "tid": 1728719, "ts": 6950381113493.626, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113493.806, "dur": 0.090 , "args": { "External id": 1382,"cbid": 200, "correlation": 5165 } }, { "ph": "f", "id": 5165, "pid": 1728719, "tid": 1728719, "ts": 6950381113493.806, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113493.997, "dur": 0.470 , "args": { "External id": 1382,"cbid": 273, "correlation": 5166 } }, { "ph": "f", "id": 5166, "pid": 1728719, "tid": 1728719, "ts": 6950381113493.997, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381142235.661, "dur": 844.215 , "args": { "External id": 1382,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5168, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5168, "pid": 0, "tid": 7, "ts": 6950381142235.661, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113494.798, "dur": 1.753 , "args": { "External id": 1382,"cbid": 211, "correlation": 5168 } }, { "ph": "s", "id": 5168, "pid": 1728719, "tid": 1728719, "ts": 6950381113494.798, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381143080.164, "dur": 9.408 , "args": { "External id": 1391,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5188, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5188, "pid": 0, "tid": 7, "ts": 6950381143080.164, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113511.138, "dur": 1.583 , "args": { "External id": 1391,"cbid": 652, "correlation": 5188 } }, { "ph": "s", "id": 5188, "pid": 1728719, "tid": 1728719, "ts": 6950381113511.138, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113521.808, "dur": 0.111 , "args": { "External id": 1392,"cbid": 200, "correlation": 5203 } }, { "ph": "f", "id": 5203, "pid": 1728719, "tid": 1728719, "ts": 6950381113521.808, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113522.019, "dur": 0.140 , "args": { "External id": 1392,"cbid": 468, "correlation": 5204 } }, { "ph": "f", "id": 5204, "pid": 1728719, "tid": 1728719, "ts": 6950381113522.019, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113522.470, "dur": 0.030 , "args": { "External id": 1392,"cbid": 1, "correlation": 5206 } }, { "ph": "f", "id": 5206, "pid": 1728719, "tid": 1728719, "ts": 6950381113522.470, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113522.680, "dur": 0.090 , "args": { "External id": 1392,"cbid": 468, "correlation": 5207 } }, { "ph": "f", "id": 5207, "pid": 1728719, "tid": 1728719, "ts": 6950381113522.680, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113522.950, "dur": 0.020 , "args": { "External id": 1392,"cbid": 1, "correlation": 5209 } }, { "ph": "f", "id": 5209, "pid": 1728719, "tid": 1728719, "ts": 6950381113522.950, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113523.061, "dur": 0.100 , "args": { "External id": 1392,"cbid": 468, "correlation": 5210 } }, { "ph": "f", "id": 5210, "pid": 1728719, "tid": 1728719, "ts": 6950381113523.061, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113523.311, "dur": 0.030 , "args": { "External id": 1392,"cbid": 1, "correlation": 5212 } }, { "ph": "f", "id": 5212, "pid": 1728719, "tid": 1728719, "ts": 6950381113523.311, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113523.421, "dur": 0.411 , "args": { "External id": 1392,"cbid": 273, "correlation": 5213 } }, { "ph": "f", "id": 5213, "pid": 1728719, "tid": 1728719, "ts": 6950381113523.421, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381143089.828, "dur": 209.150 , "args": { "External id": 1392,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5216, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5216, "pid": 0, "tid": 7, "ts": 6950381143089.828, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113524.183, "dur": 1.713 , "args": { "External id": 1392,"cbid": 211, "correlation": 5216 } }, { "ph": "s", "id": 5216, "pid": 1728719, "tid": 1728719, "ts": 6950381113524.183, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381143299.266, "dur": 47.231 , "args": { "External id": 1404,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5228, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5228, "pid": 0, "tid": 7, "ts": 6950381143299.266, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113534.652, "dur": 1.643 , "args": { "External id": 1404,"cbid": 307, "correlation": 5228 } }, { "ph": "s", "id": 5228, "pid": 1728719, "tid": 1728719, "ts": 6950381113534.652, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381143346.817, "dur": 13.056 , "args": { "External id": 1405,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5234, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5234, "pid": 0, "tid": 7, "ts": 6950381143346.817, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113546.224, "dur": 1.513 , "args": { "External id": 1405,"cbid": 652, "correlation": 5234 } }, { "ph": "s", "id": 5234, "pid": 1728719, "tid": 1728719, "ts": 6950381113546.224, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381143360.193, "dur": 13.088 , "args": { "External id": 1406,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5240, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5240, "pid": 0, "tid": 7, "ts": 6950381143360.193, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113554.579, "dur": 1.583 , "args": { "External id": 1406,"cbid": 652, "correlation": 5240 } }, { "ph": "s", "id": 5240, "pid": 1728719, "tid": 1728719, "ts": 6950381113554.579, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113566.642, "dur": 0.170 , "args": { "External id": 1407,"cbid": 200, "correlation": 5259 } }, { "ph": "f", "id": 5259, "pid": 1728719, "tid": 1728719, "ts": 6950381113566.642, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113566.932, "dur": 0.131 , "args": { "External id": 1407,"cbid": 468, "correlation": 5260 } }, { "ph": "f", "id": 5260, "pid": 1728719, "tid": 1728719, "ts": 6950381113566.932, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113567.323, "dur": 0.020 , "args": { "External id": 1407,"cbid": 1, "correlation": 5262 } }, { "ph": "f", "id": 5262, "pid": 1728719, "tid": 1728719, "ts": 6950381113567.323, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113567.413, "dur": 0.100 , "args": { "External id": 1407,"cbid": 468, "correlation": 5263 } }, { "ph": "f", "id": 5263, "pid": 1728719, "tid": 1728719, "ts": 6950381113567.413, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113567.664, "dur": 0.020 , "args": { "External id": 1407,"cbid": 1, "correlation": 5265 } }, { "ph": "f", "id": 5265, "pid": 1728719, "tid": 1728719, "ts": 6950381113567.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113567.774, "dur": 0.100 , "args": { "External id": 1407,"cbid": 468, "correlation": 5266 } }, { "ph": "f", "id": 5266, "pid": 1728719, "tid": 1728719, "ts": 6950381113567.774, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113568.034, "dur": 0.020 , "args": { "External id": 1407,"cbid": 1, "correlation": 5268 } }, { "ph": "f", "id": 5268, "pid": 1728719, "tid": 1728719, "ts": 6950381113568.034, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113568.135, "dur": 0.420 , "args": { "External id": 1407,"cbid": 273, "correlation": 5269 } }, { "ph": "f", "id": 5269, "pid": 1728719, "tid": 1728719, "ts": 6950381113568.135, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381143373.601, "dur": 657.721 , "args": { "External id": 1407,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5272, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5272, "pid": 0, "tid": 7, "ts": 6950381143373.601, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113568.926, "dur": 1.713 , "args": { "External id": 1407,"cbid": 211, "correlation": 5272 } }, { "ph": "s", "id": 5272, "pid": 1728719, "tid": 1728719, "ts": 6950381113568.926, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113580.698, "dur": 0.090 , "args": { "External id": 1419,"cbid": 200, "correlation": 5296 } }, { "ph": "f", "id": 5296, "pid": 1728719, "tid": 1728719, "ts": 6950381113580.698, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113580.888, "dur": 0.111 , "args": { "External id": 1419,"cbid": 468, "correlation": 5297 } }, { "ph": "f", "id": 5297, "pid": 1728719, "tid": 1728719, "ts": 6950381113580.888, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113581.189, "dur": 0.020 , "args": { "External id": 1419,"cbid": 1, "correlation": 5299 } }, { "ph": "f", "id": 5299, "pid": 1728719, "tid": 1728719, "ts": 6950381113581.189, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113581.269, "dur": 0.100 , "args": { "External id": 1419,"cbid": 468, "correlation": 5300 } }, { "ph": "f", "id": 5300, "pid": 1728719, "tid": 1728719, "ts": 6950381113581.269, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113581.529, "dur": 0.021 , "args": { "External id": 1419,"cbid": 1, "correlation": 5302 } }, { "ph": "f", "id": 5302, "pid": 1728719, "tid": 1728719, "ts": 6950381113581.529, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113581.640, "dur": 0.100 , "args": { "External id": 1419,"cbid": 468, "correlation": 5303 } }, { "ph": "f", "id": 5303, "pid": 1728719, "tid": 1728719, "ts": 6950381113581.640, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113581.890, "dur": 0.020 , "args": { "External id": 1419,"cbid": 1, "correlation": 5305 } }, { "ph": "f", "id": 5305, "pid": 1728719, "tid": 1728719, "ts": 6950381113581.890, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113581.990, "dur": 0.351 , "args": { "External id": 1419,"cbid": 273, "correlation": 5306 } }, { "ph": "f", "id": 5306, "pid": 1728719, "tid": 1728719, "ts": 6950381113581.990, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381144031.610, "dur": 677.593 , "args": { "External id": 1419,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5309, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5309, "pid": 0, "tid": 7, "ts": 6950381144031.610, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113582.692, "dur": 1.703 , "args": { "External id": 1419,"cbid": 211, "correlation": 5309 } }, { "ph": "s", "id": 5309, "pid": 1728719, "tid": 1728719, "ts": 6950381113582.692, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381144709.491, "dur": 155.710 , "args": { "External id": 1431,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5322, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5322, "pid": 0, "tid": 7, "ts": 6950381144709.491, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113592.029, "dur": 1.793 , "args": { "External id": 1431,"cbid": 307, "correlation": 5322 } }, { "ph": "s", "id": 5322, "pid": 1728719, "tid": 1728719, "ts": 6950381113592.029, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381144865.457, "dur": 57.919 , "args": { "External id": 1432,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5328, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5328, "pid": 0, "tid": 7, "ts": 6950381144865.457, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113600.956, "dur": 1.723 , "args": { "External id": 1432,"cbid": 652, "correlation": 5328 } }, { "ph": "s", "id": 5328, "pid": 1728719, "tid": 1728719, "ts": 6950381113600.956, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113612.217, "dur": 0.100 , "args": { "External id": 1433,"cbid": 200, "correlation": 5347 } }, { "ph": "f", "id": 5347, "pid": 1728719, "tid": 1728719, "ts": 6950381113612.217, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113612.417, "dur": 0.110 , "args": { "External id": 1433,"cbid": 468, "correlation": 5348 } }, { "ph": "f", "id": 5348, "pid": 1728719, "tid": 1728719, "ts": 6950381113612.417, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113612.718, "dur": 0.030 , "args": { "External id": 1433,"cbid": 1, "correlation": 5350 } }, { "ph": "f", "id": 5350, "pid": 1728719, "tid": 1728719, "ts": 6950381113612.718, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113612.818, "dur": 0.100 , "args": { "External id": 1433,"cbid": 468, "correlation": 5351 } }, { "ph": "f", "id": 5351, "pid": 1728719, "tid": 1728719, "ts": 6950381113612.818, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113613.058, "dur": 0.020 , "args": { "External id": 1433,"cbid": 1, "correlation": 5353 } }, { "ph": "f", "id": 5353, "pid": 1728719, "tid": 1728719, "ts": 6950381113613.058, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113613.188, "dur": 0.101 , "args": { "External id": 1433,"cbid": 468, "correlation": 5354 } }, { "ph": "f", "id": 5354, "pid": 1728719, "tid": 1728719, "ts": 6950381113613.188, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113613.449, "dur": 0.020 , "args": { "External id": 1433,"cbid": 1, "correlation": 5356 } }, { "ph": "f", "id": 5356, "pid": 1728719, "tid": 1728719, "ts": 6950381113613.449, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113613.559, "dur": 0.371 , "args": { "External id": 1433,"cbid": 273, "correlation": 5357 } }, { "ph": "f", "id": 5357, "pid": 1728719, "tid": 1728719, "ts": 6950381113613.559, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381144923.664, "dur": 677.529 , "args": { "External id": 1433,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5360, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5360, "pid": 0, "tid": 7, "ts": 6950381144923.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113614.300, "dur": 1.714 , "args": { "External id": 1433,"cbid": 211, "correlation": 5360 } }, { "ph": "s", "id": 5360, "pid": 1728719, "tid": 1728719, "ts": 6950381113614.300, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381145601.481, "dur": 40.160 , "args": { "External id": 1445,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5371, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5371, "pid": 0, "tid": 7, "ts": 6950381145601.481, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113624.980, "dur": 1.704 , "args": { "External id": 1445,"cbid": 307, "correlation": 5371 } }, { "ph": "s", "id": 5371, "pid": 1728719, "tid": 1728719, "ts": 6950381113624.980, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381145641.897, "dur": 20.543 , "args": { "External id": 1457,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5383, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5383, "pid": 0, "tid": 7, "ts": 6950381145641.897, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113706.702, "dur": 2.555 , "args": { "External id": 1457,"cbid": 307, "correlation": 5383 } }, { "ph": "s", "id": 5383, "pid": 1728719, "tid": 1728719, "ts": 6950381113706.702, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381145662.728, "dur": 12.640 , "args": { "External id": 1458,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5397, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5397, "pid": 0, "tid": 7, "ts": 6950381145662.728, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113721.039, "dur": 1.603 , "args": { "External id": 1458,"cbid": 652, "correlation": 5397 } }, { "ph": "s", "id": 5397, "pid": 1728719, "tid": 1728719, "ts": 6950381113721.039, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381145675.688, "dur": 12.384 , "args": { "External id": 1459,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5411, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5411, "pid": 0, "tid": 7, "ts": 6950381145675.688, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113732.681, "dur": 1.433 , "args": { "External id": 1459,"cbid": 652, "correlation": 5411 } }, { "ph": "s", "id": 5411, "pid": 1728719, "tid": 1728719, "ts": 6950381113732.681, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381145688.360, "dur": 10.848 , "args": { "External id": 1460,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5425, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5425, "pid": 0, "tid": 7, "ts": 6950381145688.360, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113743.220, "dur": 1.573 , "args": { "External id": 1460,"cbid": 652, "correlation": 5425 } }, { "ph": "s", "id": 5425, "pid": 1728719, "tid": 1728719, "ts": 6950381113743.220, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113754.782, "dur": 0.140 , "args": { "External id": 1461,"cbid": 200, "correlation": 5440 } }, { "ph": "f", "id": 5440, "pid": 1728719, "tid": 1728719, "ts": 6950381113754.782, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113755.052, "dur": 0.221 , "args": { "External id": 1461,"cbid": 468, "correlation": 5441 } }, { "ph": "f", "id": 5441, "pid": 1728719, "tid": 1728719, "ts": 6950381113755.052, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113755.624, "dur": 0.030 , "args": { "External id": 1461,"cbid": 1, "correlation": 5443 } }, { "ph": "f", "id": 5443, "pid": 1728719, "tid": 1728719, "ts": 6950381113755.624, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113755.724, "dur": 0.100 , "args": { "External id": 1461,"cbid": 468, "correlation": 5444 } }, { "ph": "f", "id": 5444, "pid": 1728719, "tid": 1728719, "ts": 6950381113755.724, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113755.974, "dur": 0.020 , "args": { "External id": 1461,"cbid": 1, "correlation": 5446 } }, { "ph": "f", "id": 5446, "pid": 1728719, "tid": 1728719, "ts": 6950381113755.974, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113756.094, "dur": 0.101 , "args": { "External id": 1461,"cbid": 468, "correlation": 5447 } }, { "ph": "f", "id": 5447, "pid": 1728719, "tid": 1728719, "ts": 6950381113756.094, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113756.365, "dur": 0.040 , "args": { "External id": 1461,"cbid": 1, "correlation": 5449 } }, { "ph": "f", "id": 5449, "pid": 1728719, "tid": 1728719, "ts": 6950381113756.365, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113756.505, "dur": 0.651 , "args": { "External id": 1461,"cbid": 273, "correlation": 5450 } }, { "ph": "f", "id": 5450, "pid": 1728719, "tid": 1728719, "ts": 6950381113756.505, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381145699.496, "dur": 215.710 , "args": { "External id": 1461,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5453, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5453, "pid": 0, "tid": 7, "ts": 6950381145699.496, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113757.617, "dur": 1.994 , "args": { "External id": 1461,"cbid": 211, "correlation": 5453 } }, { "ph": "s", "id": 5453, "pid": 1728719, "tid": 1728719, "ts": 6950381113757.617, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381145915.494, "dur": 8.895 , "args": { "External id": 1473,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5466, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5466, "pid": 0, "tid": 7, "ts": 6950381145915.494, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113768.427, "dur": 1.684 , "args": { "External id": 1473,"cbid": 307, "correlation": 5466 } }, { "ph": "s", "id": 5466, "pid": 1728719, "tid": 1728719, "ts": 6950381113768.427, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381145924.677, "dur": 38.208 , "args": { "External id": 1478,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5489, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5489, "pid": 0, "tid": 7, "ts": 6950381145924.677, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113788.715, "dur": 1.643 , "args": { "External id": 1478,"cbid": 307, "correlation": 5489 } }, { "ph": "s", "id": 5489, "pid": 1728719, "tid": 1728719, "ts": 6950381113788.715, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113799.125, "dur": 0.110 , "args": { "External id": 1479,"cbid": 200, "correlation": 5504 } }, { "ph": "f", "id": 5504, "pid": 1728719, "tid": 1728719, "ts": 6950381113799.125, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113799.335, "dur": 0.140 , "args": { "External id": 1479,"cbid": 468, "correlation": 5505 } }, { "ph": "f", "id": 5505, "pid": 1728719, "tid": 1728719, "ts": 6950381113799.335, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113799.696, "dur": 0.020 , "args": { "External id": 1479,"cbid": 1, "correlation": 5507 } }, { "ph": "f", "id": 5507, "pid": 1728719, "tid": 1728719, "ts": 6950381113799.696, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113799.786, "dur": 0.100 , "args": { "External id": 1479,"cbid": 468, "correlation": 5508 } }, { "ph": "f", "id": 5508, "pid": 1728719, "tid": 1728719, "ts": 6950381113799.786, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113800.026, "dur": 0.030 , "args": { "External id": 1479,"cbid": 1, "correlation": 5510 } }, { "ph": "f", "id": 5510, "pid": 1728719, "tid": 1728719, "ts": 6950381113800.026, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113800.156, "dur": 0.101 , "args": { "External id": 1479,"cbid": 468, "correlation": 5511 } }, { "ph": "f", "id": 5511, "pid": 1728719, "tid": 1728719, "ts": 6950381113800.156, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113800.407, "dur": 0.020 , "args": { "External id": 1479,"cbid": 1, "correlation": 5513 } }, { "ph": "f", "id": 5513, "pid": 1728719, "tid": 1728719, "ts": 6950381113800.407, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113800.527, "dur": 0.451 , "args": { "External id": 1479,"cbid": 273, "correlation": 5514 } }, { "ph": "f", "id": 5514, "pid": 1728719, "tid": 1728719, "ts": 6950381113800.527, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381145963.173, "dur": 225.726 , "args": { "External id": 1479,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5517, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5517, "pid": 0, "tid": 7, "ts": 6950381145963.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113801.359, "dur": 1.663 , "args": { "External id": 1479,"cbid": 211, "correlation": 5517 } }, { "ph": "s", "id": 5517, "pid": 1728719, "tid": 1728719, "ts": 6950381113801.359, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381146189.187, "dur": 8.832 , "args": { "External id": 1491,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5526, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5526, "pid": 0, "tid": 7, "ts": 6950381146189.187, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113808.893, "dur": 1.513 , "args": { "External id": 1491,"cbid": 307, "correlation": 5526 } }, { "ph": "s", "id": 5526, "pid": 1728719, "tid": 1728719, "ts": 6950381113808.893, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381146198.307, "dur": 26.751 , "args": { "External id": 1496,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5549, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5549, "pid": 0, "tid": 7, "ts": 6950381146198.307, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113825.173, "dur": 1.653 , "args": { "External id": 1496,"cbid": 307, "correlation": 5549 } }, { "ph": "s", "id": 5549, "pid": 1728719, "tid": 1728719, "ts": 6950381113825.173, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113835.292, "dur": 0.090 , "args": { "External id": 1497,"cbid": 200, "correlation": 5564 } }, { "ph": "f", "id": 5564, "pid": 1728719, "tid": 1728719, "ts": 6950381113835.292, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113835.482, "dur": 0.130 , "args": { "External id": 1497,"cbid": 468, "correlation": 5565 } }, { "ph": "f", "id": 5565, "pid": 1728719, "tid": 1728719, "ts": 6950381113835.482, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113835.813, "dur": 0.020 , "args": { "External id": 1497,"cbid": 1, "correlation": 5567 } }, { "ph": "f", "id": 5567, "pid": 1728719, "tid": 1728719, "ts": 6950381113835.813, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113835.903, "dur": 0.100 , "args": { "External id": 1497,"cbid": 468, "correlation": 5568 } }, { "ph": "f", "id": 5568, "pid": 1728719, "tid": 1728719, "ts": 6950381113835.903, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113836.143, "dur": 0.020 , "args": { "External id": 1497,"cbid": 1, "correlation": 5570 } }, { "ph": "f", "id": 5570, "pid": 1728719, "tid": 1728719, "ts": 6950381113836.143, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113836.264, "dur": 0.090 , "args": { "External id": 1497,"cbid": 468, "correlation": 5571 } }, { "ph": "f", "id": 5571, "pid": 1728719, "tid": 1728719, "ts": 6950381113836.264, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113836.524, "dur": 0.030 , "args": { "External id": 1497,"cbid": 1, "correlation": 5573 } }, { "ph": "f", "id": 5573, "pid": 1728719, "tid": 1728719, "ts": 6950381113836.524, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113836.634, "dur": 0.471 , "args": { "External id": 1497,"cbid": 273, "correlation": 5574 } }, { "ph": "f", "id": 5574, "pid": 1728719, "tid": 1728719, "ts": 6950381113836.634, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381146225.346, "dur": 230.430 , "args": { "External id": 1497,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5577, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5577, "pid": 0, "tid": 7, "ts": 6950381146225.346, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113837.456, "dur": 1.793 , "args": { "External id": 1497,"cbid": 211, "correlation": 5577 } }, { "ph": "s", "id": 5577, "pid": 1728719, "tid": 1728719, "ts": 6950381113837.456, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381146456.064, "dur": 59.199 , "args": { "External id": 1509,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5591, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5591, "pid": 0, "tid": 7, "ts": 6950381146456.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113847.214, "dur": 1.563 , "args": { "External id": 1509,"cbid": 307, "correlation": 5591 } }, { "ph": "s", "id": 5591, "pid": 1728719, "tid": 1728719, "ts": 6950381113847.214, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381146515.551, "dur": 65.663 , "args": { "External id": 1510,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5601, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5601, "pid": 0, "tid": 7, "ts": 6950381146515.551, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113856.982, "dur": 1.753 , "args": { "External id": 1510,"cbid": 307, "correlation": 5601 } }, { "ph": "s", "id": 5601, "pid": 1728719, "tid": 1728719, "ts": 6950381113856.982, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381146581.502, "dur": 41.120 , "args": { "External id": 1511,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5610, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5610, "pid": 0, "tid": 7, "ts": 6950381146581.502, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113866.610, "dur": 1.904 , "args": { "External id": 1511,"cbid": 307, "correlation": 5610 } }, { "ph": "s", "id": 5610, "pid": 1728719, "tid": 1728719, "ts": 6950381113866.610, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113882.600, "dur": 0.110 , "args": { "External id": 1519,"cbid": 200, "correlation": 5645 } }, { "ph": "f", "id": 5645, "pid": 1728719, "tid": 1728719, "ts": 6950381113882.600, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113882.780, "dur": 0.080 , "args": { "External id": 1519,"cbid": 200, "correlation": 5646 } }, { "ph": "f", "id": 5646, "pid": 1728719, "tid": 1728719, "ts": 6950381113882.780, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113882.961, "dur": 0.460 , "args": { "External id": 1519,"cbid": 273, "correlation": 5647 } }, { "ph": "f", "id": 5647, "pid": 1728719, "tid": 1728719, "ts": 6950381113882.961, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381146622.878, "dur": 844.791 , "args": { "External id": 1519,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5649, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5649, "pid": 0, "tid": 7, "ts": 6950381146622.878, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113883.822, "dur": 1.864 , "args": { "External id": 1519,"cbid": 211, "correlation": 5649 } }, { "ph": "s", "id": 5649, "pid": 1728719, "tid": 1728719, "ts": 6950381113883.822, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381147467.957, "dur": 9.504 , "args": { "External id": 1528,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5669, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5669, "pid": 0, "tid": 7, "ts": 6950381147467.957, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113900.032, "dur": 1.573 , "args": { "External id": 1528,"cbid": 652, "correlation": 5669 } }, { "ph": "s", "id": 5669, "pid": 1728719, "tid": 1728719, "ts": 6950381113900.032, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113913.908, "dur": 0.110 , "args": { "External id": 1529,"cbid": 200, "correlation": 5684 } }, { "ph": "f", "id": 5684, "pid": 1728719, "tid": 1728719, "ts": 6950381113913.908, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113914.129, "dur": 0.180 , "args": { "External id": 1529,"cbid": 468, "correlation": 5685 } }, { "ph": "f", "id": 5685, "pid": 1728719, "tid": 1728719, "ts": 6950381113914.129, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113914.620, "dur": 0.030 , "args": { "External id": 1529,"cbid": 1, "correlation": 5687 } }, { "ph": "f", "id": 5687, "pid": 1728719, "tid": 1728719, "ts": 6950381113914.620, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113914.720, "dur": 0.090 , "args": { "External id": 1529,"cbid": 468, "correlation": 5688 } }, { "ph": "f", "id": 5688, "pid": 1728719, "tid": 1728719, "ts": 6950381113914.720, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113914.960, "dur": 0.030 , "args": { "External id": 1529,"cbid": 1, "correlation": 5690 } }, { "ph": "f", "id": 5690, "pid": 1728719, "tid": 1728719, "ts": 6950381113914.960, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113915.090, "dur": 0.101 , "args": { "External id": 1529,"cbid": 468, "correlation": 5691 } }, { "ph": "f", "id": 5691, "pid": 1728719, "tid": 1728719, "ts": 6950381113915.090, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113915.371, "dur": 0.020 , "args": { "External id": 1529,"cbid": 1, "correlation": 5693 } }, { "ph": "f", "id": 5693, "pid": 1728719, "tid": 1728719, "ts": 6950381113915.371, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113915.481, "dur": 0.451 , "args": { "External id": 1529,"cbid": 273, "correlation": 5694 } }, { "ph": "f", "id": 5694, "pid": 1728719, "tid": 1728719, "ts": 6950381113915.481, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381147477.749, "dur": 210.077 , "args": { "External id": 1529,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5697, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5697, "pid": 0, "tid": 7, "ts": 6950381147477.749, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113916.333, "dur": 1.723 , "args": { "External id": 1529,"cbid": 211, "correlation": 5697 } }, { "ph": "s", "id": 5697, "pid": 1728719, "tid": 1728719, "ts": 6950381113916.333, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381147688.146, "dur": 45.600 , "args": { "External id": 1541,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5709, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5709, "pid": 0, "tid": 7, "ts": 6950381147688.146, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113926.321, "dur": 1.653 , "args": { "External id": 1541,"cbid": 307, "correlation": 5709 } }, { "ph": "s", "id": 5709, "pid": 1728719, "tid": 1728719, "ts": 6950381113926.321, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381147734.034, "dur": 13.440 , "args": { "External id": 1542,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5715, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5715, "pid": 0, "tid": 7, "ts": 6950381147734.034, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113934.547, "dur": 1.492 , "args": { "External id": 1542,"cbid": 652, "correlation": 5715 } }, { "ph": "s", "id": 5715, "pid": 1728719, "tid": 1728719, "ts": 6950381113934.547, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381147747.794, "dur": 13.248 , "args": { "External id": 1543,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5721, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5721, "pid": 0, "tid": 7, "ts": 6950381147747.794, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113942.862, "dur": 1.513 , "args": { "External id": 1543,"cbid": 652, "correlation": 5721 } }, { "ph": "s", "id": 5721, "pid": 1728719, "tid": 1728719, "ts": 6950381113942.862, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113954.273, "dur": 0.070 , "args": { "External id": 1544,"cbid": 200, "correlation": 5740 } }, { "ph": "f", "id": 5740, "pid": 1728719, "tid": 1728719, "ts": 6950381113954.273, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113954.444, "dur": 0.110 , "args": { "External id": 1544,"cbid": 468, "correlation": 5741 } }, { "ph": "f", "id": 5741, "pid": 1728719, "tid": 1728719, "ts": 6950381113954.444, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113954.774, "dur": 0.020 , "args": { "External id": 1544,"cbid": 1, "correlation": 5743 } }, { "ph": "f", "id": 5743, "pid": 1728719, "tid": 1728719, "ts": 6950381113954.774, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113954.864, "dur": 0.091 , "args": { "External id": 1544,"cbid": 468, "correlation": 5744 } }, { "ph": "f", "id": 5744, "pid": 1728719, "tid": 1728719, "ts": 6950381113954.864, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113955.105, "dur": 0.040 , "args": { "External id": 1544,"cbid": 1, "correlation": 5746 } }, { "ph": "f", "id": 5746, "pid": 1728719, "tid": 1728719, "ts": 6950381113955.105, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113955.235, "dur": 0.100 , "args": { "External id": 1544,"cbid": 468, "correlation": 5747 } }, { "ph": "f", "id": 5747, "pid": 1728719, "tid": 1728719, "ts": 6950381113955.235, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113955.486, "dur": 0.030 , "args": { "External id": 1544,"cbid": 1, "correlation": 5749 } }, { "ph": "f", "id": 5749, "pid": 1728719, "tid": 1728719, "ts": 6950381113955.486, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113955.596, "dur": 0.380 , "args": { "External id": 1544,"cbid": 273, "correlation": 5750 } }, { "ph": "f", "id": 5750, "pid": 1728719, "tid": 1728719, "ts": 6950381113955.596, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381147761.362, "dur": 666.584 , "args": { "External id": 1544,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5753, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5753, "pid": 0, "tid": 7, "ts": 6950381147761.362, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113956.327, "dur": 1.693 , "args": { "External id": 1544,"cbid": 211, "correlation": 5753 } }, { "ph": "s", "id": 5753, "pid": 1728719, "tid": 1728719, "ts": 6950381113956.327, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113968.039, "dur": 0.080 , "args": { "External id": 1556,"cbid": 200, "correlation": 5777 } }, { "ph": "f", "id": 5777, "pid": 1728719, "tid": 1728719, "ts": 6950381113968.039, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113968.209, "dur": 0.110 , "args": { "External id": 1556,"cbid": 468, "correlation": 5778 } }, { "ph": "f", "id": 5778, "pid": 1728719, "tid": 1728719, "ts": 6950381113968.209, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113968.510, "dur": 0.030 , "args": { "External id": 1556,"cbid": 1, "correlation": 5780 } }, { "ph": "f", "id": 5780, "pid": 1728719, "tid": 1728719, "ts": 6950381113968.510, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113968.680, "dur": 0.100 , "args": { "External id": 1556,"cbid": 468, "correlation": 5781 } }, { "ph": "f", "id": 5781, "pid": 1728719, "tid": 1728719, "ts": 6950381113968.680, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113968.921, "dur": 0.020 , "args": { "External id": 1556,"cbid": 1, "correlation": 5783 } }, { "ph": "f", "id": 5783, "pid": 1728719, "tid": 1728719, "ts": 6950381113968.921, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113969.041, "dur": 0.100 , "args": { "External id": 1556,"cbid": 468, "correlation": 5784 } }, { "ph": "f", "id": 5784, "pid": 1728719, "tid": 1728719, "ts": 6950381113969.041, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113969.291, "dur": 0.020 , "args": { "External id": 1556,"cbid": 1, "correlation": 5786 } }, { "ph": "f", "id": 5786, "pid": 1728719, "tid": 1728719, "ts": 6950381113969.291, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113969.401, "dur": 0.331 , "args": { "External id": 1556,"cbid": 273, "correlation": 5787 } }, { "ph": "f", "id": 5787, "pid": 1728719, "tid": 1728719, "ts": 6950381113969.401, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381148428.234, "dur": 678.489 , "args": { "External id": 1556,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5790, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5790, "pid": 0, "tid": 7, "ts": 6950381148428.234, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113970.083, "dur": 1.643 , "args": { "External id": 1556,"cbid": 211, "correlation": 5790 } }, { "ph": "s", "id": 5790, "pid": 1728719, "tid": 1728719, "ts": 6950381113970.083, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381149107.043, "dur": 154.622 , "args": { "External id": 1568,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5803, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5803, "pid": 0, "tid": 7, "ts": 6950381149107.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381113979.120, "dur": 1.613 , "args": { "External id": 1568,"cbid": 307, "correlation": 5803 } }, { "ph": "s", "id": 5803, "pid": 1728719, "tid": 1728719, "ts": 6950381113979.120, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381149261.921, "dur": 57.984 , "args": { "External id": 1569,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5809, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5809, "pid": 0, "tid": 7, "ts": 6950381149261.921, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381113987.215, "dur": 1.533 , "args": { "External id": 1569,"cbid": 652, "correlation": 5809 } }, { "ph": "s", "id": 5809, "pid": 1728719, "tid": 1728719, "ts": 6950381113987.215, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113998.375, "dur": 0.091 , "args": { "External id": 1570,"cbid": 200, "correlation": 5828 } }, { "ph": "f", "id": 5828, "pid": 1728719, "tid": 1728719, "ts": 6950381113998.375, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113998.566, "dur": 0.120 , "args": { "External id": 1570,"cbid": 468, "correlation": 5829 } }, { "ph": "f", "id": 5829, "pid": 1728719, "tid": 1728719, "ts": 6950381113998.566, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113998.886, "dur": 0.020 , "args": { "External id": 1570,"cbid": 1, "correlation": 5831 } }, { "ph": "f", "id": 5831, "pid": 1728719, "tid": 1728719, "ts": 6950381113998.886, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113999.067, "dur": 0.090 , "args": { "External id": 1570,"cbid": 468, "correlation": 5832 } }, { "ph": "f", "id": 5832, "pid": 1728719, "tid": 1728719, "ts": 6950381113999.067, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113999.297, "dur": 0.030 , "args": { "External id": 1570,"cbid": 1, "correlation": 5834 } }, { "ph": "f", "id": 5834, "pid": 1728719, "tid": 1728719, "ts": 6950381113999.297, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113999.417, "dur": 0.091 , "args": { "External id": 1570,"cbid": 468, "correlation": 5835 } }, { "ph": "f", "id": 5835, "pid": 1728719, "tid": 1728719, "ts": 6950381113999.417, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381113999.668, "dur": 0.030 , "args": { "External id": 1570,"cbid": 1, "correlation": 5837 } }, { "ph": "f", "id": 5837, "pid": 1728719, "tid": 1728719, "ts": 6950381113999.668, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381113999.798, "dur": 0.371 , "args": { "External id": 1570,"cbid": 273, "correlation": 5838 } }, { "ph": "f", "id": 5838, "pid": 1728719, "tid": 1728719, "ts": 6950381113999.798, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381149320.193, "dur": 674.040 , "args": { "External id": 1570,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5841, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5841, "pid": 0, "tid": 7, "ts": 6950381149320.193, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114000.519, "dur": 1.724 , "args": { "External id": 1570,"cbid": 211, "correlation": 5841 } }, { "ph": "s", "id": 5841, "pid": 1728719, "tid": 1728719, "ts": 6950381114000.519, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381149994.553, "dur": 40.288 , "args": { "External id": 1582,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5852, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5852, "pid": 0, "tid": 7, "ts": 6950381149994.553, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114010.268, "dur": 1.603 , "args": { "External id": 1582,"cbid": 307, "correlation": 5852 } }, { "ph": "s", "id": 5852, "pid": 1728719, "tid": 1728719, "ts": 6950381114010.268, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381150035.097, "dur": 20.736 , "args": { "External id": 1594,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5864, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5864, "pid": 0, "tid": 7, "ts": 6950381150035.097, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114090.707, "dur": 2.956 , "args": { "External id": 1594,"cbid": 307, "correlation": 5864 } }, { "ph": "s", "id": 5864, "pid": 1728719, "tid": 1728719, "ts": 6950381114090.707, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381150056.153, "dur": 12.671 , "args": { "External id": 1595,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5878, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5878, "pid": 0, "tid": 7, "ts": 6950381150056.153, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114105.094, "dur": 1.753 , "args": { "External id": 1595,"cbid": 652, "correlation": 5878 } }, { "ph": "s", "id": 5878, "pid": 1728719, "tid": 1728719, "ts": 6950381114105.094, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381150069.112, "dur": 12.480 , "args": { "External id": 1596,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5892, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5892, "pid": 0, "tid": 7, "ts": 6950381150069.112, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114117.387, "dur": 1.533 , "args": { "External id": 1596,"cbid": 652, "correlation": 5892 } }, { "ph": "s", "id": 5892, "pid": 1728719, "tid": 1728719, "ts": 6950381114117.387, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381150081.880, "dur": 11.392 , "args": { "External id": 1597,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5906, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5906, "pid": 0, "tid": 7, "ts": 6950381150081.880, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114128.508, "dur": 1.522 , "args": { "External id": 1597,"cbid": 652, "correlation": 5906 } }, { "ph": "s", "id": 5906, "pid": 1728719, "tid": 1728719, "ts": 6950381114128.508, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114140.069, "dur": 0.120 , "args": { "External id": 1598,"cbid": 200, "correlation": 5921 } }, { "ph": "f", "id": 5921, "pid": 1728719, "tid": 1728719, "ts": 6950381114140.069, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114140.320, "dur": 0.180 , "args": { "External id": 1598,"cbid": 468, "correlation": 5922 } }, { "ph": "f", "id": 5922, "pid": 1728719, "tid": 1728719, "ts": 6950381114140.320, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114140.881, "dur": 0.050 , "args": { "External id": 1598,"cbid": 1, "correlation": 5924 } }, { "ph": "f", "id": 5924, "pid": 1728719, "tid": 1728719, "ts": 6950381114140.881, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114141.021, "dur": 0.110 , "args": { "External id": 1598,"cbid": 468, "correlation": 5925 } }, { "ph": "f", "id": 5925, "pid": 1728719, "tid": 1728719, "ts": 6950381114141.021, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114141.291, "dur": 0.020 , "args": { "External id": 1598,"cbid": 1, "correlation": 5927 } }, { "ph": "f", "id": 5927, "pid": 1728719, "tid": 1728719, "ts": 6950381114141.291, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114141.402, "dur": 0.100 , "args": { "External id": 1598,"cbid": 468, "correlation": 5928 } }, { "ph": "f", "id": 5928, "pid": 1728719, "tid": 1728719, "ts": 6950381114141.402, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114141.652, "dur": 0.030 , "args": { "External id": 1598,"cbid": 1, "correlation": 5930 } }, { "ph": "f", "id": 5930, "pid": 1728719, "tid": 1728719, "ts": 6950381114141.652, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114141.812, "dur": 0.712 , "args": { "External id": 1598,"cbid": 273, "correlation": 5931 } }, { "ph": "f", "id": 5931, "pid": 1728719, "tid": 1728719, "ts": 6950381114141.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381150093.560, "dur": 215.358 , "args": { "External id": 1598,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5934, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5934, "pid": 0, "tid": 7, "ts": 6950381150093.560, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114142.975, "dur": 1.903 , "args": { "External id": 1598,"cbid": 211, "correlation": 5934 } }, { "ph": "s", "id": 5934, "pid": 1728719, "tid": 1728719, "ts": 6950381114142.975, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381150309.238, "dur": 8.832 , "args": { "External id": 1610,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5947, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5947, "pid": 0, "tid": 7, "ts": 6950381150309.238, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114153.664, "dur": 1.764 , "args": { "External id": 1610,"cbid": 307, "correlation": 5947 } }, { "ph": "s", "id": 5947, "pid": 1728719, "tid": 1728719, "ts": 6950381114153.664, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381150318.358, "dur": 37.823 , "args": { "External id": 1615,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5970, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5970, "pid": 0, "tid": 7, "ts": 6950381150318.358, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114172.860, "dur": 1.713 , "args": { "External id": 1615,"cbid": 307, "correlation": 5970 } }, { "ph": "s", "id": 5970, "pid": 1728719, "tid": 1728719, "ts": 6950381114172.860, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114183.290, "dur": 0.100 , "args": { "External id": 1616,"cbid": 200, "correlation": 5985 } }, { "ph": "f", "id": 5985, "pid": 1728719, "tid": 1728719, "ts": 6950381114183.290, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114183.480, "dur": 0.130 , "args": { "External id": 1616,"cbid": 468, "correlation": 5986 } }, { "ph": "f", "id": 5986, "pid": 1728719, "tid": 1728719, "ts": 6950381114183.480, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114183.841, "dur": 0.020 , "args": { "External id": 1616,"cbid": 1, "correlation": 5988 } }, { "ph": "f", "id": 5988, "pid": 1728719, "tid": 1728719, "ts": 6950381114183.841, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114183.931, "dur": 0.100 , "args": { "External id": 1616,"cbid": 468, "correlation": 5989 } }, { "ph": "f", "id": 5989, "pid": 1728719, "tid": 1728719, "ts": 6950381114183.931, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114184.201, "dur": 0.020 , "args": { "External id": 1616,"cbid": 1, "correlation": 5991 } }, { "ph": "f", "id": 5991, "pid": 1728719, "tid": 1728719, "ts": 6950381114184.201, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114184.312, "dur": 0.090 , "args": { "External id": 1616,"cbid": 468, "correlation": 5992 } }, { "ph": "f", "id": 5992, "pid": 1728719, "tid": 1728719, "ts": 6950381114184.312, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114184.562, "dur": 0.020 , "args": { "External id": 1616,"cbid": 1, "correlation": 5994 } }, { "ph": "f", "id": 5994, "pid": 1728719, "tid": 1728719, "ts": 6950381114184.562, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114184.682, "dur": 0.391 , "args": { "External id": 1616,"cbid": 273, "correlation": 5995 } }, { "ph": "f", "id": 5995, "pid": 1728719, "tid": 1728719, "ts": 6950381114184.682, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381150356.469, "dur": 226.878 , "args": { "External id": 1616,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 5998, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 5998, "pid": 0, "tid": 7, "ts": 6950381150356.469, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114185.474, "dur": 1.713 , "args": { "External id": 1616,"cbid": 211, "correlation": 5998 } }, { "ph": "s", "id": 5998, "pid": 1728719, "tid": 1728719, "ts": 6950381114185.474, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381150583.635, "dur": 8.832 , "args": { "External id": 1628,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6007, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6007, "pid": 0, "tid": 7, "ts": 6950381150583.635, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114193.068, "dur": 1.633 , "args": { "External id": 1628,"cbid": 307, "correlation": 6007 } }, { "ph": "s", "id": 6007, "pid": 1728719, "tid": 1728719, "ts": 6950381114193.068, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381150592.755, "dur": 26.176 , "args": { "External id": 1633,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6030, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6030, "pid": 0, "tid": 7, "ts": 6950381150592.755, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114209.128, "dur": 1.673 , "args": { "External id": 1633,"cbid": 307, "correlation": 6030 } }, { "ph": "s", "id": 6030, "pid": 1728719, "tid": 1728719, "ts": 6950381114209.128, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114219.196, "dur": 0.101 , "args": { "External id": 1634,"cbid": 200, "correlation": 6045 } }, { "ph": "f", "id": 6045, "pid": 1728719, "tid": 1728719, "ts": 6950381114219.196, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114219.397, "dur": 0.140 , "args": { "External id": 1634,"cbid": 468, "correlation": 6046 } }, { "ph": "f", "id": 6046, "pid": 1728719, "tid": 1728719, "ts": 6950381114219.397, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114219.747, "dur": 0.021 , "args": { "External id": 1634,"cbid": 1, "correlation": 6048 } }, { "ph": "f", "id": 6048, "pid": 1728719, "tid": 1728719, "ts": 6950381114219.747, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114219.838, "dur": 0.100 , "args": { "External id": 1634,"cbid": 468, "correlation": 6049 } }, { "ph": "f", "id": 6049, "pid": 1728719, "tid": 1728719, "ts": 6950381114219.838, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114220.078, "dur": 0.030 , "args": { "External id": 1634,"cbid": 1, "correlation": 6051 } }, { "ph": "f", "id": 6051, "pid": 1728719, "tid": 1728719, "ts": 6950381114220.078, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114220.198, "dur": 0.101 , "args": { "External id": 1634,"cbid": 468, "correlation": 6052 } }, { "ph": "f", "id": 6052, "pid": 1728719, "tid": 1728719, "ts": 6950381114220.198, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114220.439, "dur": 0.030 , "args": { "External id": 1634,"cbid": 1, "correlation": 6054 } }, { "ph": "f", "id": 6054, "pid": 1728719, "tid": 1728719, "ts": 6950381114220.439, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114220.549, "dur": 0.391 , "args": { "External id": 1634,"cbid": 273, "correlation": 6055 } }, { "ph": "f", "id": 6055, "pid": 1728719, "tid": 1728719, "ts": 6950381114220.549, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381150619.187, "dur": 231.165 , "args": { "External id": 1634,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6058, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6058, "pid": 0, "tid": 7, "ts": 6950381150619.187, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114221.340, "dur": 1.744 , "args": { "External id": 1634,"cbid": 211, "correlation": 6058 } }, { "ph": "s", "id": 6058, "pid": 1728719, "tid": 1728719, "ts": 6950381114221.340, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381150850.640, "dur": 60.511 , "args": { "External id": 1646,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6072, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6072, "pid": 0, "tid": 7, "ts": 6950381150850.640, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114231.119, "dur": 1.623 , "args": { "External id": 1646,"cbid": 307, "correlation": 6072 } }, { "ph": "s", "id": 6072, "pid": 1728719, "tid": 1728719, "ts": 6950381114231.119, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381150911.407, "dur": 62.432 , "args": { "External id": 1647,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6082, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6082, "pid": 0, "tid": 7, "ts": 6950381150911.407, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114241.348, "dur": 1.703 , "args": { "External id": 1647,"cbid": 307, "correlation": 6082 } }, { "ph": "s", "id": 6082, "pid": 1728719, "tid": 1728719, "ts": 6950381114241.348, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381150974.127, "dur": 38.911 , "args": { "External id": 1648,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6091, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6091, "pid": 0, "tid": 7, "ts": 6950381150974.127, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114250.845, "dur": 1.563 , "args": { "External id": 1648,"cbid": 307, "correlation": 6091 } }, { "ph": "s", "id": 6091, "pid": 1728719, "tid": 1728719, "ts": 6950381114250.845, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114266.444, "dur": 0.131 , "args": { "External id": 1656,"cbid": 200, "correlation": 6126 } }, { "ph": "f", "id": 6126, "pid": 1728719, "tid": 1728719, "ts": 6950381114266.444, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114266.655, "dur": 0.070 , "args": { "External id": 1656,"cbid": 200, "correlation": 6127 } }, { "ph": "f", "id": 6127, "pid": 1728719, "tid": 1728719, "ts": 6950381114266.655, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114266.825, "dur": 0.481 , "args": { "External id": 1656,"cbid": 273, "correlation": 6128 } }, { "ph": "f", "id": 6128, "pid": 1728719, "tid": 1728719, "ts": 6950381114266.825, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381151013.326, "dur": 846.199 , "args": { "External id": 1656,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6130, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6130, "pid": 0, "tid": 7, "ts": 6950381151013.326, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114267.597, "dur": 1.953 , "args": { "External id": 1656,"cbid": 211, "correlation": 6130 } }, { "ph": "s", "id": 6130, "pid": 1728719, "tid": 1728719, "ts": 6950381114267.597, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381151859.813, "dur": 9.472 , "args": { "External id": 1665,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6150, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6150, "pid": 0, "tid": 7, "ts": 6950381151859.813, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114283.907, "dur": 1.523 , "args": { "External id": 1665,"cbid": 652, "correlation": 6150 } }, { "ph": "s", "id": 6150, "pid": 1728719, "tid": 1728719, "ts": 6950381114283.907, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114294.537, "dur": 0.110 , "args": { "External id": 1666,"cbid": 200, "correlation": 6165 } }, { "ph": "f", "id": 6165, "pid": 1728719, "tid": 1728719, "ts": 6950381114294.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114294.737, "dur": 0.150 , "args": { "External id": 1666,"cbid": 468, "correlation": 6166 } }, { "ph": "f", "id": 6166, "pid": 1728719, "tid": 1728719, "ts": 6950381114294.737, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114295.168, "dur": 0.020 , "args": { "External id": 1666,"cbid": 1, "correlation": 6168 } }, { "ph": "f", "id": 6168, "pid": 1728719, "tid": 1728719, "ts": 6950381114295.168, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114295.258, "dur": 0.100 , "args": { "External id": 1666,"cbid": 468, "correlation": 6169 } }, { "ph": "f", "id": 6169, "pid": 1728719, "tid": 1728719, "ts": 6950381114295.258, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114295.509, "dur": 0.020 , "args": { "External id": 1666,"cbid": 1, "correlation": 6171 } }, { "ph": "f", "id": 6171, "pid": 1728719, "tid": 1728719, "ts": 6950381114295.509, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114295.619, "dur": 0.100 , "args": { "External id": 1666,"cbid": 468, "correlation": 6172 } }, { "ph": "f", "id": 6172, "pid": 1728719, "tid": 1728719, "ts": 6950381114295.619, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114295.879, "dur": 0.020 , "args": { "External id": 1666,"cbid": 1, "correlation": 6174 } }, { "ph": "f", "id": 6174, "pid": 1728719, "tid": 1728719, "ts": 6950381114295.879, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114295.979, "dur": 0.401 , "args": { "External id": 1666,"cbid": 273, "correlation": 6175 } }, { "ph": "f", "id": 6175, "pid": 1728719, "tid": 1728719, "ts": 6950381114295.979, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381151869.573, "dur": 207.774 , "args": { "External id": 1666,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6178, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6178, "pid": 0, "tid": 7, "ts": 6950381151869.573, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114296.751, "dur": 1.803 , "args": { "External id": 1666,"cbid": 211, "correlation": 6178 } }, { "ph": "s", "id": 6178, "pid": 1728719, "tid": 1728719, "ts": 6950381114296.751, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381152077.635, "dur": 46.079 , "args": { "External id": 1678,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6190, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6190, "pid": 0, "tid": 7, "ts": 6950381152077.635, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114306.319, "dur": 1.693 , "args": { "External id": 1678,"cbid": 307, "correlation": 6190 } }, { "ph": "s", "id": 6190, "pid": 1728719, "tid": 1728719, "ts": 6950381114306.319, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381152124.002, "dur": 13.280 , "args": { "External id": 1679,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6196, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6196, "pid": 0, "tid": 7, "ts": 6950381152124.002, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114314.163, "dur": 1.513 , "args": { "External id": 1679,"cbid": 652, "correlation": 6196 } }, { "ph": "s", "id": 6196, "pid": 1728719, "tid": 1728719, "ts": 6950381114314.163, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381152137.602, "dur": 12.864 , "args": { "External id": 1680,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6202, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6202, "pid": 0, "tid": 7, "ts": 6950381152137.602, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114322.268, "dur": 1.543 , "args": { "External id": 1680,"cbid": 652, "correlation": 6202 } }, { "ph": "s", "id": 6202, "pid": 1728719, "tid": 1728719, "ts": 6950381114322.268, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114333.830, "dur": 0.100 , "args": { "External id": 1681,"cbid": 200, "correlation": 6221 } }, { "ph": "f", "id": 6221, "pid": 1728719, "tid": 1728719, "ts": 6950381114333.830, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114334.020, "dur": 0.120 , "args": { "External id": 1681,"cbid": 468, "correlation": 6222 } }, { "ph": "f", "id": 6222, "pid": 1728719, "tid": 1728719, "ts": 6950381114334.020, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114334.351, "dur": 0.030 , "args": { "External id": 1681,"cbid": 1, "correlation": 6224 } }, { "ph": "f", "id": 6224, "pid": 1728719, "tid": 1728719, "ts": 6950381114334.351, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114334.451, "dur": 0.100 , "args": { "External id": 1681,"cbid": 468, "correlation": 6225 } }, { "ph": "f", "id": 6225, "pid": 1728719, "tid": 1728719, "ts": 6950381114334.451, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114334.701, "dur": 0.021 , "args": { "External id": 1681,"cbid": 1, "correlation": 6227 } }, { "ph": "f", "id": 6227, "pid": 1728719, "tid": 1728719, "ts": 6950381114334.701, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114334.812, "dur": 0.090 , "args": { "External id": 1681,"cbid": 468, "correlation": 6228 } }, { "ph": "f", "id": 6228, "pid": 1728719, "tid": 1728719, "ts": 6950381114334.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114335.062, "dur": 0.020 , "args": { "External id": 1681,"cbid": 1, "correlation": 6230 } }, { "ph": "f", "id": 6230, "pid": 1728719, "tid": 1728719, "ts": 6950381114335.062, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114335.162, "dur": 0.391 , "args": { "External id": 1681,"cbid": 273, "correlation": 6231 } }, { "ph": "f", "id": 6231, "pid": 1728719, "tid": 1728719, "ts": 6950381114335.162, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381152150.754, "dur": 658.201 , "args": { "External id": 1681,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6234, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6234, "pid": 0, "tid": 7, "ts": 6950381152150.754, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114335.914, "dur": 1.713 , "args": { "External id": 1681,"cbid": 211, "correlation": 6234 } }, { "ph": "s", "id": 6234, "pid": 1728719, "tid": 1728719, "ts": 6950381114335.914, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114347.616, "dur": 0.090 , "args": { "External id": 1693,"cbid": 200, "correlation": 6258 } }, { "ph": "f", "id": 6258, "pid": 1728719, "tid": 1728719, "ts": 6950381114347.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114347.786, "dur": 0.120 , "args": { "External id": 1693,"cbid": 468, "correlation": 6259 } }, { "ph": "f", "id": 6259, "pid": 1728719, "tid": 1728719, "ts": 6950381114347.786, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114348.086, "dur": 0.020 , "args": { "External id": 1693,"cbid": 1, "correlation": 6261 } }, { "ph": "f", "id": 6261, "pid": 1728719, "tid": 1728719, "ts": 6950381114348.086, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114348.177, "dur": 0.100 , "args": { "External id": 1693,"cbid": 468, "correlation": 6262 } }, { "ph": "f", "id": 6262, "pid": 1728719, "tid": 1728719, "ts": 6950381114348.177, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114348.417, "dur": 0.020 , "args": { "External id": 1693,"cbid": 1, "correlation": 6264 } }, { "ph": "f", "id": 6264, "pid": 1728719, "tid": 1728719, "ts": 6950381114348.417, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114348.527, "dur": 0.090 , "args": { "External id": 1693,"cbid": 468, "correlation": 6265 } }, { "ph": "f", "id": 6265, "pid": 1728719, "tid": 1728719, "ts": 6950381114348.527, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114348.768, "dur": 0.020 , "args": { "External id": 1693,"cbid": 1, "correlation": 6267 } }, { "ph": "f", "id": 6267, "pid": 1728719, "tid": 1728719, "ts": 6950381114348.768, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114348.878, "dur": 0.330 , "args": { "External id": 1693,"cbid": 273, "correlation": 6268 } }, { "ph": "f", "id": 6268, "pid": 1728719, "tid": 1728719, "ts": 6950381114348.878, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381152809.275, "dur": 676.057 , "args": { "External id": 1693,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6271, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6271, "pid": 0, "tid": 7, "ts": 6950381152809.275, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114349.559, "dur": 1.613 , "args": { "External id": 1693,"cbid": 211, "correlation": 6271 } }, { "ph": "s", "id": 6271, "pid": 1728719, "tid": 1728719, "ts": 6950381114349.559, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381153485.620, "dur": 153.566 , "args": { "External id": 1705,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6284, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6284, "pid": 0, "tid": 7, "ts": 6950381153485.620, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114358.636, "dur": 1.653 , "args": { "External id": 1705,"cbid": 307, "correlation": 6284 } }, { "ph": "s", "id": 6284, "pid": 1728719, "tid": 1728719, "ts": 6950381114358.636, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381153639.442, "dur": 58.335 , "args": { "External id": 1706,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6290, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6290, "pid": 0, "tid": 7, "ts": 6950381153639.442, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114367.052, "dur": 1.643 , "args": { "External id": 1706,"cbid": 652, "correlation": 6290 } }, { "ph": "s", "id": 6290, "pid": 1728719, "tid": 1728719, "ts": 6950381114367.052, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114378.633, "dur": 0.080 , "args": { "External id": 1707,"cbid": 200, "correlation": 6309 } }, { "ph": "f", "id": 6309, "pid": 1728719, "tid": 1728719, "ts": 6950381114378.633, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114378.804, "dur": 0.110 , "args": { "External id": 1707,"cbid": 468, "correlation": 6310 } }, { "ph": "f", "id": 6310, "pid": 1728719, "tid": 1728719, "ts": 6950381114378.804, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114379.124, "dur": 0.020 , "args": { "External id": 1707,"cbid": 1, "correlation": 6312 } }, { "ph": "f", "id": 6312, "pid": 1728719, "tid": 1728719, "ts": 6950381114379.124, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114379.284, "dur": 0.091 , "args": { "External id": 1707,"cbid": 468, "correlation": 6313 } }, { "ph": "f", "id": 6313, "pid": 1728719, "tid": 1728719, "ts": 6950381114379.284, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114379.515, "dur": 0.020 , "args": { "External id": 1707,"cbid": 1, "correlation": 6315 } }, { "ph": "f", "id": 6315, "pid": 1728719, "tid": 1728719, "ts": 6950381114379.515, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114379.645, "dur": 0.090 , "args": { "External id": 1707,"cbid": 468, "correlation": 6316 } }, { "ph": "f", "id": 6316, "pid": 1728719, "tid": 1728719, "ts": 6950381114379.645, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114379.896, "dur": 0.020 , "args": { "External id": 1707,"cbid": 1, "correlation": 6318 } }, { "ph": "f", "id": 6318, "pid": 1728719, "tid": 1728719, "ts": 6950381114379.896, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114380.016, "dur": 0.350 , "args": { "External id": 1707,"cbid": 273, "correlation": 6319 } }, { "ph": "f", "id": 6319, "pid": 1728719, "tid": 1728719, "ts": 6950381114380.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381153698.065, "dur": 680.249 , "args": { "External id": 1707,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6322, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6322, "pid": 0, "tid": 7, "ts": 6950381153698.065, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114380.737, "dur": 1.874 , "args": { "External id": 1707,"cbid": 211, "correlation": 6322 } }, { "ph": "s", "id": 6322, "pid": 1728719, "tid": 1728719, "ts": 6950381114380.737, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381154378.602, "dur": 40.288 , "args": { "External id": 1719,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6333, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6333, "pid": 0, "tid": 7, "ts": 6950381154378.602, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114390.896, "dur": 1.663 , "args": { "External id": 1719,"cbid": 307, "correlation": 6333 } }, { "ph": "s", "id": 6333, "pid": 1728719, "tid": 1728719, "ts": 6950381114390.896, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381154419.146, "dur": 21.055 , "args": { "External id": 1731,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6345, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6345, "pid": 0, "tid": 7, "ts": 6950381154419.146, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114470.635, "dur": 2.644 , "args": { "External id": 1731,"cbid": 307, "correlation": 6345 } }, { "ph": "s", "id": 6345, "pid": 1728719, "tid": 1728719, "ts": 6950381114470.635, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381154440.521, "dur": 12.992 , "args": { "External id": 1732,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6359, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6359, "pid": 0, "tid": 7, "ts": 6950381154440.521, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114485.021, "dur": 1.764 , "args": { "External id": 1732,"cbid": 652, "correlation": 6359 } }, { "ph": "s", "id": 6359, "pid": 1728719, "tid": 1728719, "ts": 6950381114485.021, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381154453.833, "dur": 12.512 , "args": { "External id": 1733,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6373, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6373, "pid": 0, "tid": 7, "ts": 6950381154453.833, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114496.773, "dur": 1.533 , "args": { "External id": 1733,"cbid": 652, "correlation": 6373 } }, { "ph": "s", "id": 6373, "pid": 1728719, "tid": 1728719, "ts": 6950381114496.773, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381154466.633, "dur": 11.136 , "args": { "External id": 1734,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6387, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6387, "pid": 0, "tid": 7, "ts": 6950381154466.633, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114507.503, "dur": 1.523 , "args": { "External id": 1734,"cbid": 652, "correlation": 6387 } }, { "ph": "s", "id": 6387, "pid": 1728719, "tid": 1728719, "ts": 6950381114507.503, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114519.025, "dur": 0.130 , "args": { "External id": 1735,"cbid": 200, "correlation": 6402 } }, { "ph": "f", "id": 6402, "pid": 1728719, "tid": 1728719, "ts": 6950381114519.025, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114519.315, "dur": 0.180 , "args": { "External id": 1735,"cbid": 468, "correlation": 6403 } }, { "ph": "f", "id": 6403, "pid": 1728719, "tid": 1728719, "ts": 6950381114519.315, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114519.846, "dur": 0.030 , "args": { "External id": 1735,"cbid": 1, "correlation": 6405 } }, { "ph": "f", "id": 6405, "pid": 1728719, "tid": 1728719, "ts": 6950381114519.846, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114519.946, "dur": 0.100 , "args": { "External id": 1735,"cbid": 468, "correlation": 6406 } }, { "ph": "f", "id": 6406, "pid": 1728719, "tid": 1728719, "ts": 6950381114519.946, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114520.197, "dur": 0.020 , "args": { "External id": 1735,"cbid": 1, "correlation": 6408 } }, { "ph": "f", "id": 6408, "pid": 1728719, "tid": 1728719, "ts": 6950381114520.197, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114520.317, "dur": 0.100 , "args": { "External id": 1735,"cbid": 468, "correlation": 6409 } }, { "ph": "f", "id": 6409, "pid": 1728719, "tid": 1728719, "ts": 6950381114520.317, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114520.577, "dur": 0.021 , "args": { "External id": 1735,"cbid": 1, "correlation": 6411 } }, { "ph": "f", "id": 6411, "pid": 1728719, "tid": 1728719, "ts": 6950381114520.577, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114520.718, "dur": 0.741 , "args": { "External id": 1735,"cbid": 273, "correlation": 6412 } }, { "ph": "f", "id": 6412, "pid": 1728719, "tid": 1728719, "ts": 6950381114520.718, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381154478.057, "dur": 216.477 , "args": { "External id": 1735,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6415, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6415, "pid": 0, "tid": 7, "ts": 6950381154478.057, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114521.910, "dur": 1.984 , "args": { "External id": 1735,"cbid": 211, "correlation": 6415 } }, { "ph": "s", "id": 6415, "pid": 1728719, "tid": 1728719, "ts": 6950381114521.910, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381154694.855, "dur": 8.831 , "args": { "External id": 1747,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6428, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6428, "pid": 0, "tid": 7, "ts": 6950381154694.855, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114533.471, "dur": 1.634 , "args": { "External id": 1747,"cbid": 307, "correlation": 6428 } }, { "ph": "s", "id": 6428, "pid": 1728719, "tid": 1728719, "ts": 6950381114533.471, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381154703.974, "dur": 39.040 , "args": { "External id": 1752,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6451, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6451, "pid": 0, "tid": 7, "ts": 6950381154703.974, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114554.881, "dur": 1.663 , "args": { "External id": 1752,"cbid": 307, "correlation": 6451 } }, { "ph": "s", "id": 6451, "pid": 1728719, "tid": 1728719, "ts": 6950381114554.881, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114565.211, "dur": 0.110 , "args": { "External id": 1753,"cbid": 200, "correlation": 6466 } }, { "ph": "f", "id": 6466, "pid": 1728719, "tid": 1728719, "ts": 6950381114565.211, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114565.441, "dur": 0.130 , "args": { "External id": 1753,"cbid": 468, "correlation": 6467 } }, { "ph": "f", "id": 6467, "pid": 1728719, "tid": 1728719, "ts": 6950381114565.441, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114565.822, "dur": 0.040 , "args": { "External id": 1753,"cbid": 1, "correlation": 6469 } }, { "ph": "f", "id": 6469, "pid": 1728719, "tid": 1728719, "ts": 6950381114565.822, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114565.942, "dur": 0.100 , "args": { "External id": 1753,"cbid": 468, "correlation": 6470 } }, { "ph": "f", "id": 6470, "pid": 1728719, "tid": 1728719, "ts": 6950381114565.942, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114566.182, "dur": 0.030 , "args": { "External id": 1753,"cbid": 1, "correlation": 6472 } }, { "ph": "f", "id": 6472, "pid": 1728719, "tid": 1728719, "ts": 6950381114566.182, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114566.313, "dur": 0.100 , "args": { "External id": 1753,"cbid": 468, "correlation": 6473 } }, { "ph": "f", "id": 6473, "pid": 1728719, "tid": 1728719, "ts": 6950381114566.313, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114566.573, "dur": 0.020 , "args": { "External id": 1753,"cbid": 1, "correlation": 6475 } }, { "ph": "f", "id": 6475, "pid": 1728719, "tid": 1728719, "ts": 6950381114566.573, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114566.683, "dur": 0.451 , "args": { "External id": 1753,"cbid": 273, "correlation": 6476 } }, { "ph": "f", "id": 6476, "pid": 1728719, "tid": 1728719, "ts": 6950381114566.683, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381154743.302, "dur": 226.366 , "args": { "External id": 1753,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6479, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6479, "pid": 0, "tid": 7, "ts": 6950381154743.302, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114567.515, "dur": 1.843 , "args": { "External id": 1753,"cbid": 211, "correlation": 6479 } }, { "ph": "s", "id": 6479, "pid": 1728719, "tid": 1728719, "ts": 6950381114567.515, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381154969.956, "dur": 8.799 , "args": { "External id": 1765,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6488, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6488, "pid": 0, "tid": 7, "ts": 6950381154969.956, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114575.219, "dur": 1.563 , "args": { "External id": 1765,"cbid": 307, "correlation": 6488 } }, { "ph": "s", "id": 6488, "pid": 1728719, "tid": 1728719, "ts": 6950381114575.219, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381154979.043, "dur": 27.232 , "args": { "External id": 1770,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6511, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6511, "pid": 0, "tid": 7, "ts": 6950381154979.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114590.938, "dur": 1.543 , "args": { "External id": 1770,"cbid": 307, "correlation": 6511 } }, { "ph": "s", "id": 6511, "pid": 1728719, "tid": 1728719, "ts": 6950381114590.938, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114601.408, "dur": 0.090 , "args": { "External id": 1771,"cbid": 200, "correlation": 6526 } }, { "ph": "f", "id": 6526, "pid": 1728719, "tid": 1728719, "ts": 6950381114601.408, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114601.578, "dur": 0.130 , "args": { "External id": 1771,"cbid": 468, "correlation": 6527 } }, { "ph": "f", "id": 6527, "pid": 1728719, "tid": 1728719, "ts": 6950381114601.578, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114601.919, "dur": 0.020 , "args": { "External id": 1771,"cbid": 1, "correlation": 6529 } }, { "ph": "f", "id": 6529, "pid": 1728719, "tid": 1728719, "ts": 6950381114601.919, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114602.009, "dur": 0.100 , "args": { "External id": 1771,"cbid": 468, "correlation": 6530 } }, { "ph": "f", "id": 6530, "pid": 1728719, "tid": 1728719, "ts": 6950381114602.009, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114602.260, "dur": 0.020 , "args": { "External id": 1771,"cbid": 1, "correlation": 6532 } }, { "ph": "f", "id": 6532, "pid": 1728719, "tid": 1728719, "ts": 6950381114602.260, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114602.370, "dur": 0.100 , "args": { "External id": 1771,"cbid": 468, "correlation": 6533 } }, { "ph": "f", "id": 6533, "pid": 1728719, "tid": 1728719, "ts": 6950381114602.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114602.620, "dur": 0.020 , "args": { "External id": 1771,"cbid": 1, "correlation": 6535 } }, { "ph": "f", "id": 6535, "pid": 1728719, "tid": 1728719, "ts": 6950381114602.620, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114602.740, "dur": 0.381 , "args": { "External id": 1771,"cbid": 273, "correlation": 6536 } }, { "ph": "f", "id": 6536, "pid": 1728719, "tid": 1728719, "ts": 6950381114602.740, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381155006.563, "dur": 230.590 , "args": { "External id": 1771,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6539, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6539, "pid": 0, "tid": 7, "ts": 6950381155006.563, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114603.472, "dur": 1.693 , "args": { "External id": 1771,"cbid": 211, "correlation": 6539 } }, { "ph": "s", "id": 6539, "pid": 1728719, "tid": 1728719, "ts": 6950381114603.472, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381155237.441, "dur": 59.135 , "args": { "External id": 1783,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6553, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6553, "pid": 0, "tid": 7, "ts": 6950381155237.441, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114612.879, "dur": 1.774 , "args": { "External id": 1783,"cbid": 307, "correlation": 6553 } }, { "ph": "s", "id": 6553, "pid": 1728719, "tid": 1728719, "ts": 6950381114612.879, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381155296.864, "dur": 65.215 , "args": { "External id": 1784,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6563, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6563, "pid": 0, "tid": 7, "ts": 6950381155296.864, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114623.559, "dur": 1.673 , "args": { "External id": 1784,"cbid": 307, "correlation": 6563 } }, { "ph": "s", "id": 6563, "pid": 1728719, "tid": 1728719, "ts": 6950381114623.559, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381155362.367, "dur": 39.584 , "args": { "External id": 1785,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6572, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6572, "pid": 0, "tid": 7, "ts": 6950381155362.367, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114632.836, "dur": 1.623 , "args": { "External id": 1785,"cbid": 307, "correlation": 6572 } }, { "ph": "s", "id": 6572, "pid": 1728719, "tid": 1728719, "ts": 6950381114632.836, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114648.055, "dur": 0.120 , "args": { "External id": 1793,"cbid": 200, "correlation": 6607 } }, { "ph": "f", "id": 6607, "pid": 1728719, "tid": 1728719, "ts": 6950381114648.055, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114648.245, "dur": 0.080 , "args": { "External id": 1793,"cbid": 200, "correlation": 6608 } }, { "ph": "f", "id": 6608, "pid": 1728719, "tid": 1728719, "ts": 6950381114648.245, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114648.415, "dur": 0.461 , "args": { "External id": 1793,"cbid": 273, "correlation": 6609 } }, { "ph": "f", "id": 6609, "pid": 1728719, "tid": 1728719, "ts": 6950381114648.415, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381155402.239, "dur": 850.103 , "args": { "External id": 1793,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6611, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6611, "pid": 0, "tid": 7, "ts": 6950381155402.239, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114649.207, "dur": 1.813 , "args": { "External id": 1793,"cbid": 211, "correlation": 6611 } }, { "ph": "s", "id": 6611, "pid": 1728719, "tid": 1728719, "ts": 6950381114649.207, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381156252.598, "dur": 9.568 , "args": { "External id": 1802,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6631, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6631, "pid": 0, "tid": 7, "ts": 6950381156252.598, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114665.708, "dur": 1.593 , "args": { "External id": 1802,"cbid": 652, "correlation": 6631 } }, { "ph": "s", "id": 6631, "pid": 1728719, "tid": 1728719, "ts": 6950381114665.708, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114676.688, "dur": 0.110 , "args": { "External id": 1803,"cbid": 200, "correlation": 6646 } }, { "ph": "f", "id": 6646, "pid": 1728719, "tid": 1728719, "ts": 6950381114676.688, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114676.898, "dur": 0.171 , "args": { "External id": 1803,"cbid": 468, "correlation": 6647 } }, { "ph": "f", "id": 6647, "pid": 1728719, "tid": 1728719, "ts": 6950381114676.898, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114677.329, "dur": 0.020 , "args": { "External id": 1803,"cbid": 1, "correlation": 6649 } }, { "ph": "f", "id": 6649, "pid": 1728719, "tid": 1728719, "ts": 6950381114677.329, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114677.429, "dur": 0.111 , "args": { "External id": 1803,"cbid": 468, "correlation": 6650 } }, { "ph": "f", "id": 6650, "pid": 1728719, "tid": 1728719, "ts": 6950381114677.429, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114677.700, "dur": 0.020 , "args": { "External id": 1803,"cbid": 1, "correlation": 6652 } }, { "ph": "f", "id": 6652, "pid": 1728719, "tid": 1728719, "ts": 6950381114677.700, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114677.810, "dur": 0.100 , "args": { "External id": 1803,"cbid": 468, "correlation": 6653 } }, { "ph": "f", "id": 6653, "pid": 1728719, "tid": 1728719, "ts": 6950381114677.810, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114678.091, "dur": 0.020 , "args": { "External id": 1803,"cbid": 1, "correlation": 6655 } }, { "ph": "f", "id": 6655, "pid": 1728719, "tid": 1728719, "ts": 6950381114678.091, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114678.191, "dur": 0.401 , "args": { "External id": 1803,"cbid": 273, "correlation": 6656 } }, { "ph": "f", "id": 6656, "pid": 1728719, "tid": 1728719, "ts": 6950381114678.191, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381156262.422, "dur": 209.181 , "args": { "External id": 1803,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6659, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6659, "pid": 0, "tid": 7, "ts": 6950381156262.422, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114678.962, "dur": 1.774 , "args": { "External id": 1803,"cbid": 211, "correlation": 6659 } }, { "ph": "s", "id": 6659, "pid": 1728719, "tid": 1728719, "ts": 6950381114678.962, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381156471.891, "dur": 46.784 , "args": { "External id": 1815,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6671, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6671, "pid": 0, "tid": 7, "ts": 6950381156471.891, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114688.851, "dur": 1.673 , "args": { "External id": 1815,"cbid": 307, "correlation": 6671 } }, { "ph": "s", "id": 6671, "pid": 1728719, "tid": 1728719, "ts": 6950381114688.851, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381156518.995, "dur": 13.183 , "args": { "External id": 1816,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6677, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6677, "pid": 0, "tid": 7, "ts": 6950381156518.995, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114696.866, "dur": 1.522 , "args": { "External id": 1816,"cbid": 652, "correlation": 6677 } }, { "ph": "s", "id": 6677, "pid": 1728719, "tid": 1728719, "ts": 6950381114696.866, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381156532.466, "dur": 13.024 , "args": { "External id": 1817,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6683, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6683, "pid": 0, "tid": 7, "ts": 6950381156532.466, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114705.291, "dur": 1.433 , "args": { "External id": 1817,"cbid": 652, "correlation": 6683 } }, { "ph": "s", "id": 6683, "pid": 1728719, "tid": 1728719, "ts": 6950381114705.291, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114716.783, "dur": 0.120 , "args": { "External id": 1818,"cbid": 200, "correlation": 6702 } }, { "ph": "f", "id": 6702, "pid": 1728719, "tid": 1728719, "ts": 6950381114716.783, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114717.013, "dur": 0.120 , "args": { "External id": 1818,"cbid": 468, "correlation": 6703 } }, { "ph": "f", "id": 6703, "pid": 1728719, "tid": 1728719, "ts": 6950381114717.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114717.354, "dur": 0.030 , "args": { "External id": 1818,"cbid": 1, "correlation": 6705 } }, { "ph": "f", "id": 6705, "pid": 1728719, "tid": 1728719, "ts": 6950381114717.354, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114717.524, "dur": 0.100 , "args": { "External id": 1818,"cbid": 468, "correlation": 6706 } }, { "ph": "f", "id": 6706, "pid": 1728719, "tid": 1728719, "ts": 6950381114717.524, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114717.775, "dur": 0.020 , "args": { "External id": 1818,"cbid": 1, "correlation": 6708 } }, { "ph": "f", "id": 6708, "pid": 1728719, "tid": 1728719, "ts": 6950381114717.775, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114717.885, "dur": 0.100 , "args": { "External id": 1818,"cbid": 468, "correlation": 6709 } }, { "ph": "f", "id": 6709, "pid": 1728719, "tid": 1728719, "ts": 6950381114717.885, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114718.145, "dur": 0.020 , "args": { "External id": 1818,"cbid": 1, "correlation": 6711 } }, { "ph": "f", "id": 6711, "pid": 1728719, "tid": 1728719, "ts": 6950381114718.145, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114718.245, "dur": 0.371 , "args": { "External id": 1818,"cbid": 273, "correlation": 6712 } }, { "ph": "f", "id": 6712, "pid": 1728719, "tid": 1728719, "ts": 6950381114718.245, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381156545.778, "dur": 660.441 , "args": { "External id": 1818,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6715, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6715, "pid": 0, "tid": 7, "ts": 6950381156545.778, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114718.967, "dur": 1.703 , "args": { "External id": 1818,"cbid": 211, "correlation": 6715 } }, { "ph": "s", "id": 6715, "pid": 1728719, "tid": 1728719, "ts": 6950381114718.967, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114730.959, "dur": 0.090 , "args": { "External id": 1830,"cbid": 200, "correlation": 6739 } }, { "ph": "f", "id": 6739, "pid": 1728719, "tid": 1728719, "ts": 6950381114730.959, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114731.149, "dur": 0.111 , "args": { "External id": 1830,"cbid": 468, "correlation": 6740 } }, { "ph": "f", "id": 6740, "pid": 1728719, "tid": 1728719, "ts": 6950381114731.149, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114731.460, "dur": 0.020 , "args": { "External id": 1830,"cbid": 1, "correlation": 6742 } }, { "ph": "f", "id": 6742, "pid": 1728719, "tid": 1728719, "ts": 6950381114731.460, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114731.550, "dur": 0.100 , "args": { "External id": 1830,"cbid": 468, "correlation": 6743 } }, { "ph": "f", "id": 6743, "pid": 1728719, "tid": 1728719, "ts": 6950381114731.550, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114731.801, "dur": 0.020 , "args": { "External id": 1830,"cbid": 1, "correlation": 6745 } }, { "ph": "f", "id": 6745, "pid": 1728719, "tid": 1728719, "ts": 6950381114731.801, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114731.911, "dur": 0.100 , "args": { "External id": 1830,"cbid": 468, "correlation": 6746 } }, { "ph": "f", "id": 6746, "pid": 1728719, "tid": 1728719, "ts": 6950381114731.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114732.161, "dur": 0.020 , "args": { "External id": 1830,"cbid": 1, "correlation": 6748 } }, { "ph": "f", "id": 6748, "pid": 1728719, "tid": 1728719, "ts": 6950381114732.161, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114732.262, "dur": 0.330 , "args": { "External id": 1830,"cbid": 273, "correlation": 6749 } }, { "ph": "f", "id": 6749, "pid": 1728719, "tid": 1728719, "ts": 6950381114732.262, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381157206.507, "dur": 676.633 , "args": { "External id": 1830,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6752, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6752, "pid": 0, "tid": 7, "ts": 6950381157206.507, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114732.933, "dur": 1.783 , "args": { "External id": 1830,"cbid": 211, "correlation": 6752 } }, { "ph": "s", "id": 6752, "pid": 1728719, "tid": 1728719, "ts": 6950381114732.933, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381157883.428, "dur": 153.918 , "args": { "External id": 1842,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6765, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6765, "pid": 0, "tid": 7, "ts": 6950381157883.428, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114742.300, "dur": 1.693 , "args": { "External id": 1842,"cbid": 307, "correlation": 6765 } }, { "ph": "s", "id": 6765, "pid": 1728719, "tid": 1728719, "ts": 6950381114742.300, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381158037.602, "dur": 58.368 , "args": { "External id": 1843,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6771, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6771, "pid": 0, "tid": 7, "ts": 6950381158037.602, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114750.756, "dur": 1.533 , "args": { "External id": 1843,"cbid": 652, "correlation": 6771 } }, { "ph": "s", "id": 6771, "pid": 1728719, "tid": 1728719, "ts": 6950381114750.756, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114761.907, "dur": 0.090 , "args": { "External id": 1844,"cbid": 200, "correlation": 6790 } }, { "ph": "f", "id": 6790, "pid": 1728719, "tid": 1728719, "ts": 6950381114761.907, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114762.097, "dur": 0.120 , "args": { "External id": 1844,"cbid": 468, "correlation": 6791 } }, { "ph": "f", "id": 6791, "pid": 1728719, "tid": 1728719, "ts": 6950381114762.097, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114762.438, "dur": 0.020 , "args": { "External id": 1844,"cbid": 1, "correlation": 6793 } }, { "ph": "f", "id": 6793, "pid": 1728719, "tid": 1728719, "ts": 6950381114762.438, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114762.528, "dur": 0.100 , "args": { "External id": 1844,"cbid": 468, "correlation": 6794 } }, { "ph": "f", "id": 6794, "pid": 1728719, "tid": 1728719, "ts": 6950381114762.528, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114762.778, "dur": 0.020 , "args": { "External id": 1844,"cbid": 1, "correlation": 6796 } }, { "ph": "f", "id": 6796, "pid": 1728719, "tid": 1728719, "ts": 6950381114762.778, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114762.889, "dur": 0.090 , "args": { "External id": 1844,"cbid": 468, "correlation": 6797 } }, { "ph": "f", "id": 6797, "pid": 1728719, "tid": 1728719, "ts": 6950381114762.889, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114763.149, "dur": 0.020 , "args": { "External id": 1844,"cbid": 1, "correlation": 6799 } }, { "ph": "f", "id": 6799, "pid": 1728719, "tid": 1728719, "ts": 6950381114763.149, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114763.279, "dur": 0.361 , "args": { "External id": 1844,"cbid": 273, "correlation": 6800 } }, { "ph": "f", "id": 6800, "pid": 1728719, "tid": 1728719, "ts": 6950381114763.279, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381158096.257, "dur": 679.705 , "args": { "External id": 1844,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6803, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6803, "pid": 0, "tid": 7, "ts": 6950381158096.257, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114763.991, "dur": 1.663 , "args": { "External id": 1844,"cbid": 211, "correlation": 6803 } }, { "ph": "s", "id": 6803, "pid": 1728719, "tid": 1728719, "ts": 6950381114763.991, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381158776.250, "dur": 40.096 , "args": { "External id": 1856,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6814, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6814, "pid": 0, "tid": 7, "ts": 6950381158776.250, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114773.689, "dur": 1.733 , "args": { "External id": 1856,"cbid": 307, "correlation": 6814 } }, { "ph": "s", "id": 6814, "pid": 1728719, "tid": 1728719, "ts": 6950381114773.689, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381158816.602, "dur": 21.024 , "args": { "External id": 1868,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6826, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6826, "pid": 0, "tid": 7, "ts": 6950381158816.602, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114853.076, "dur": 2.465 , "args": { "External id": 1868,"cbid": 307, "correlation": 6826 } }, { "ph": "s", "id": 6826, "pid": 1728719, "tid": 1728719, "ts": 6950381114853.076, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381158837.946, "dur": 13.343 , "args": { "External id": 1869,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6840, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6840, "pid": 0, "tid": 7, "ts": 6950381158837.946, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114867.273, "dur": 1.663 , "args": { "External id": 1869,"cbid": 652, "correlation": 6840 } }, { "ph": "s", "id": 6840, "pid": 1728719, "tid": 1728719, "ts": 6950381114867.273, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381158851.609, "dur": 12.256 , "args": { "External id": 1870,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6854, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6854, "pid": 0, "tid": 7, "ts": 6950381158851.609, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114878.834, "dur": 1.613 , "args": { "External id": 1870,"cbid": 652, "correlation": 6854 } }, { "ph": "s", "id": 6854, "pid": 1728719, "tid": 1728719, "ts": 6950381114878.834, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381158864.153, "dur": 10.720 , "args": { "External id": 1871,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6868, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6868, "pid": 0, "tid": 7, "ts": 6950381158864.153, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381114889.314, "dur": 1.503 , "args": { "External id": 1871,"cbid": 652, "correlation": 6868 } }, { "ph": "s", "id": 6868, "pid": 1728719, "tid": 1728719, "ts": 6950381114889.314, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114900.925, "dur": 0.161 , "args": { "External id": 1872,"cbid": 200, "correlation": 6883 } }, { "ph": "f", "id": 6883, "pid": 1728719, "tid": 1728719, "ts": 6950381114900.925, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114901.236, "dur": 0.170 , "args": { "External id": 1872,"cbid": 468, "correlation": 6884 } }, { "ph": "f", "id": 6884, "pid": 1728719, "tid": 1728719, "ts": 6950381114901.236, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114901.757, "dur": 0.030 , "args": { "External id": 1872,"cbid": 1, "correlation": 6886 } }, { "ph": "f", "id": 6886, "pid": 1728719, "tid": 1728719, "ts": 6950381114901.757, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114901.857, "dur": 0.100 , "args": { "External id": 1872,"cbid": 468, "correlation": 6887 } }, { "ph": "f", "id": 6887, "pid": 1728719, "tid": 1728719, "ts": 6950381114901.857, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114902.118, "dur": 0.030 , "args": { "External id": 1872,"cbid": 1, "correlation": 6889 } }, { "ph": "f", "id": 6889, "pid": 1728719, "tid": 1728719, "ts": 6950381114902.118, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114902.258, "dur": 0.100 , "args": { "External id": 1872,"cbid": 468, "correlation": 6890 } }, { "ph": "f", "id": 6890, "pid": 1728719, "tid": 1728719, "ts": 6950381114902.258, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114902.508, "dur": 0.020 , "args": { "External id": 1872,"cbid": 1, "correlation": 6892 } }, { "ph": "f", "id": 6892, "pid": 1728719, "tid": 1728719, "ts": 6950381114902.508, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114902.649, "dur": 0.671 , "args": { "External id": 1872,"cbid": 273, "correlation": 6893 } }, { "ph": "f", "id": 6893, "pid": 1728719, "tid": 1728719, "ts": 6950381114902.649, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381158875.161, "dur": 216.254 , "args": { "External id": 1872,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6896, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6896, "pid": 0, "tid": 7, "ts": 6950381158875.161, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114903.761, "dur": 1.783 , "args": { "External id": 1872,"cbid": 211, "correlation": 6896 } }, { "ph": "s", "id": 6896, "pid": 1728719, "tid": 1728719, "ts": 6950381114903.761, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381159091.735, "dur": 8.832 , "args": { "External id": 1884,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6909, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6909, "pid": 0, "tid": 7, "ts": 6950381159091.735, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114914.110, "dur": 1.823 , "args": { "External id": 1884,"cbid": 307, "correlation": 6909 } }, { "ph": "s", "id": 6909, "pid": 1728719, "tid": 1728719, "ts": 6950381114914.110, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381159100.855, "dur": 36.991 , "args": { "External id": 1889,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6932, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6932, "pid": 0, "tid": 7, "ts": 6950381159100.855, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114933.165, "dur": 1.613 , "args": { "External id": 1889,"cbid": 307, "correlation": 6932 } }, { "ph": "s", "id": 6932, "pid": 1728719, "tid": 1728719, "ts": 6950381114933.165, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114943.364, "dur": 0.111 , "args": { "External id": 1890,"cbid": 200, "correlation": 6947 } }, { "ph": "f", "id": 6947, "pid": 1728719, "tid": 1728719, "ts": 6950381114943.364, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114943.575, "dur": 0.130 , "args": { "External id": 1890,"cbid": 468, "correlation": 6948 } }, { "ph": "f", "id": 6948, "pid": 1728719, "tid": 1728719, "ts": 6950381114943.575, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114943.936, "dur": 0.030 , "args": { "External id": 1890,"cbid": 1, "correlation": 6950 } }, { "ph": "f", "id": 6950, "pid": 1728719, "tid": 1728719, "ts": 6950381114943.936, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114944.036, "dur": 0.100 , "args": { "External id": 1890,"cbid": 468, "correlation": 6951 } }, { "ph": "f", "id": 6951, "pid": 1728719, "tid": 1728719, "ts": 6950381114944.036, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114944.276, "dur": 0.020 , "args": { "External id": 1890,"cbid": 1, "correlation": 6953 } }, { "ph": "f", "id": 6953, "pid": 1728719, "tid": 1728719, "ts": 6950381114944.276, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114944.396, "dur": 0.091 , "args": { "External id": 1890,"cbid": 468, "correlation": 6954 } }, { "ph": "f", "id": 6954, "pid": 1728719, "tid": 1728719, "ts": 6950381114944.396, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114944.647, "dur": 0.020 , "args": { "External id": 1890,"cbid": 1, "correlation": 6956 } }, { "ph": "f", "id": 6956, "pid": 1728719, "tid": 1728719, "ts": 6950381114944.647, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114944.767, "dur": 0.411 , "args": { "External id": 1890,"cbid": 273, "correlation": 6957 } }, { "ph": "f", "id": 6957, "pid": 1728719, "tid": 1728719, "ts": 6950381114944.767, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381159138.102, "dur": 227.358 , "args": { "External id": 1890,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6960, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6960, "pid": 0, "tid": 7, "ts": 6950381159138.102, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114945.579, "dur": 1.693 , "args": { "External id": 1890,"cbid": 211, "correlation": 6960 } }, { "ph": "s", "id": 6960, "pid": 1728719, "tid": 1728719, "ts": 6950381114945.579, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381159365.748, "dur": 8.896 , "args": { "External id": 1902,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6969, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6969, "pid": 0, "tid": 7, "ts": 6950381159365.748, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114953.123, "dur": 1.583 , "args": { "External id": 1902,"cbid": 307, "correlation": 6969 } }, { "ph": "s", "id": 6969, "pid": 1728719, "tid": 1728719, "ts": 6950381114953.123, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381159374.932, "dur": 28.127 , "args": { "External id": 1907,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 6992, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 6992, "pid": 0, "tid": 7, "ts": 6950381159374.932, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114969.423, "dur": 1.493 , "args": { "External id": 1907,"cbid": 307, "correlation": 6992 } }, { "ph": "s", "id": 6992, "pid": 1728719, "tid": 1728719, "ts": 6950381114969.423, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114979.311, "dur": 0.091 , "args": { "External id": 1908,"cbid": 200, "correlation": 7007 } }, { "ph": "f", "id": 7007, "pid": 1728719, "tid": 1728719, "ts": 6950381114979.311, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114979.492, "dur": 0.130 , "args": { "External id": 1908,"cbid": 468, "correlation": 7008 } }, { "ph": "f", "id": 7008, "pid": 1728719, "tid": 1728719, "ts": 6950381114979.492, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114979.812, "dur": 0.020 , "args": { "External id": 1908,"cbid": 1, "correlation": 7010 } }, { "ph": "f", "id": 7010, "pid": 1728719, "tid": 1728719, "ts": 6950381114979.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114979.912, "dur": 0.111 , "args": { "External id": 1908,"cbid": 468, "correlation": 7011 } }, { "ph": "f", "id": 7011, "pid": 1728719, "tid": 1728719, "ts": 6950381114979.912, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114980.173, "dur": 0.020 , "args": { "External id": 1908,"cbid": 1, "correlation": 7013 } }, { "ph": "f", "id": 7013, "pid": 1728719, "tid": 1728719, "ts": 6950381114980.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114980.303, "dur": 0.100 , "args": { "External id": 1908,"cbid": 468, "correlation": 7014 } }, { "ph": "f", "id": 7014, "pid": 1728719, "tid": 1728719, "ts": 6950381114980.303, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381114980.554, "dur": 0.020 , "args": { "External id": 1908,"cbid": 1, "correlation": 7016 } }, { "ph": "f", "id": 7016, "pid": 1728719, "tid": 1728719, "ts": 6950381114980.554, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381114980.664, "dur": 0.391 , "args": { "External id": 1908,"cbid": 273, "correlation": 7017 } }, { "ph": "f", "id": 7017, "pid": 1728719, "tid": 1728719, "ts": 6950381114980.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381159403.348, "dur": 228.381 , "args": { "External id": 1908,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7020, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7020, "pid": 0, "tid": 7, "ts": 6950381159403.348, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114981.435, "dur": 1.834 , "args": { "External id": 1908,"cbid": 211, "correlation": 7020 } }, { "ph": "s", "id": 7020, "pid": 1728719, "tid": 1728719, "ts": 6950381114981.435, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381159632.017, "dur": 59.967 , "args": { "External id": 1920,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7034, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7034, "pid": 0, "tid": 7, "ts": 6950381159632.017, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381114990.833, "dur": 1.543 , "args": { "External id": 1920,"cbid": 307, "correlation": 7034 } }, { "ph": "s", "id": 7034, "pid": 1728719, "tid": 1728719, "ts": 6950381114990.833, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381159692.240, "dur": 65.376 , "args": { "External id": 1921,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7044, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7044, "pid": 0, "tid": 7, "ts": 6950381159692.240, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115001.092, "dur": 1.713 , "args": { "External id": 1921,"cbid": 307, "correlation": 7044 } }, { "ph": "s", "id": 7044, "pid": 1728719, "tid": 1728719, "ts": 6950381115001.092, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381159757.872, "dur": 38.623 , "args": { "External id": 1922,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7053, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7053, "pid": 0, "tid": 7, "ts": 6950381159757.872, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115010.880, "dur": 1.613 , "args": { "External id": 1922,"cbid": 307, "correlation": 7053 } }, { "ph": "s", "id": 7053, "pid": 1728719, "tid": 1728719, "ts": 6950381115010.880, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115026.199, "dur": 0.160 , "args": { "External id": 1930,"cbid": 200, "correlation": 7088 } }, { "ph": "f", "id": 7088, "pid": 1728719, "tid": 1728719, "ts": 6950381115026.199, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115026.439, "dur": 0.080 , "args": { "External id": 1930,"cbid": 200, "correlation": 7089 } }, { "ph": "f", "id": 7089, "pid": 1728719, "tid": 1728719, "ts": 6950381115026.439, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115026.639, "dur": 0.511 , "args": { "External id": 1930,"cbid": 273, "correlation": 7090 } }, { "ph": "f", "id": 7090, "pid": 1728719, "tid": 1728719, "ts": 6950381115026.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381159796.783, "dur": 848.759 , "args": { "External id": 1930,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7092, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7092, "pid": 0, "tid": 7, "ts": 6950381159796.783, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115027.461, "dur": 1.823 , "args": { "External id": 1930,"cbid": 211, "correlation": 7092 } }, { "ph": "s", "id": 7092, "pid": 1728719, "tid": 1728719, "ts": 6950381115027.461, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381160645.830, "dur": 9.312 , "args": { "External id": 1939,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7112, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7112, "pid": 0, "tid": 7, "ts": 6950381160645.830, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115044.282, "dur": 1.603 , "args": { "External id": 1939,"cbid": 652, "correlation": 7112 } }, { "ph": "s", "id": 7112, "pid": 1728719, "tid": 1728719, "ts": 6950381115044.282, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115055.243, "dur": 0.120 , "args": { "External id": 1940,"cbid": 200, "correlation": 7127 } }, { "ph": "f", "id": 7127, "pid": 1728719, "tid": 1728719, "ts": 6950381115055.243, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115055.453, "dur": 0.160 , "args": { "External id": 1940,"cbid": 468, "correlation": 7128 } }, { "ph": "f", "id": 7128, "pid": 1728719, "tid": 1728719, "ts": 6950381115055.453, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115055.964, "dur": 0.040 , "args": { "External id": 1940,"cbid": 1, "correlation": 7130 } }, { "ph": "f", "id": 7130, "pid": 1728719, "tid": 1728719, "ts": 6950381115055.964, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115056.074, "dur": 0.110 , "args": { "External id": 1940,"cbid": 468, "correlation": 7131 } }, { "ph": "f", "id": 7131, "pid": 1728719, "tid": 1728719, "ts": 6950381115056.074, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115056.335, "dur": 0.020 , "args": { "External id": 1940,"cbid": 1, "correlation": 7133 } }, { "ph": "f", "id": 7133, "pid": 1728719, "tid": 1728719, "ts": 6950381115056.335, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115056.465, "dur": 0.110 , "args": { "External id": 1940,"cbid": 468, "correlation": 7134 } }, { "ph": "f", "id": 7134, "pid": 1728719, "tid": 1728719, "ts": 6950381115056.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115056.735, "dur": 0.021 , "args": { "External id": 1940,"cbid": 1, "correlation": 7136 } }, { "ph": "f", "id": 7136, "pid": 1728719, "tid": 1728719, "ts": 6950381115056.735, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115056.856, "dur": 0.410 , "args": { "External id": 1940,"cbid": 273, "correlation": 7137 } }, { "ph": "f", "id": 7137, "pid": 1728719, "tid": 1728719, "ts": 6950381115056.856, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381160655.398, "dur": 208.222 , "args": { "External id": 1940,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7140, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7140, "pid": 0, "tid": 7, "ts": 6950381160655.398, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115057.627, "dur": 1.894 , "args": { "External id": 1940,"cbid": 211, "correlation": 7140 } }, { "ph": "s", "id": 7140, "pid": 1728719, "tid": 1728719, "ts": 6950381115057.627, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381160863.908, "dur": 46.687 , "args": { "External id": 1952,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7152, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7152, "pid": 0, "tid": 7, "ts": 6950381160863.908, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115067.486, "dur": 1.833 , "args": { "External id": 1952,"cbid": 307, "correlation": 7152 } }, { "ph": "s", "id": 7152, "pid": 1728719, "tid": 1728719, "ts": 6950381115067.486, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381160910.883, "dur": 13.152 , "args": { "External id": 1953,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7158, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7158, "pid": 0, "tid": 7, "ts": 6950381160910.883, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115075.631, "dur": 1.482 , "args": { "External id": 1953,"cbid": 652, "correlation": 7158 } }, { "ph": "s", "id": 7158, "pid": 1728719, "tid": 1728719, "ts": 6950381115075.631, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381160924.355, "dur": 12.416 , "args": { "External id": 1954,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7164, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7164, "pid": 0, "tid": 7, "ts": 6950381160924.355, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115083.926, "dur": 1.473 , "args": { "External id": 1954,"cbid": 652, "correlation": 7164 } }, { "ph": "s", "id": 7164, "pid": 1728719, "tid": 1728719, "ts": 6950381115083.926, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115095.678, "dur": 0.100 , "args": { "External id": 1955,"cbid": 200, "correlation": 7183 } }, { "ph": "f", "id": 7183, "pid": 1728719, "tid": 1728719, "ts": 6950381115095.678, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115095.888, "dur": 0.121 , "args": { "External id": 1955,"cbid": 468, "correlation": 7184 } }, { "ph": "f", "id": 7184, "pid": 1728719, "tid": 1728719, "ts": 6950381115095.888, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115096.239, "dur": 0.020 , "args": { "External id": 1955,"cbid": 1, "correlation": 7186 } }, { "ph": "f", "id": 7186, "pid": 1728719, "tid": 1728719, "ts": 6950381115096.239, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115096.359, "dur": 0.100 , "args": { "External id": 1955,"cbid": 468, "correlation": 7187 } }, { "ph": "f", "id": 7187, "pid": 1728719, "tid": 1728719, "ts": 6950381115096.359, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115096.600, "dur": 0.020 , "args": { "External id": 1955,"cbid": 1, "correlation": 7189 } }, { "ph": "f", "id": 7189, "pid": 1728719, "tid": 1728719, "ts": 6950381115096.600, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115096.710, "dur": 0.100 , "args": { "External id": 1955,"cbid": 468, "correlation": 7190 } }, { "ph": "f", "id": 7190, "pid": 1728719, "tid": 1728719, "ts": 6950381115096.710, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115096.960, "dur": 0.020 , "args": { "External id": 1955,"cbid": 1, "correlation": 7192 } }, { "ph": "f", "id": 7192, "pid": 1728719, "tid": 1728719, "ts": 6950381115096.960, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115097.071, "dur": 0.380 , "args": { "External id": 1955,"cbid": 273, "correlation": 7193 } }, { "ph": "f", "id": 7193, "pid": 1728719, "tid": 1728719, "ts": 6950381115097.071, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381160937.059, "dur": 662.073 , "args": { "External id": 1955,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7196, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7196, "pid": 0, "tid": 7, "ts": 6950381160937.059, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115097.812, "dur": 1.713 , "args": { "External id": 1955,"cbid": 211, "correlation": 7196 } }, { "ph": "s", "id": 7196, "pid": 1728719, "tid": 1728719, "ts": 6950381115097.812, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115109.574, "dur": 0.080 , "args": { "External id": 1967,"cbid": 200, "correlation": 7220 } }, { "ph": "f", "id": 7220, "pid": 1728719, "tid": 1728719, "ts": 6950381115109.574, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115109.754, "dur": 0.120 , "args": { "External id": 1967,"cbid": 468, "correlation": 7221 } }, { "ph": "f", "id": 7221, "pid": 1728719, "tid": 1728719, "ts": 6950381115109.754, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115110.055, "dur": 0.020 , "args": { "External id": 1967,"cbid": 1, "correlation": 7223 } }, { "ph": "f", "id": 7223, "pid": 1728719, "tid": 1728719, "ts": 6950381115110.055, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115110.155, "dur": 0.100 , "args": { "External id": 1967,"cbid": 468, "correlation": 7224 } }, { "ph": "f", "id": 7224, "pid": 1728719, "tid": 1728719, "ts": 6950381115110.155, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115110.395, "dur": 0.030 , "args": { "External id": 1967,"cbid": 1, "correlation": 7226 } }, { "ph": "f", "id": 7226, "pid": 1728719, "tid": 1728719, "ts": 6950381115110.395, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115110.516, "dur": 0.090 , "args": { "External id": 1967,"cbid": 468, "correlation": 7227 } }, { "ph": "f", "id": 7227, "pid": 1728719, "tid": 1728719, "ts": 6950381115110.516, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115110.786, "dur": 0.020 , "args": { "External id": 1967,"cbid": 1, "correlation": 7229 } }, { "ph": "f", "id": 7229, "pid": 1728719, "tid": 1728719, "ts": 6950381115110.786, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115110.886, "dur": 0.331 , "args": { "External id": 1967,"cbid": 273, "correlation": 7230 } }, { "ph": "f", "id": 7230, "pid": 1728719, "tid": 1728719, "ts": 6950381115110.886, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381161599.452, "dur": 677.944 , "args": { "External id": 1967,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7233, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7233, "pid": 0, "tid": 7, "ts": 6950381161599.452, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115111.578, "dur": 1.733 , "args": { "External id": 1967,"cbid": 211, "correlation": 7233 } }, { "ph": "s", "id": 7233, "pid": 1728719, "tid": 1728719, "ts": 6950381115111.578, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381162277.684, "dur": 153.759 , "args": { "External id": 1979,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7246, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7246, "pid": 0, "tid": 7, "ts": 6950381162277.684, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115120.644, "dur": 1.633 , "args": { "External id": 1979,"cbid": 307, "correlation": 7246 } }, { "ph": "s", "id": 7246, "pid": 1728719, "tid": 1728719, "ts": 6950381115120.644, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381162431.731, "dur": 58.175 , "args": { "External id": 1980,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7252, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7252, "pid": 0, "tid": 7, "ts": 6950381162431.731, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115128.589, "dur": 1.513 , "args": { "External id": 1980,"cbid": 652, "correlation": 7252 } }, { "ph": "s", "id": 7252, "pid": 1728719, "tid": 1728719, "ts": 6950381115128.589, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115139.990, "dur": 0.091 , "args": { "External id": 1981,"cbid": 200, "correlation": 7271 } }, { "ph": "f", "id": 7271, "pid": 1728719, "tid": 1728719, "ts": 6950381115139.990, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115140.612, "dur": 0.110 , "args": { "External id": 1981,"cbid": 468, "correlation": 7272 } }, { "ph": "f", "id": 7272, "pid": 1728719, "tid": 1728719, "ts": 6950381115140.612, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115141.113, "dur": 0.020 , "args": { "External id": 1981,"cbid": 1, "correlation": 7274 } }, { "ph": "f", "id": 7274, "pid": 1728719, "tid": 1728719, "ts": 6950381115141.113, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115141.203, "dur": 0.100 , "args": { "External id": 1981,"cbid": 468, "correlation": 7275 } }, { "ph": "f", "id": 7275, "pid": 1728719, "tid": 1728719, "ts": 6950381115141.203, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115141.463, "dur": 0.020 , "args": { "External id": 1981,"cbid": 1, "correlation": 7277 } }, { "ph": "f", "id": 7277, "pid": 1728719, "tid": 1728719, "ts": 6950381115141.463, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115141.593, "dur": 0.101 , "args": { "External id": 1981,"cbid": 468, "correlation": 7278 } }, { "ph": "f", "id": 7278, "pid": 1728719, "tid": 1728719, "ts": 6950381115141.593, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115141.844, "dur": 0.020 , "args": { "External id": 1981,"cbid": 1, "correlation": 7280 } }, { "ph": "f", "id": 7280, "pid": 1728719, "tid": 1728719, "ts": 6950381115141.844, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115141.974, "dur": 0.551 , "args": { "External id": 1981,"cbid": 273, "correlation": 7281 } }, { "ph": "f", "id": 7281, "pid": 1728719, "tid": 1728719, "ts": 6950381115141.974, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381162490.194, "dur": 678.233 , "args": { "External id": 1981,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7284, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7284, "pid": 0, "tid": 7, "ts": 6950381162490.194, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115142.896, "dur": 1.643 , "args": { "External id": 1981,"cbid": 211, "correlation": 7284 } }, { "ph": "s", "id": 7284, "pid": 1728719, "tid": 1728719, "ts": 6950381115142.896, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381163168.715, "dur": 40.607 , "args": { "External id": 1993,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7295, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7295, "pid": 0, "tid": 7, "ts": 6950381163168.715, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115152.644, "dur": 1.623 , "args": { "External id": 1993,"cbid": 307, "correlation": 7295 } }, { "ph": "s", "id": 7295, "pid": 1728719, "tid": 1728719, "ts": 6950381115152.644, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381163209.578, "dur": 21.280 , "args": { "External id": 2005,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7307, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7307, "pid": 0, "tid": 7, "ts": 6950381163209.578, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115231.531, "dur": 2.374 , "args": { "External id": 2005,"cbid": 307, "correlation": 7307 } }, { "ph": "s", "id": 7307, "pid": 1728719, "tid": 1728719, "ts": 6950381115231.531, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381163231.146, "dur": 12.992 , "args": { "External id": 2006,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7321, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7321, "pid": 0, "tid": 7, "ts": 6950381163231.146, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115245.767, "dur": 1.683 , "args": { "External id": 2006,"cbid": 652, "correlation": 7321 } }, { "ph": "s", "id": 7321, "pid": 1728719, "tid": 1728719, "ts": 6950381115245.767, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381163244.458, "dur": 12.224 , "args": { "External id": 2007,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7335, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7335, "pid": 0, "tid": 7, "ts": 6950381163244.458, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115257.820, "dur": 1.452 , "args": { "External id": 2007,"cbid": 652, "correlation": 7335 } }, { "ph": "s", "id": 7335, "pid": 1728719, "tid": 1728719, "ts": 6950381115257.820, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381163256.970, "dur": 10.431 , "args": { "External id": 2008,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7349, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7349, "pid": 0, "tid": 7, "ts": 6950381163256.970, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115268.670, "dur": 1.473 , "args": { "External id": 2008,"cbid": 652, "correlation": 7349 } }, { "ph": "s", "id": 7349, "pid": 1728719, "tid": 1728719, "ts": 6950381115268.670, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115280.081, "dur": 0.130 , "args": { "External id": 2009,"cbid": 200, "correlation": 7364 } }, { "ph": "f", "id": 7364, "pid": 1728719, "tid": 1728719, "ts": 6950381115280.081, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115280.362, "dur": 0.230 , "args": { "External id": 2009,"cbid": 468, "correlation": 7365 } }, { "ph": "f", "id": 7365, "pid": 1728719, "tid": 1728719, "ts": 6950381115280.362, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115280.943, "dur": 0.030 , "args": { "External id": 2009,"cbid": 1, "correlation": 7367 } }, { "ph": "f", "id": 7367, "pid": 1728719, "tid": 1728719, "ts": 6950381115280.943, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115281.043, "dur": 0.100 , "args": { "External id": 2009,"cbid": 468, "correlation": 7368 } }, { "ph": "f", "id": 7368, "pid": 1728719, "tid": 1728719, "ts": 6950381115281.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115281.314, "dur": 0.030 , "args": { "External id": 2009,"cbid": 1, "correlation": 7370 } }, { "ph": "f", "id": 7370, "pid": 1728719, "tid": 1728719, "ts": 6950381115281.314, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115281.434, "dur": 0.100 , "args": { "External id": 2009,"cbid": 468, "correlation": 7371 } }, { "ph": "f", "id": 7371, "pid": 1728719, "tid": 1728719, "ts": 6950381115281.434, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115281.704, "dur": 0.040 , "args": { "External id": 2009,"cbid": 1, "correlation": 7373 } }, { "ph": "f", "id": 7373, "pid": 1728719, "tid": 1728719, "ts": 6950381115281.704, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115281.875, "dur": 0.631 , "args": { "External id": 2009,"cbid": 273, "correlation": 7374 } }, { "ph": "f", "id": 7374, "pid": 1728719, "tid": 1728719, "ts": 6950381115281.875, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381163267.689, "dur": 217.182 , "args": { "External id": 2009,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7377, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7377, "pid": 0, "tid": 7, "ts": 6950381163267.689, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115282.967, "dur": 1.803 , "args": { "External id": 2009,"cbid": 211, "correlation": 7377 } }, { "ph": "s", "id": 7377, "pid": 1728719, "tid": 1728719, "ts": 6950381115282.967, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381163485.159, "dur": 8.832 , "args": { "External id": 2021,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7390, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7390, "pid": 0, "tid": 7, "ts": 6950381163485.159, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115293.486, "dur": 1.703 , "args": { "External id": 2021,"cbid": 307, "correlation": 7390 } }, { "ph": "s", "id": 7390, "pid": 1728719, "tid": 1728719, "ts": 6950381115293.486, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381163494.279, "dur": 37.088 , "args": { "External id": 2026,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7413, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7413, "pid": 0, "tid": 7, "ts": 6950381163494.279, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115312.321, "dur": 1.663 , "args": { "External id": 2026,"cbid": 307, "correlation": 7413 } }, { "ph": "s", "id": 7413, "pid": 1728719, "tid": 1728719, "ts": 6950381115312.321, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115322.630, "dur": 0.121 , "args": { "External id": 2027,"cbid": 200, "correlation": 7428 } }, { "ph": "f", "id": 7428, "pid": 1728719, "tid": 1728719, "ts": 6950381115322.630, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115322.841, "dur": 0.140 , "args": { "External id": 2027,"cbid": 468, "correlation": 7429 } }, { "ph": "f", "id": 7429, "pid": 1728719, "tid": 1728719, "ts": 6950381115322.841, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115323.222, "dur": 0.030 , "args": { "External id": 2027,"cbid": 1, "correlation": 7431 } }, { "ph": "f", "id": 7431, "pid": 1728719, "tid": 1728719, "ts": 6950381115323.222, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115323.332, "dur": 0.110 , "args": { "External id": 2027,"cbid": 468, "correlation": 7432 } }, { "ph": "f", "id": 7432, "pid": 1728719, "tid": 1728719, "ts": 6950381115323.332, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115323.602, "dur": 0.020 , "args": { "External id": 2027,"cbid": 1, "correlation": 7434 } }, { "ph": "f", "id": 7434, "pid": 1728719, "tid": 1728719, "ts": 6950381115323.602, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115323.722, "dur": 0.091 , "args": { "External id": 2027,"cbid": 468, "correlation": 7435 } }, { "ph": "f", "id": 7435, "pid": 1728719, "tid": 1728719, "ts": 6950381115323.722, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115323.973, "dur": 0.020 , "args": { "External id": 2027,"cbid": 1, "correlation": 7437 } }, { "ph": "f", "id": 7437, "pid": 1728719, "tid": 1728719, "ts": 6950381115323.973, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115324.093, "dur": 0.471 , "args": { "External id": 2027,"cbid": 273, "correlation": 7438 } }, { "ph": "f", "id": 7438, "pid": 1728719, "tid": 1728719, "ts": 6950381115324.093, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381163531.655, "dur": 227.325 , "args": { "External id": 2027,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7441, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7441, "pid": 0, "tid": 7, "ts": 6950381163531.655, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115324.945, "dur": 1.733 , "args": { "External id": 2027,"cbid": 211, "correlation": 7441 } }, { "ph": "s", "id": 7441, "pid": 1728719, "tid": 1728719, "ts": 6950381115324.945, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381163759.268, "dur": 8.864 , "args": { "External id": 2039,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7450, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7450, "pid": 0, "tid": 7, "ts": 6950381163759.268, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115332.419, "dur": 1.502 , "args": { "External id": 2039,"cbid": 307, "correlation": 7450 } }, { "ph": "s", "id": 7450, "pid": 1728719, "tid": 1728719, "ts": 6950381115332.419, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381163768.420, "dur": 25.312 , "args": { "External id": 2044,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7473, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7473, "pid": 0, "tid": 7, "ts": 6950381163768.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115348.068, "dur": 1.773 , "args": { "External id": 2044,"cbid": 307, "correlation": 7473 } }, { "ph": "s", "id": 7473, "pid": 1728719, "tid": 1728719, "ts": 6950381115348.068, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115361.533, "dur": 0.100 , "args": { "External id": 2045,"cbid": 200, "correlation": 7488 } }, { "ph": "f", "id": 7488, "pid": 1728719, "tid": 1728719, "ts": 6950381115361.533, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115361.713, "dur": 0.130 , "args": { "External id": 2045,"cbid": 468, "correlation": 7489 } }, { "ph": "f", "id": 7489, "pid": 1728719, "tid": 1728719, "ts": 6950381115361.713, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115362.064, "dur": 0.030 , "args": { "External id": 2045,"cbid": 1, "correlation": 7491 } }, { "ph": "f", "id": 7491, "pid": 1728719, "tid": 1728719, "ts": 6950381115362.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115362.154, "dur": 0.110 , "args": { "External id": 2045,"cbid": 468, "correlation": 7492 } }, { "ph": "f", "id": 7492, "pid": 1728719, "tid": 1728719, "ts": 6950381115362.154, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115362.404, "dur": 0.021 , "args": { "External id": 2045,"cbid": 1, "correlation": 7494 } }, { "ph": "f", "id": 7494, "pid": 1728719, "tid": 1728719, "ts": 6950381115362.404, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115362.525, "dur": 0.090 , "args": { "External id": 2045,"cbid": 468, "correlation": 7495 } }, { "ph": "f", "id": 7495, "pid": 1728719, "tid": 1728719, "ts": 6950381115362.525, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115362.795, "dur": 0.020 , "args": { "External id": 2045,"cbid": 1, "correlation": 7497 } }, { "ph": "f", "id": 7497, "pid": 1728719, "tid": 1728719, "ts": 6950381115362.795, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115362.895, "dur": 0.411 , "args": { "External id": 2045,"cbid": 273, "correlation": 7498 } }, { "ph": "f", "id": 7498, "pid": 1728719, "tid": 1728719, "ts": 6950381115362.895, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381163794.020, "dur": 229.597 , "args": { "External id": 2045,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7501, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7501, "pid": 0, "tid": 7, "ts": 6950381163794.020, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115363.667, "dur": 1.703 , "args": { "External id": 2045,"cbid": 211, "correlation": 7501 } }, { "ph": "s", "id": 7501, "pid": 1728719, "tid": 1728719, "ts": 6950381115363.667, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381164023.905, "dur": 59.136 , "args": { "External id": 2057,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7515, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7515, "pid": 0, "tid": 7, "ts": 6950381164023.905, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115372.784, "dur": 1.623 , "args": { "External id": 2057,"cbid": 307, "correlation": 7515 } }, { "ph": "s", "id": 7515, "pid": 1728719, "tid": 1728719, "ts": 6950381115372.784, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381164083.329, "dur": 66.399 , "args": { "External id": 2058,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7525, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7525, "pid": 0, "tid": 7, "ts": 6950381164083.329, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115382.832, "dur": 1.704 , "args": { "External id": 2058,"cbid": 307, "correlation": 7525 } }, { "ph": "s", "id": 7525, "pid": 1728719, "tid": 1728719, "ts": 6950381115382.832, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381164149.984, "dur": 38.592 , "args": { "External id": 2059,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7534, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7534, "pid": 0, "tid": 7, "ts": 6950381164149.984, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115392.250, "dur": 1.693 , "args": { "External id": 2059,"cbid": 307, "correlation": 7534 } }, { "ph": "s", "id": 7534, "pid": 1728719, "tid": 1728719, "ts": 6950381115392.250, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115407.268, "dur": 0.100 , "args": { "External id": 2067,"cbid": 200, "correlation": 7569 } }, { "ph": "f", "id": 7569, "pid": 1728719, "tid": 1728719, "ts": 6950381115407.268, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115407.428, "dur": 0.070 , "args": { "External id": 2067,"cbid": 200, "correlation": 7570 } }, { "ph": "f", "id": 7570, "pid": 1728719, "tid": 1728719, "ts": 6950381115407.428, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115407.589, "dur": 0.511 , "args": { "External id": 2067,"cbid": 273, "correlation": 7571 } }, { "ph": "f", "id": 7571, "pid": 1728719, "tid": 1728719, "ts": 6950381115407.589, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381164188.864, "dur": 852.630 , "args": { "External id": 2067,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7573, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7573, "pid": 0, "tid": 7, "ts": 6950381164188.864, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115408.380, "dur": 1.934 , "args": { "External id": 2067,"cbid": 211, "correlation": 7573 } }, { "ph": "s", "id": 7573, "pid": 1728719, "tid": 1728719, "ts": 6950381115408.380, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381165041.750, "dur": 9.440 , "args": { "External id": 2076,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7593, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7593, "pid": 0, "tid": 7, "ts": 6950381165041.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115425.011, "dur": 1.613 , "args": { "External id": 2076,"cbid": 652, "correlation": 7593 } }, { "ph": "s", "id": 7593, "pid": 1728719, "tid": 1728719, "ts": 6950381115425.011, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115435.861, "dur": 0.120 , "args": { "External id": 2077,"cbid": 200, "correlation": 7608 } }, { "ph": "f", "id": 7608, "pid": 1728719, "tid": 1728719, "ts": 6950381115435.861, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115436.082, "dur": 0.180 , "args": { "External id": 2077,"cbid": 468, "correlation": 7609 } }, { "ph": "f", "id": 7609, "pid": 1728719, "tid": 1728719, "ts": 6950381115436.082, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115436.532, "dur": 0.031 , "args": { "External id": 2077,"cbid": 1, "correlation": 7611 } }, { "ph": "f", "id": 7611, "pid": 1728719, "tid": 1728719, "ts": 6950381115436.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115436.633, "dur": 0.090 , "args": { "External id": 2077,"cbid": 468, "correlation": 7612 } }, { "ph": "f", "id": 7612, "pid": 1728719, "tid": 1728719, "ts": 6950381115436.633, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115436.883, "dur": 0.020 , "args": { "External id": 2077,"cbid": 1, "correlation": 7614 } }, { "ph": "f", "id": 7614, "pid": 1728719, "tid": 1728719, "ts": 6950381115436.883, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115437.003, "dur": 0.101 , "args": { "External id": 2077,"cbid": 468, "correlation": 7615 } }, { "ph": "f", "id": 7615, "pid": 1728719, "tid": 1728719, "ts": 6950381115437.003, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115437.264, "dur": 0.020 , "args": { "External id": 2077,"cbid": 1, "correlation": 7617 } }, { "ph": "f", "id": 7617, "pid": 1728719, "tid": 1728719, "ts": 6950381115437.264, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115437.364, "dur": 0.411 , "args": { "External id": 2077,"cbid": 273, "correlation": 7618 } }, { "ph": "f", "id": 7618, "pid": 1728719, "tid": 1728719, "ts": 6950381115437.364, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381165051.446, "dur": 208.222 , "args": { "External id": 2077,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7621, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7621, "pid": 0, "tid": 7, "ts": 6950381165051.446, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115438.145, "dur": 1.734 , "args": { "External id": 2077,"cbid": 211, "correlation": 7621 } }, { "ph": "s", "id": 7621, "pid": 1728719, "tid": 1728719, "ts": 6950381115438.145, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381165259.956, "dur": 46.015 , "args": { "External id": 2089,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7633, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7633, "pid": 0, "tid": 7, "ts": 6950381165259.956, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115447.773, "dur": 1.643 , "args": { "External id": 2089,"cbid": 307, "correlation": 7633 } }, { "ph": "s", "id": 7633, "pid": 1728719, "tid": 1728719, "ts": 6950381115447.773, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381165306.259, "dur": 13.184 , "args": { "External id": 2090,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7639, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7639, "pid": 0, "tid": 7, "ts": 6950381165306.259, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115455.838, "dur": 1.463 , "args": { "External id": 2090,"cbid": 652, "correlation": 7639 } }, { "ph": "s", "id": 7639, "pid": 1728719, "tid": 1728719, "ts": 6950381115455.838, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381165319.763, "dur": 12.544 , "args": { "External id": 2091,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7645, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7645, "pid": 0, "tid": 7, "ts": 6950381165319.763, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115464.274, "dur": 1.593 , "args": { "External id": 2091,"cbid": 652, "correlation": 7645 } }, { "ph": "s", "id": 7645, "pid": 1728719, "tid": 1728719, "ts": 6950381115464.274, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115475.806, "dur": 0.110 , "args": { "External id": 2092,"cbid": 200, "correlation": 7664 } }, { "ph": "f", "id": 7664, "pid": 1728719, "tid": 1728719, "ts": 6950381115475.806, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115476.016, "dur": 0.120 , "args": { "External id": 2092,"cbid": 468, "correlation": 7665 } }, { "ph": "f", "id": 7665, "pid": 1728719, "tid": 1728719, "ts": 6950381115476.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115476.357, "dur": 0.030 , "args": { "External id": 2092,"cbid": 1, "correlation": 7667 } }, { "ph": "f", "id": 7667, "pid": 1728719, "tid": 1728719, "ts": 6950381115476.357, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115476.457, "dur": 0.110 , "args": { "External id": 2092,"cbid": 468, "correlation": 7668 } }, { "ph": "f", "id": 7668, "pid": 1728719, "tid": 1728719, "ts": 6950381115476.457, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115476.707, "dur": 0.020 , "args": { "External id": 2092,"cbid": 1, "correlation": 7670 } }, { "ph": "f", "id": 7670, "pid": 1728719, "tid": 1728719, "ts": 6950381115476.707, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115476.817, "dur": 0.101 , "args": { "External id": 2092,"cbid": 468, "correlation": 7671 } }, { "ph": "f", "id": 7671, "pid": 1728719, "tid": 1728719, "ts": 6950381115476.817, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115477.078, "dur": 0.020 , "args": { "External id": 2092,"cbid": 1, "correlation": 7673 } }, { "ph": "f", "id": 7673, "pid": 1728719, "tid": 1728719, "ts": 6950381115477.078, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115477.178, "dur": 0.391 , "args": { "External id": 2092,"cbid": 273, "correlation": 7674 } }, { "ph": "f", "id": 7674, "pid": 1728719, "tid": 1728719, "ts": 6950381115477.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381165332.627, "dur": 663.193 , "args": { "External id": 2092,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7677, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7677, "pid": 0, "tid": 7, "ts": 6950381165332.627, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115477.940, "dur": 1.703 , "args": { "External id": 2092,"cbid": 211, "correlation": 7677 } }, { "ph": "s", "id": 7677, "pid": 1728719, "tid": 1728719, "ts": 6950381115477.940, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115489.651, "dur": 0.091 , "args": { "External id": 2104,"cbid": 200, "correlation": 7701 } }, { "ph": "f", "id": 7701, "pid": 1728719, "tid": 1728719, "ts": 6950381115489.651, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115489.832, "dur": 0.110 , "args": { "External id": 2104,"cbid": 468, "correlation": 7702 } }, { "ph": "f", "id": 7702, "pid": 1728719, "tid": 1728719, "ts": 6950381115489.832, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115490.142, "dur": 0.020 , "args": { "External id": 2104,"cbid": 1, "correlation": 7704 } }, { "ph": "f", "id": 7704, "pid": 1728719, "tid": 1728719, "ts": 6950381115490.142, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115490.242, "dur": 0.101 , "args": { "External id": 2104,"cbid": 468, "correlation": 7705 } }, { "ph": "f", "id": 7705, "pid": 1728719, "tid": 1728719, "ts": 6950381115490.242, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115490.493, "dur": 0.020 , "args": { "External id": 2104,"cbid": 1, "correlation": 7707 } }, { "ph": "f", "id": 7707, "pid": 1728719, "tid": 1728719, "ts": 6950381115490.493, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115490.593, "dur": 0.100 , "args": { "External id": 2104,"cbid": 468, "correlation": 7708 } }, { "ph": "f", "id": 7708, "pid": 1728719, "tid": 1728719, "ts": 6950381115490.593, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115490.864, "dur": 0.020 , "args": { "External id": 2104,"cbid": 1, "correlation": 7710 } }, { "ph": "f", "id": 7710, "pid": 1728719, "tid": 1728719, "ts": 6950381115490.864, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115490.964, "dur": 0.330 , "args": { "External id": 2104,"cbid": 273, "correlation": 7711 } }, { "ph": "f", "id": 7711, "pid": 1728719, "tid": 1728719, "ts": 6950381115490.964, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381165996.076, "dur": 672.889 , "args": { "External id": 2104,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7714, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7714, "pid": 0, "tid": 7, "ts": 6950381165996.076, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115491.635, "dur": 1.603 , "args": { "External id": 2104,"cbid": 211, "correlation": 7714 } }, { "ph": "s", "id": 7714, "pid": 1728719, "tid": 1728719, "ts": 6950381115491.635, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381166669.253, "dur": 154.782 , "args": { "External id": 2116,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7727, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7727, "pid": 0, "tid": 7, "ts": 6950381166669.253, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115500.712, "dur": 1.543 , "args": { "External id": 2116,"cbid": 307, "correlation": 7727 } }, { "ph": "s", "id": 7727, "pid": 1728719, "tid": 1728719, "ts": 6950381115500.712, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381166824.291, "dur": 57.887 , "args": { "External id": 2117,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7733, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7733, "pid": 0, "tid": 7, "ts": 6950381166824.291, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115509.067, "dur": 1.553 , "args": { "External id": 2117,"cbid": 652, "correlation": 7733 } }, { "ph": "s", "id": 7733, "pid": 1728719, "tid": 1728719, "ts": 6950381115509.067, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115520.198, "dur": 0.100 , "args": { "External id": 2118,"cbid": 200, "correlation": 7752 } }, { "ph": "f", "id": 7752, "pid": 1728719, "tid": 1728719, "ts": 6950381115520.198, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115520.389, "dur": 0.110 , "args": { "External id": 2118,"cbid": 468, "correlation": 7753 } }, { "ph": "f", "id": 7753, "pid": 1728719, "tid": 1728719, "ts": 6950381115520.389, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115520.719, "dur": 0.020 , "args": { "External id": 2118,"cbid": 1, "correlation": 7755 } }, { "ph": "f", "id": 7755, "pid": 1728719, "tid": 1728719, "ts": 6950381115520.719, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115520.809, "dur": 0.101 , "args": { "External id": 2118,"cbid": 468, "correlation": 7756 } }, { "ph": "f", "id": 7756, "pid": 1728719, "tid": 1728719, "ts": 6950381115520.809, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115521.060, "dur": 0.020 , "args": { "External id": 2118,"cbid": 1, "correlation": 7758 } }, { "ph": "f", "id": 7758, "pid": 1728719, "tid": 1728719, "ts": 6950381115521.060, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115521.170, "dur": 0.090 , "args": { "External id": 2118,"cbid": 468, "correlation": 7759 } }, { "ph": "f", "id": 7759, "pid": 1728719, "tid": 1728719, "ts": 6950381115521.170, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115521.420, "dur": 0.021 , "args": { "External id": 2118,"cbid": 1, "correlation": 7761 } }, { "ph": "f", "id": 7761, "pid": 1728719, "tid": 1728719, "ts": 6950381115521.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115521.541, "dur": 0.370 , "args": { "External id": 2118,"cbid": 273, "correlation": 7762 } }, { "ph": "f", "id": 7762, "pid": 1728719, "tid": 1728719, "ts": 6950381115521.541, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381166882.466, "dur": 664.249 , "args": { "External id": 2118,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7765, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7765, "pid": 0, "tid": 7, "ts": 6950381166882.466, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115522.252, "dur": 1.693 , "args": { "External id": 2118,"cbid": 211, "correlation": 7765 } }, { "ph": "s", "id": 7765, "pid": 1728719, "tid": 1728719, "ts": 6950381115522.252, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381167547.003, "dur": 41.120 , "args": { "External id": 2130,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7776, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7776, "pid": 0, "tid": 7, "ts": 6950381167547.003, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115532.782, "dur": 1.613 , "args": { "External id": 2130,"cbid": 307, "correlation": 7776 } }, { "ph": "s", "id": 7776, "pid": 1728719, "tid": 1728719, "ts": 6950381115532.782, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381167588.411, "dur": 21.152 , "args": { "External id": 2142,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7788, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7788, "pid": 0, "tid": 7, "ts": 6950381167588.411, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115615.205, "dur": 2.745 , "args": { "External id": 2142,"cbid": 307, "correlation": 7788 } }, { "ph": "s", "id": 7788, "pid": 1728719, "tid": 1728719, "ts": 6950381115615.205, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381167609.851, "dur": 12.383 , "args": { "External id": 2143,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7802, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7802, "pid": 0, "tid": 7, "ts": 6950381167609.851, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115629.732, "dur": 1.603 , "args": { "External id": 2143,"cbid": 652, "correlation": 7802 } }, { "ph": "s", "id": 7802, "pid": 1728719, "tid": 1728719, "ts": 6950381115629.732, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381167622.522, "dur": 12.224 , "args": { "External id": 2144,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7816, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7816, "pid": 0, "tid": 7, "ts": 6950381167622.522, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115641.414, "dur": 1.493 , "args": { "External id": 2144,"cbid": 652, "correlation": 7816 } }, { "ph": "s", "id": 7816, "pid": 1728719, "tid": 1728719, "ts": 6950381115641.414, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381167635.034, "dur": 10.976 , "args": { "External id": 2145,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7830, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7830, "pid": 0, "tid": 7, "ts": 6950381167635.034, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115651.833, "dur": 1.533 , "args": { "External id": 2145,"cbid": 652, "correlation": 7830 } }, { "ph": "s", "id": 7830, "pid": 1728719, "tid": 1728719, "ts": 6950381115651.833, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115663.405, "dur": 0.180 , "args": { "External id": 2146,"cbid": 200, "correlation": 7845 } }, { "ph": "f", "id": 7845, "pid": 1728719, "tid": 1728719, "ts": 6950381115663.405, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115663.775, "dur": 0.191 , "args": { "External id": 2146,"cbid": 468, "correlation": 7846 } }, { "ph": "f", "id": 7846, "pid": 1728719, "tid": 1728719, "ts": 6950381115663.775, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115664.357, "dur": 0.030 , "args": { "External id": 2146,"cbid": 1, "correlation": 7848 } }, { "ph": "f", "id": 7848, "pid": 1728719, "tid": 1728719, "ts": 6950381115664.357, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115664.457, "dur": 0.100 , "args": { "External id": 2146,"cbid": 468, "correlation": 7849 } }, { "ph": "f", "id": 7849, "pid": 1728719, "tid": 1728719, "ts": 6950381115664.457, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115664.707, "dur": 0.020 , "args": { "External id": 2146,"cbid": 1, "correlation": 7851 } }, { "ph": "f", "id": 7851, "pid": 1728719, "tid": 1728719, "ts": 6950381115664.707, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115664.837, "dur": 0.091 , "args": { "External id": 2146,"cbid": 468, "correlation": 7852 } }, { "ph": "f", "id": 7852, "pid": 1728719, "tid": 1728719, "ts": 6950381115664.837, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115665.088, "dur": 0.020 , "args": { "External id": 2146,"cbid": 1, "correlation": 7854 } }, { "ph": "f", "id": 7854, "pid": 1728719, "tid": 1728719, "ts": 6950381115665.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115665.248, "dur": 0.621 , "args": { "External id": 2146,"cbid": 273, "correlation": 7855 } }, { "ph": "f", "id": 7855, "pid": 1728719, "tid": 1728719, "ts": 6950381115665.248, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381167646.298, "dur": 217.278 , "args": { "External id": 2146,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7858, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7858, "pid": 0, "tid": 7, "ts": 6950381167646.298, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115666.320, "dur": 1.854 , "args": { "External id": 2146,"cbid": 211, "correlation": 7858 } }, { "ph": "s", "id": 7858, "pid": 1728719, "tid": 1728719, "ts": 6950381115666.320, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381167863.896, "dur": 8.768 , "args": { "External id": 2158,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7871, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7871, "pid": 0, "tid": 7, "ts": 6950381167863.896, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115676.760, "dur": 1.713 , "args": { "External id": 2158,"cbid": 307, "correlation": 7871 } }, { "ph": "s", "id": 7871, "pid": 1728719, "tid": 1728719, "ts": 6950381115676.760, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381167872.952, "dur": 37.599 , "args": { "External id": 2163,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7894, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7894, "pid": 0, "tid": 7, "ts": 6950381167872.952, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115695.474, "dur": 1.644 , "args": { "External id": 2163,"cbid": 307, "correlation": 7894 } }, { "ph": "s", "id": 7894, "pid": 1728719, "tid": 1728719, "ts": 6950381115695.474, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115705.704, "dur": 0.120 , "args": { "External id": 2164,"cbid": 200, "correlation": 7909 } }, { "ph": "f", "id": 7909, "pid": 1728719, "tid": 1728719, "ts": 6950381115705.704, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115705.924, "dur": 0.130 , "args": { "External id": 2164,"cbid": 468, "correlation": 7910 } }, { "ph": "f", "id": 7910, "pid": 1728719, "tid": 1728719, "ts": 6950381115705.924, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115706.275, "dur": 0.030 , "args": { "External id": 2164,"cbid": 1, "correlation": 7912 } }, { "ph": "f", "id": 7912, "pid": 1728719, "tid": 1728719, "ts": 6950381115706.275, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115706.365, "dur": 0.100 , "args": { "External id": 2164,"cbid": 468, "correlation": 7913 } }, { "ph": "f", "id": 7913, "pid": 1728719, "tid": 1728719, "ts": 6950381115706.365, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115706.625, "dur": 0.020 , "args": { "External id": 2164,"cbid": 1, "correlation": 7915 } }, { "ph": "f", "id": 7915, "pid": 1728719, "tid": 1728719, "ts": 6950381115706.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115706.745, "dur": 0.091 , "args": { "External id": 2164,"cbid": 468, "correlation": 7916 } }, { "ph": "f", "id": 7916, "pid": 1728719, "tid": 1728719, "ts": 6950381115706.745, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115706.986, "dur": 0.030 , "args": { "External id": 2164,"cbid": 1, "correlation": 7918 } }, { "ph": "f", "id": 7918, "pid": 1728719, "tid": 1728719, "ts": 6950381115706.986, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115707.096, "dur": 0.561 , "args": { "External id": 2164,"cbid": 273, "correlation": 7919 } }, { "ph": "f", "id": 7919, "pid": 1728719, "tid": 1728719, "ts": 6950381115707.096, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381167910.839, "dur": 227.454 , "args": { "External id": 2164,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7922, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7922, "pid": 0, "tid": 7, "ts": 6950381167910.839, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115708.028, "dur": 1.693 , "args": { "External id": 2164,"cbid": 211, "correlation": 7922 } }, { "ph": "s", "id": 7922, "pid": 1728719, "tid": 1728719, "ts": 6950381115708.028, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381168138.581, "dur": 8.928 , "args": { "External id": 2176,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7931, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7931, "pid": 0, "tid": 7, "ts": 6950381168138.581, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115715.602, "dur": 1.553 , "args": { "External id": 2176,"cbid": 307, "correlation": 7931 } }, { "ph": "s", "id": 7931, "pid": 1728719, "tid": 1728719, "ts": 6950381115715.602, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381168147.765, "dur": 25.983 , "args": { "External id": 2181,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7954, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7954, "pid": 0, "tid": 7, "ts": 6950381168147.765, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115731.091, "dur": 1.583 , "args": { "External id": 2181,"cbid": 307, "correlation": 7954 } }, { "ph": "s", "id": 7954, "pid": 1728719, "tid": 1728719, "ts": 6950381115731.091, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115741.180, "dur": 0.100 , "args": { "External id": 2182,"cbid": 200, "correlation": 7969 } }, { "ph": "f", "id": 7969, "pid": 1728719, "tid": 1728719, "ts": 6950381115741.180, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115741.370, "dur": 0.120 , "args": { "External id": 2182,"cbid": 468, "correlation": 7970 } }, { "ph": "f", "id": 7970, "pid": 1728719, "tid": 1728719, "ts": 6950381115741.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115741.690, "dur": 0.021 , "args": { "External id": 2182,"cbid": 1, "correlation": 7972 } }, { "ph": "f", "id": 7972, "pid": 1728719, "tid": 1728719, "ts": 6950381115741.690, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115741.781, "dur": 0.090 , "args": { "External id": 2182,"cbid": 468, "correlation": 7973 } }, { "ph": "f", "id": 7973, "pid": 1728719, "tid": 1728719, "ts": 6950381115741.781, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115742.021, "dur": 0.020 , "args": { "External id": 2182,"cbid": 1, "correlation": 7975 } }, { "ph": "f", "id": 7975, "pid": 1728719, "tid": 1728719, "ts": 6950381115742.021, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115742.131, "dur": 0.100 , "args": { "External id": 2182,"cbid": 468, "correlation": 7976 } }, { "ph": "f", "id": 7976, "pid": 1728719, "tid": 1728719, "ts": 6950381115742.131, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115742.392, "dur": 0.020 , "args": { "External id": 2182,"cbid": 1, "correlation": 7978 } }, { "ph": "f", "id": 7978, "pid": 1728719, "tid": 1728719, "ts": 6950381115742.392, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115742.492, "dur": 0.381 , "args": { "External id": 2182,"cbid": 273, "correlation": 7979 } }, { "ph": "f", "id": 7979, "pid": 1728719, "tid": 1728719, "ts": 6950381115742.492, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381168174.036, "dur": 229.822 , "args": { "External id": 2182,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7982, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7982, "pid": 0, "tid": 7, "ts": 6950381168174.036, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115743.223, "dur": 1.834 , "args": { "External id": 2182,"cbid": 211, "correlation": 7982 } }, { "ph": "s", "id": 7982, "pid": 1728719, "tid": 1728719, "ts": 6950381115743.223, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381168404.178, "dur": 60.351 , "args": { "External id": 2194,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 7996, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 7996, "pid": 0, "tid": 7, "ts": 6950381168404.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115752.461, "dur": 1.532 , "args": { "External id": 2194,"cbid": 307, "correlation": 7996 } }, { "ph": "s", "id": 7996, "pid": 1728719, "tid": 1728719, "ts": 6950381115752.461, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381168464.817, "dur": 64.320 , "args": { "External id": 2195,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8006, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8006, "pid": 0, "tid": 7, "ts": 6950381168464.817, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115762.159, "dur": 1.733 , "args": { "External id": 2195,"cbid": 307, "correlation": 8006 } }, { "ph": "s", "id": 8006, "pid": 1728719, "tid": 1728719, "ts": 6950381115762.159, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381168529.425, "dur": 39.455 , "args": { "External id": 2196,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8015, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8015, "pid": 0, "tid": 7, "ts": 6950381168529.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115771.807, "dur": 1.743 , "args": { "External id": 2196,"cbid": 307, "correlation": 8015 } }, { "ph": "s", "id": 8015, "pid": 1728719, "tid": 1728719, "ts": 6950381115771.807, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115787.195, "dur": 0.140 , "args": { "External id": 2204,"cbid": 200, "correlation": 8050 } }, { "ph": "f", "id": 8050, "pid": 1728719, "tid": 1728719, "ts": 6950381115787.195, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115787.416, "dur": 0.080 , "args": { "External id": 2204,"cbid": 200, "correlation": 8051 } }, { "ph": "f", "id": 8051, "pid": 1728719, "tid": 1728719, "ts": 6950381115787.416, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115787.616, "dur": 0.491 , "args": { "External id": 2204,"cbid": 273, "correlation": 8052 } }, { "ph": "f", "id": 8052, "pid": 1728719, "tid": 1728719, "ts": 6950381115787.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381168569.168, "dur": 851.447 , "args": { "External id": 2204,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8054, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8054, "pid": 0, "tid": 7, "ts": 6950381168569.168, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115788.498, "dur": 2.043 , "args": { "External id": 2204,"cbid": 211, "correlation": 8054 } }, { "ph": "s", "id": 8054, "pid": 1728719, "tid": 1728719, "ts": 6950381115788.498, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381169420.871, "dur": 9.216 , "args": { "External id": 2213,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8074, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8074, "pid": 0, "tid": 7, "ts": 6950381169420.871, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115804.768, "dur": 1.593 , "args": { "External id": 2213,"cbid": 652, "correlation": 8074 } }, { "ph": "s", "id": 8074, "pid": 1728719, "tid": 1728719, "ts": 6950381115804.768, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115815.588, "dur": 0.110 , "args": { "External id": 2214,"cbid": 200, "correlation": 8089 } }, { "ph": "f", "id": 8089, "pid": 1728719, "tid": 1728719, "ts": 6950381115815.588, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115815.818, "dur": 0.151 , "args": { "External id": 2214,"cbid": 468, "correlation": 8090 } }, { "ph": "f", "id": 8090, "pid": 1728719, "tid": 1728719, "ts": 6950381115815.818, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115816.229, "dur": 0.030 , "args": { "External id": 2214,"cbid": 1, "correlation": 8092 } }, { "ph": "f", "id": 8092, "pid": 1728719, "tid": 1728719, "ts": 6950381115816.229, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115816.339, "dur": 0.101 , "args": { "External id": 2214,"cbid": 468, "correlation": 8093 } }, { "ph": "f", "id": 8093, "pid": 1728719, "tid": 1728719, "ts": 6950381115816.339, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115816.590, "dur": 0.030 , "args": { "External id": 2214,"cbid": 1, "correlation": 8095 } }, { "ph": "f", "id": 8095, "pid": 1728719, "tid": 1728719, "ts": 6950381115816.590, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115816.710, "dur": 0.090 , "args": { "External id": 2214,"cbid": 468, "correlation": 8096 } }, { "ph": "f", "id": 8096, "pid": 1728719, "tid": 1728719, "ts": 6950381115816.710, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115816.981, "dur": 0.030 , "args": { "External id": 2214,"cbid": 1, "correlation": 8098 } }, { "ph": "f", "id": 8098, "pid": 1728719, "tid": 1728719, "ts": 6950381115816.981, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115817.091, "dur": 0.401 , "args": { "External id": 2214,"cbid": 273, "correlation": 8099 } }, { "ph": "f", "id": 8099, "pid": 1728719, "tid": 1728719, "ts": 6950381115817.091, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381169430.343, "dur": 207.198 , "args": { "External id": 2214,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8102, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8102, "pid": 0, "tid": 7, "ts": 6950381169430.343, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115817.892, "dur": 1.734 , "args": { "External id": 2214,"cbid": 211, "correlation": 8102 } }, { "ph": "s", "id": 8102, "pid": 1728719, "tid": 1728719, "ts": 6950381115817.892, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381169637.829, "dur": 46.463 , "args": { "External id": 2226,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8114, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8114, "pid": 0, "tid": 7, "ts": 6950381169637.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115827.580, "dur": 1.714 , "args": { "External id": 2226,"cbid": 307, "correlation": 8114 } }, { "ph": "s", "id": 8114, "pid": 1728719, "tid": 1728719, "ts": 6950381115827.580, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381169684.612, "dur": 13.120 , "args": { "External id": 2227,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8120, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8120, "pid": 0, "tid": 7, "ts": 6950381169684.612, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115836.116, "dur": 1.503 , "args": { "External id": 2227,"cbid": 652, "correlation": 8120 } }, { "ph": "s", "id": 8120, "pid": 1728719, "tid": 1728719, "ts": 6950381115836.116, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381169698.052, "dur": 13.184 , "args": { "External id": 2228,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8126, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8126, "pid": 0, "tid": 7, "ts": 6950381169698.052, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115844.432, "dur": 1.683 , "args": { "External id": 2228,"cbid": 652, "correlation": 8126 } }, { "ph": "s", "id": 8126, "pid": 1728719, "tid": 1728719, "ts": 6950381115844.432, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115856.184, "dur": 0.100 , "args": { "External id": 2229,"cbid": 200, "correlation": 8145 } }, { "ph": "f", "id": 8145, "pid": 1728719, "tid": 1728719, "ts": 6950381115856.184, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115856.394, "dur": 0.120 , "args": { "External id": 2229,"cbid": 468, "correlation": 8146 } }, { "ph": "f", "id": 8146, "pid": 1728719, "tid": 1728719, "ts": 6950381115856.394, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115856.725, "dur": 0.020 , "args": { "External id": 2229,"cbid": 1, "correlation": 8148 } }, { "ph": "f", "id": 8148, "pid": 1728719, "tid": 1728719, "ts": 6950381115856.725, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115856.825, "dur": 0.100 , "args": { "External id": 2229,"cbid": 468, "correlation": 8149 } }, { "ph": "f", "id": 8149, "pid": 1728719, "tid": 1728719, "ts": 6950381115856.825, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115857.085, "dur": 0.030 , "args": { "External id": 2229,"cbid": 1, "correlation": 8151 } }, { "ph": "f", "id": 8151, "pid": 1728719, "tid": 1728719, "ts": 6950381115857.085, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115857.216, "dur": 0.100 , "args": { "External id": 2229,"cbid": 468, "correlation": 8152 } }, { "ph": "f", "id": 8152, "pid": 1728719, "tid": 1728719, "ts": 6950381115857.216, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115857.476, "dur": 0.030 , "args": { "External id": 2229,"cbid": 1, "correlation": 8154 } }, { "ph": "f", "id": 8154, "pid": 1728719, "tid": 1728719, "ts": 6950381115857.476, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115857.586, "dur": 0.401 , "args": { "External id": 2229,"cbid": 273, "correlation": 8155 } }, { "ph": "f", "id": 8155, "pid": 1728719, "tid": 1728719, "ts": 6950381115857.586, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381169711.556, "dur": 664.344 , "args": { "External id": 2229,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8158, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8158, "pid": 0, "tid": 7, "ts": 6950381169711.556, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115858.348, "dur": 1.703 , "args": { "External id": 2229,"cbid": 211, "correlation": 8158 } }, { "ph": "s", "id": 8158, "pid": 1728719, "tid": 1728719, "ts": 6950381115858.348, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115869.969, "dur": 0.080 , "args": { "External id": 2241,"cbid": 200, "correlation": 8182 } }, { "ph": "f", "id": 8182, "pid": 1728719, "tid": 1728719, "ts": 6950381115869.969, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115870.150, "dur": 0.110 , "args": { "External id": 2241,"cbid": 468, "correlation": 8183 } }, { "ph": "f", "id": 8183, "pid": 1728719, "tid": 1728719, "ts": 6950381115870.150, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115870.460, "dur": 0.020 , "args": { "External id": 2241,"cbid": 1, "correlation": 8185 } }, { "ph": "f", "id": 8185, "pid": 1728719, "tid": 1728719, "ts": 6950381115870.460, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115870.560, "dur": 0.101 , "args": { "External id": 2241,"cbid": 468, "correlation": 8186 } }, { "ph": "f", "id": 8186, "pid": 1728719, "tid": 1728719, "ts": 6950381115870.560, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115870.801, "dur": 0.030 , "args": { "External id": 2241,"cbid": 1, "correlation": 8188 } }, { "ph": "f", "id": 8188, "pid": 1728719, "tid": 1728719, "ts": 6950381115870.801, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115870.931, "dur": 0.100 , "args": { "External id": 2241,"cbid": 468, "correlation": 8189 } }, { "ph": "f", "id": 8189, "pid": 1728719, "tid": 1728719, "ts": 6950381115870.931, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115871.192, "dur": 0.020 , "args": { "External id": 2241,"cbid": 1, "correlation": 8191 } }, { "ph": "f", "id": 8191, "pid": 1728719, "tid": 1728719, "ts": 6950381115871.192, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115871.292, "dur": 0.320 , "args": { "External id": 2241,"cbid": 273, "correlation": 8192 } }, { "ph": "f", "id": 8192, "pid": 1728719, "tid": 1728719, "ts": 6950381115871.292, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381170376.189, "dur": 673.816 , "args": { "External id": 2241,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8195, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8195, "pid": 0, "tid": 7, "ts": 6950381170376.189, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115871.983, "dur": 1.633 , "args": { "External id": 2241,"cbid": 211, "correlation": 8195 } }, { "ph": "s", "id": 8195, "pid": 1728719, "tid": 1728719, "ts": 6950381115871.983, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381171050.325, "dur": 154.879 , "args": { "External id": 2253,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8208, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8208, "pid": 0, "tid": 7, "ts": 6950381171050.325, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115881.000, "dur": 1.613 , "args": { "External id": 2253,"cbid": 307, "correlation": 8208 } }, { "ph": "s", "id": 8208, "pid": 1728719, "tid": 1728719, "ts": 6950381115881.000, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381171205.459, "dur": 57.664 , "args": { "External id": 2254,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8214, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8214, "pid": 0, "tid": 7, "ts": 6950381171205.459, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381115889.105, "dur": 1.583 , "args": { "External id": 2254,"cbid": 652, "correlation": 8214 } }, { "ph": "s", "id": 8214, "pid": 1728719, "tid": 1728719, "ts": 6950381115889.105, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115900.326, "dur": 0.100 , "args": { "External id": 2255,"cbid": 200, "correlation": 8233 } }, { "ph": "f", "id": 8233, "pid": 1728719, "tid": 1728719, "ts": 6950381115900.326, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115900.536, "dur": 0.120 , "args": { "External id": 2255,"cbid": 468, "correlation": 8234 } }, { "ph": "f", "id": 8234, "pid": 1728719, "tid": 1728719, "ts": 6950381115900.536, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115900.847, "dur": 0.020 , "args": { "External id": 2255,"cbid": 1, "correlation": 8236 } }, { "ph": "f", "id": 8236, "pid": 1728719, "tid": 1728719, "ts": 6950381115900.847, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115900.937, "dur": 0.090 , "args": { "External id": 2255,"cbid": 468, "correlation": 8237 } }, { "ph": "f", "id": 8237, "pid": 1728719, "tid": 1728719, "ts": 6950381115900.937, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115901.167, "dur": 0.030 , "args": { "External id": 2255,"cbid": 1, "correlation": 8239 } }, { "ph": "f", "id": 8239, "pid": 1728719, "tid": 1728719, "ts": 6950381115901.167, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115901.288, "dur": 0.090 , "args": { "External id": 2255,"cbid": 468, "correlation": 8240 } }, { "ph": "f", "id": 8240, "pid": 1728719, "tid": 1728719, "ts": 6950381115901.288, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381115901.528, "dur": 0.020 , "args": { "External id": 2255,"cbid": 1, "correlation": 8242 } }, { "ph": "f", "id": 8242, "pid": 1728719, "tid": 1728719, "ts": 6950381115901.528, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381115901.638, "dur": 0.351 , "args": { "External id": 2255,"cbid": 273, "correlation": 8243 } }, { "ph": "f", "id": 8243, "pid": 1728719, "tid": 1728719, "ts": 6950381115901.638, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381171263.379, "dur": 675.865 , "args": { "External id": 2255,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8246, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8246, "pid": 0, "tid": 7, "ts": 6950381171263.379, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115902.360, "dur": 1.663 , "args": { "External id": 2255,"cbid": 211, "correlation": 8246 } }, { "ph": "s", "id": 8246, "pid": 1728719, "tid": 1728719, "ts": 6950381115902.360, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381171939.532, "dur": 39.871 , "args": { "External id": 2267,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8257, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8257, "pid": 0, "tid": 7, "ts": 6950381171939.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115911.937, "dur": 1.603 , "args": { "External id": 2267,"cbid": 307, "correlation": 8257 } }, { "ph": "s", "id": 8257, "pid": 1728719, "tid": 1728719, "ts": 6950381115911.937, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381171979.691, "dur": 20.640 , "args": { "External id": 2279,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8269, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8269, "pid": 0, "tid": 7, "ts": 6950381171979.691, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381115990.954, "dur": 2.345 , "args": { "External id": 2279,"cbid": 307, "correlation": 8269 } }, { "ph": "s", "id": 8269, "pid": 1728719, "tid": 1728719, "ts": 6950381115990.954, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381172000.651, "dur": 13.024 , "args": { "External id": 2280,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8283, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8283, "pid": 0, "tid": 7, "ts": 6950381172000.651, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116004.800, "dur": 1.563 , "args": { "External id": 2280,"cbid": 652, "correlation": 8283 } }, { "ph": "s", "id": 8283, "pid": 1728719, "tid": 1728719, "ts": 6950381116004.800, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381172013.963, "dur": 12.384 , "args": { "External id": 2281,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8297, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8297, "pid": 0, "tid": 7, "ts": 6950381172013.963, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116016.292, "dur": 1.603 , "args": { "External id": 2281,"cbid": 652, "correlation": 8297 } }, { "ph": "s", "id": 8297, "pid": 1728719, "tid": 1728719, "ts": 6950381116016.292, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381172026.667, "dur": 10.687 , "args": { "External id": 2282,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8311, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8311, "pid": 0, "tid": 7, "ts": 6950381172026.667, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116026.671, "dur": 1.463 , "args": { "External id": 2282,"cbid": 652, "correlation": 8311 } }, { "ph": "s", "id": 8311, "pid": 1728719, "tid": 1728719, "ts": 6950381116026.671, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116037.922, "dur": 0.150 , "args": { "External id": 2283,"cbid": 200, "correlation": 8326 } }, { "ph": "f", "id": 8326, "pid": 1728719, "tid": 1728719, "ts": 6950381116037.922, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116038.212, "dur": 0.191 , "args": { "External id": 2283,"cbid": 468, "correlation": 8327 } }, { "ph": "f", "id": 8327, "pid": 1728719, "tid": 1728719, "ts": 6950381116038.212, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116038.703, "dur": 0.040 , "args": { "External id": 2283,"cbid": 1, "correlation": 8329 } }, { "ph": "f", "id": 8329, "pid": 1728719, "tid": 1728719, "ts": 6950381116038.703, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116038.814, "dur": 0.110 , "args": { "External id": 2283,"cbid": 468, "correlation": 8330 } }, { "ph": "f", "id": 8330, "pid": 1728719, "tid": 1728719, "ts": 6950381116038.814, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116039.084, "dur": 0.020 , "args": { "External id": 2283,"cbid": 1, "correlation": 8332 } }, { "ph": "f", "id": 8332, "pid": 1728719, "tid": 1728719, "ts": 6950381116039.084, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116039.224, "dur": 0.101 , "args": { "External id": 2283,"cbid": 468, "correlation": 8333 } }, { "ph": "f", "id": 8333, "pid": 1728719, "tid": 1728719, "ts": 6950381116039.224, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116039.485, "dur": 0.020 , "args": { "External id": 2283,"cbid": 1, "correlation": 8335 } }, { "ph": "f", "id": 8335, "pid": 1728719, "tid": 1728719, "ts": 6950381116039.485, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116039.635, "dur": 0.631 , "args": { "External id": 2283,"cbid": 273, "correlation": 8336 } }, { "ph": "f", "id": 8336, "pid": 1728719, "tid": 1728719, "ts": 6950381116039.635, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381172037.642, "dur": 214.686 , "args": { "External id": 2283,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8339, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8339, "pid": 0, "tid": 7, "ts": 6950381172037.642, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116040.707, "dur": 1.823 , "args": { "External id": 2283,"cbid": 211, "correlation": 8339 } }, { "ph": "s", "id": 8339, "pid": 1728719, "tid": 1728719, "ts": 6950381116040.707, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381172252.648, "dur": 8.768 , "args": { "External id": 2295,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8352, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8352, "pid": 0, "tid": 7, "ts": 6950381172252.648, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116051.106, "dur": 1.674 , "args": { "External id": 2295,"cbid": 307, "correlation": 8352 } }, { "ph": "s", "id": 8352, "pid": 1728719, "tid": 1728719, "ts": 6950381116051.106, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381172261.704, "dur": 36.832 , "args": { "External id": 2300,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8375, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8375, "pid": 0, "tid": 7, "ts": 6950381172261.704, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116069.701, "dur": 1.643 , "args": { "External id": 2300,"cbid": 307, "correlation": 8375 } }, { "ph": "s", "id": 8375, "pid": 1728719, "tid": 1728719, "ts": 6950381116069.701, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116080.231, "dur": 0.100 , "args": { "External id": 2301,"cbid": 200, "correlation": 8390 } }, { "ph": "f", "id": 8390, "pid": 1728719, "tid": 1728719, "ts": 6950381116080.231, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116080.441, "dur": 0.140 , "args": { "External id": 2301,"cbid": 468, "correlation": 8391 } }, { "ph": "f", "id": 8391, "pid": 1728719, "tid": 1728719, "ts": 6950381116080.441, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116080.812, "dur": 0.030 , "args": { "External id": 2301,"cbid": 1, "correlation": 8393 } }, { "ph": "f", "id": 8393, "pid": 1728719, "tid": 1728719, "ts": 6950381116080.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116080.922, "dur": 0.100 , "args": { "External id": 2301,"cbid": 468, "correlation": 8394 } }, { "ph": "f", "id": 8394, "pid": 1728719, "tid": 1728719, "ts": 6950381116080.922, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116081.162, "dur": 0.030 , "args": { "External id": 2301,"cbid": 1, "correlation": 8396 } }, { "ph": "f", "id": 8396, "pid": 1728719, "tid": 1728719, "ts": 6950381116081.162, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116081.283, "dur": 0.090 , "args": { "External id": 2301,"cbid": 468, "correlation": 8397 } }, { "ph": "f", "id": 8397, "pid": 1728719, "tid": 1728719, "ts": 6950381116081.283, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116081.533, "dur": 0.020 , "args": { "External id": 2301,"cbid": 1, "correlation": 8399 } }, { "ph": "f", "id": 8399, "pid": 1728719, "tid": 1728719, "ts": 6950381116081.533, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116081.643, "dur": 0.411 , "args": { "External id": 2301,"cbid": 273, "correlation": 8400 } }, { "ph": "f", "id": 8400, "pid": 1728719, "tid": 1728719, "ts": 6950381116081.643, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381172298.824, "dur": 228.061 , "args": { "External id": 2301,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8403, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8403, "pid": 0, "tid": 7, "ts": 6950381172298.824, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116082.445, "dur": 1.713 , "args": { "External id": 2301,"cbid": 211, "correlation": 8403 } }, { "ph": "s", "id": 8403, "pid": 1728719, "tid": 1728719, "ts": 6950381116082.445, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381172527.173, "dur": 8.832 , "args": { "External id": 2313,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8412, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8412, "pid": 0, "tid": 7, "ts": 6950381172527.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116089.979, "dur": 1.563 , "args": { "External id": 2313,"cbid": 307, "correlation": 8412 } }, { "ph": "s", "id": 8412, "pid": 1728719, "tid": 1728719, "ts": 6950381116089.979, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381172536.293, "dur": 28.000 , "args": { "External id": 2318,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8435, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8435, "pid": 0, "tid": 7, "ts": 6950381172536.293, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116105.418, "dur": 1.522 , "args": { "External id": 2318,"cbid": 307, "correlation": 8435 } }, { "ph": "s", "id": 8435, "pid": 1728719, "tid": 1728719, "ts": 6950381116105.418, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116115.266, "dur": 0.100 , "args": { "External id": 2319,"cbid": 200, "correlation": 8450 } }, { "ph": "f", "id": 8450, "pid": 1728719, "tid": 1728719, "ts": 6950381116115.266, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116115.446, "dur": 0.130 , "args": { "External id": 2319,"cbid": 468, "correlation": 8451 } }, { "ph": "f", "id": 8451, "pid": 1728719, "tid": 1728719, "ts": 6950381116115.446, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116115.777, "dur": 0.020 , "args": { "External id": 2319,"cbid": 1, "correlation": 8453 } }, { "ph": "f", "id": 8453, "pid": 1728719, "tid": 1728719, "ts": 6950381116115.777, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116115.867, "dur": 0.100 , "args": { "External id": 2319,"cbid": 468, "correlation": 8454 } }, { "ph": "f", "id": 8454, "pid": 1728719, "tid": 1728719, "ts": 6950381116115.867, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116116.117, "dur": 0.020 , "args": { "External id": 2319,"cbid": 1, "correlation": 8456 } }, { "ph": "f", "id": 8456, "pid": 1728719, "tid": 1728719, "ts": 6950381116116.117, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116116.228, "dur": 0.090 , "args": { "External id": 2319,"cbid": 468, "correlation": 8457 } }, { "ph": "f", "id": 8457, "pid": 1728719, "tid": 1728719, "ts": 6950381116116.228, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116116.488, "dur": 0.020 , "args": { "External id": 2319,"cbid": 1, "correlation": 8459 } }, { "ph": "f", "id": 8459, "pid": 1728719, "tid": 1728719, "ts": 6950381116116.488, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116116.598, "dur": 0.421 , "args": { "External id": 2319,"cbid": 273, "correlation": 8460 } }, { "ph": "f", "id": 8460, "pid": 1728719, "tid": 1728719, "ts": 6950381116116.598, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381172564.581, "dur": 232.925 , "args": { "External id": 2319,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8463, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8463, "pid": 0, "tid": 7, "ts": 6950381172564.581, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116117.380, "dur": 1.673 , "args": { "External id": 2319,"cbid": 211, "correlation": 8463 } }, { "ph": "s", "id": 8463, "pid": 1728719, "tid": 1728719, "ts": 6950381116117.380, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381172797.794, "dur": 59.680 , "args": { "External id": 2331,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8477, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8477, "pid": 0, "tid": 7, "ts": 6950381172797.794, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116126.517, "dur": 1.543 , "args": { "External id": 2331,"cbid": 307, "correlation": 8477 } }, { "ph": "s", "id": 8477, "pid": 1728719, "tid": 1728719, "ts": 6950381116126.517, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381172857.762, "dur": 65.247 , "args": { "External id": 2332,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8487, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8487, "pid": 0, "tid": 7, "ts": 6950381172857.762, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116136.836, "dur": 1.713 , "args": { "External id": 2332,"cbid": 307, "correlation": 8487 } }, { "ph": "s", "id": 8487, "pid": 1728719, "tid": 1728719, "ts": 6950381116136.836, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381172923.297, "dur": 42.528 , "args": { "External id": 2333,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8496, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8496, "pid": 0, "tid": 7, "ts": 6950381172923.297, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116146.965, "dur": 1.613 , "args": { "External id": 2333,"cbid": 307, "correlation": 8496 } }, { "ph": "s", "id": 8496, "pid": 1728719, "tid": 1728719, "ts": 6950381116146.965, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116161.722, "dur": 0.121 , "args": { "External id": 2341,"cbid": 200, "correlation": 8531 } }, { "ph": "f", "id": 8531, "pid": 1728719, "tid": 1728719, "ts": 6950381116161.722, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116161.903, "dur": 0.150 , "args": { "External id": 2341,"cbid": 200, "correlation": 8532 } }, { "ph": "f", "id": 8532, "pid": 1728719, "tid": 1728719, "ts": 6950381116161.903, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116162.153, "dur": 0.501 , "args": { "External id": 2341,"cbid": 273, "correlation": 8533 } }, { "ph": "f", "id": 8533, "pid": 1728719, "tid": 1728719, "ts": 6950381116162.153, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381172966.113, "dur": 850.614 , "args": { "External id": 2341,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8535, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8535, "pid": 0, "tid": 7, "ts": 6950381172966.113, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116163.025, "dur": 1.753 , "args": { "External id": 2341,"cbid": 211, "correlation": 8535 } }, { "ph": "s", "id": 8535, "pid": 1728719, "tid": 1728719, "ts": 6950381116163.025, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381173816.983, "dur": 9.440 , "args": { "External id": 2350,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8555, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8555, "pid": 0, "tid": 7, "ts": 6950381173816.983, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116178.594, "dur": 1.593 , "args": { "External id": 2350,"cbid": 652, "correlation": 8555 } }, { "ph": "s", "id": 8555, "pid": 1728719, "tid": 1728719, "ts": 6950381116178.594, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116189.354, "dur": 0.110 , "args": { "External id": 2351,"cbid": 200, "correlation": 8570 } }, { "ph": "f", "id": 8570, "pid": 1728719, "tid": 1728719, "ts": 6950381116189.354, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116189.564, "dur": 0.130 , "args": { "External id": 2351,"cbid": 468, "correlation": 8571 } }, { "ph": "f", "id": 8571, "pid": 1728719, "tid": 1728719, "ts": 6950381116189.564, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116189.945, "dur": 0.030 , "args": { "External id": 2351,"cbid": 1, "correlation": 8573 } }, { "ph": "f", "id": 8573, "pid": 1728719, "tid": 1728719, "ts": 6950381116189.945, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116190.045, "dur": 0.100 , "args": { "External id": 2351,"cbid": 468, "correlation": 8574 } }, { "ph": "f", "id": 8574, "pid": 1728719, "tid": 1728719, "ts": 6950381116190.045, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116190.316, "dur": 0.020 , "args": { "External id": 2351,"cbid": 1, "correlation": 8576 } }, { "ph": "f", "id": 8576, "pid": 1728719, "tid": 1728719, "ts": 6950381116190.316, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116190.426, "dur": 0.100 , "args": { "External id": 2351,"cbid": 468, "correlation": 8577 } }, { "ph": "f", "id": 8577, "pid": 1728719, "tid": 1728719, "ts": 6950381116190.426, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116190.676, "dur": 0.020 , "args": { "External id": 2351,"cbid": 1, "correlation": 8579 } }, { "ph": "f", "id": 8579, "pid": 1728719, "tid": 1728719, "ts": 6950381116190.676, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116190.786, "dur": 0.381 , "args": { "External id": 2351,"cbid": 273, "correlation": 8580 } }, { "ph": "f", "id": 8580, "pid": 1728719, "tid": 1728719, "ts": 6950381116190.786, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381173826.679, "dur": 208.990 , "args": { "External id": 2351,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8583, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8583, "pid": 0, "tid": 7, "ts": 6950381173826.679, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116191.558, "dur": 1.733 , "args": { "External id": 2351,"cbid": 211, "correlation": 8583 } }, { "ph": "s", "id": 8583, "pid": 1728719, "tid": 1728719, "ts": 6950381116191.558, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381174035.989, "dur": 45.823 , "args": { "External id": 2363,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8595, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8595, "pid": 0, "tid": 7, "ts": 6950381174035.989, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116201.366, "dur": 1.683 , "args": { "External id": 2363,"cbid": 307, "correlation": 8595 } }, { "ph": "s", "id": 8595, "pid": 1728719, "tid": 1728719, "ts": 6950381116201.366, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381174082.100, "dur": 13.760 , "args": { "External id": 2364,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8601, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8601, "pid": 0, "tid": 7, "ts": 6950381174082.100, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116209.211, "dur": 1.733 , "args": { "External id": 2364,"cbid": 652, "correlation": 8601 } }, { "ph": "s", "id": 8601, "pid": 1728719, "tid": 1728719, "ts": 6950381116209.211, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381174096.180, "dur": 12.896 , "args": { "External id": 2365,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8607, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8607, "pid": 0, "tid": 7, "ts": 6950381174096.180, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116217.957, "dur": 1.493 , "args": { "External id": 2365,"cbid": 652, "correlation": 8607 } }, { "ph": "s", "id": 8607, "pid": 1728719, "tid": 1728719, "ts": 6950381116217.957, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116229.549, "dur": 0.100 , "args": { "External id": 2366,"cbid": 200, "correlation": 8626 } }, { "ph": "f", "id": 8626, "pid": 1728719, "tid": 1728719, "ts": 6950381116229.549, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116229.749, "dur": 0.110 , "args": { "External id": 2366,"cbid": 468, "correlation": 8627 } }, { "ph": "f", "id": 8627, "pid": 1728719, "tid": 1728719, "ts": 6950381116229.749, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116230.060, "dur": 0.030 , "args": { "External id": 2366,"cbid": 1, "correlation": 8629 } }, { "ph": "f", "id": 8629, "pid": 1728719, "tid": 1728719, "ts": 6950381116230.060, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116230.160, "dur": 0.100 , "args": { "External id": 2366,"cbid": 468, "correlation": 8630 } }, { "ph": "f", "id": 8630, "pid": 1728719, "tid": 1728719, "ts": 6950381116230.160, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116230.420, "dur": 0.020 , "args": { "External id": 2366,"cbid": 1, "correlation": 8632 } }, { "ph": "f", "id": 8632, "pid": 1728719, "tid": 1728719, "ts": 6950381116230.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116230.530, "dur": 0.101 , "args": { "External id": 2366,"cbid": 468, "correlation": 8633 } }, { "ph": "f", "id": 8633, "pid": 1728719, "tid": 1728719, "ts": 6950381116230.530, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116230.801, "dur": 0.020 , "args": { "External id": 2366,"cbid": 1, "correlation": 8635 } }, { "ph": "f", "id": 8635, "pid": 1728719, "tid": 1728719, "ts": 6950381116230.801, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116230.921, "dur": 0.371 , "args": { "External id": 2366,"cbid": 273, "correlation": 8636 } }, { "ph": "f", "id": 8636, "pid": 1728719, "tid": 1728719, "ts": 6950381116230.921, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381174109.364, "dur": 661.273 , "args": { "External id": 2366,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8639, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8639, "pid": 0, "tid": 7, "ts": 6950381174109.364, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116231.693, "dur": 1.713 , "args": { "External id": 2366,"cbid": 211, "correlation": 8639 } }, { "ph": "s", "id": 8639, "pid": 1728719, "tid": 1728719, "ts": 6950381116231.693, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116243.475, "dur": 0.080 , "args": { "External id": 2378,"cbid": 200, "correlation": 8663 } }, { "ph": "f", "id": 8663, "pid": 1728719, "tid": 1728719, "ts": 6950381116243.475, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116243.645, "dur": 0.110 , "args": { "External id": 2378,"cbid": 468, "correlation": 8664 } }, { "ph": "f", "id": 8664, "pid": 1728719, "tid": 1728719, "ts": 6950381116243.645, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116243.935, "dur": 0.030 , "args": { "External id": 2378,"cbid": 1, "correlation": 8666 } }, { "ph": "f", "id": 8666, "pid": 1728719, "tid": 1728719, "ts": 6950381116243.935, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116244.026, "dur": 0.100 , "args": { "External id": 2378,"cbid": 468, "correlation": 8667 } }, { "ph": "f", "id": 8667, "pid": 1728719, "tid": 1728719, "ts": 6950381116244.026, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116244.266, "dur": 0.030 , "args": { "External id": 2378,"cbid": 1, "correlation": 8669 } }, { "ph": "f", "id": 8669, "pid": 1728719, "tid": 1728719, "ts": 6950381116244.266, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116244.386, "dur": 0.090 , "args": { "External id": 2378,"cbid": 468, "correlation": 8670 } }, { "ph": "f", "id": 8670, "pid": 1728719, "tid": 1728719, "ts": 6950381116244.386, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116244.637, "dur": 0.020 , "args": { "External id": 2378,"cbid": 1, "correlation": 8672 } }, { "ph": "f", "id": 8672, "pid": 1728719, "tid": 1728719, "ts": 6950381116244.637, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116244.737, "dur": 0.341 , "args": { "External id": 2378,"cbid": 273, "correlation": 8673 } }, { "ph": "f", "id": 8673, "pid": 1728719, "tid": 1728719, "ts": 6950381116244.737, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381174770.925, "dur": 666.937 , "args": { "External id": 2378,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8676, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8676, "pid": 0, "tid": 7, "ts": 6950381174770.925, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116245.428, "dur": 1.693 , "args": { "External id": 2378,"cbid": 211, "correlation": 8676 } }, { "ph": "s", "id": 8676, "pid": 1728719, "tid": 1728719, "ts": 6950381116245.428, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381175438.182, "dur": 154.526 , "args": { "External id": 2390,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8689, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8689, "pid": 0, "tid": 7, "ts": 6950381175438.182, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116254.565, "dur": 1.603 , "args": { "External id": 2390,"cbid": 307, "correlation": 8689 } }, { "ph": "s", "id": 8689, "pid": 1728719, "tid": 1728719, "ts": 6950381116254.565, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381175592.964, "dur": 58.623 , "args": { "External id": 2391,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8695, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8695, "pid": 0, "tid": 7, "ts": 6950381175592.964, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116262.560, "dur": 1.593 , "args": { "External id": 2391,"cbid": 652, "correlation": 8695 } }, { "ph": "s", "id": 8695, "pid": 1728719, "tid": 1728719, "ts": 6950381116262.560, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116273.791, "dur": 0.080 , "args": { "External id": 2392,"cbid": 200, "correlation": 8714 } }, { "ph": "f", "id": 8714, "pid": 1728719, "tid": 1728719, "ts": 6950381116273.791, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116273.961, "dur": 0.120 , "args": { "External id": 2392,"cbid": 468, "correlation": 8715 } }, { "ph": "f", "id": 8715, "pid": 1728719, "tid": 1728719, "ts": 6950381116273.961, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116274.302, "dur": 0.020 , "args": { "External id": 2392,"cbid": 1, "correlation": 8717 } }, { "ph": "f", "id": 8717, "pid": 1728719, "tid": 1728719, "ts": 6950381116274.302, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116274.392, "dur": 0.090 , "args": { "External id": 2392,"cbid": 468, "correlation": 8718 } }, { "ph": "f", "id": 8718, "pid": 1728719, "tid": 1728719, "ts": 6950381116274.392, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116274.633, "dur": 0.020 , "args": { "External id": 2392,"cbid": 1, "correlation": 8720 } }, { "ph": "f", "id": 8720, "pid": 1728719, "tid": 1728719, "ts": 6950381116274.633, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116274.753, "dur": 0.100 , "args": { "External id": 2392,"cbid": 468, "correlation": 8721 } }, { "ph": "f", "id": 8721, "pid": 1728719, "tid": 1728719, "ts": 6950381116274.753, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116275.003, "dur": 0.020 , "args": { "External id": 2392,"cbid": 1, "correlation": 8723 } }, { "ph": "f", "id": 8723, "pid": 1728719, "tid": 1728719, "ts": 6950381116275.003, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116275.113, "dur": 0.371 , "args": { "External id": 2392,"cbid": 273, "correlation": 8724 } }, { "ph": "f", "id": 8724, "pid": 1728719, "tid": 1728719, "ts": 6950381116275.113, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381175651.875, "dur": 679.257 , "args": { "External id": 2392,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8727, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8727, "pid": 0, "tid": 7, "ts": 6950381175651.875, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116275.835, "dur": 1.683 , "args": { "External id": 2392,"cbid": 211, "correlation": 8727 } }, { "ph": "s", "id": 8727, "pid": 1728719, "tid": 1728719, "ts": 6950381116275.835, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381176331.420, "dur": 40.032 , "args": { "External id": 2404,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8738, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8738, "pid": 0, "tid": 7, "ts": 6950381176331.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116285.773, "dur": 1.603 , "args": { "External id": 2404,"cbid": 307, "correlation": 8738 } }, { "ph": "s", "id": 8738, "pid": 1728719, "tid": 1728719, "ts": 6950381116285.773, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381176371.740, "dur": 21.119 , "args": { "External id": 2416,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8750, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8750, "pid": 0, "tid": 7, "ts": 6950381176371.740, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116365.412, "dur": 2.414 , "args": { "External id": 2416,"cbid": 307, "correlation": 8750 } }, { "ph": "s", "id": 8750, "pid": 1728719, "tid": 1728719, "ts": 6950381116365.412, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381176393.179, "dur": 12.512 , "args": { "External id": 2417,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8764, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8764, "pid": 0, "tid": 7, "ts": 6950381176393.179, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116379.237, "dur": 1.603 , "args": { "External id": 2417,"cbid": 652, "correlation": 8764 } }, { "ph": "s", "id": 8764, "pid": 1728719, "tid": 1728719, "ts": 6950381116379.237, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381176406.011, "dur": 12.448 , "args": { "External id": 2418,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8778, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8778, "pid": 0, "tid": 7, "ts": 6950381176406.011, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116390.909, "dur": 1.603 , "args": { "External id": 2418,"cbid": 652, "correlation": 8778 } }, { "ph": "s", "id": 8778, "pid": 1728719, "tid": 1728719, "ts": 6950381116390.909, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381176418.747, "dur": 10.688 , "args": { "External id": 2419,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8792, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8792, "pid": 0, "tid": 7, "ts": 6950381176418.747, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116401.729, "dur": 1.483 , "args": { "External id": 2419,"cbid": 652, "correlation": 8792 } }, { "ph": "s", "id": 8792, "pid": 1728719, "tid": 1728719, "ts": 6950381116401.729, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116413.140, "dur": 0.141 , "args": { "External id": 2420,"cbid": 200, "correlation": 8807 } }, { "ph": "f", "id": 8807, "pid": 1728719, "tid": 1728719, "ts": 6950381116413.140, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116413.411, "dur": 0.210 , "args": { "External id": 2420,"cbid": 468, "correlation": 8808 } }, { "ph": "f", "id": 8808, "pid": 1728719, "tid": 1728719, "ts": 6950381116413.411, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116413.942, "dur": 0.030 , "args": { "External id": 2420,"cbid": 1, "correlation": 8810 } }, { "ph": "f", "id": 8810, "pid": 1728719, "tid": 1728719, "ts": 6950381116413.942, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116414.042, "dur": 0.110 , "args": { "External id": 2420,"cbid": 468, "correlation": 8811 } }, { "ph": "f", "id": 8811, "pid": 1728719, "tid": 1728719, "ts": 6950381116414.042, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116414.303, "dur": 0.020 , "args": { "External id": 2420,"cbid": 1, "correlation": 8813 } }, { "ph": "f", "id": 8813, "pid": 1728719, "tid": 1728719, "ts": 6950381116414.303, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116414.433, "dur": 0.100 , "args": { "External id": 2420,"cbid": 468, "correlation": 8814 } }, { "ph": "f", "id": 8814, "pid": 1728719, "tid": 1728719, "ts": 6950381116414.433, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116414.693, "dur": 0.020 , "args": { "External id": 2420,"cbid": 1, "correlation": 8816 } }, { "ph": "f", "id": 8816, "pid": 1728719, "tid": 1728719, "ts": 6950381116414.693, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116414.854, "dur": 0.631 , "args": { "External id": 2420,"cbid": 273, "correlation": 8817 } }, { "ph": "f", "id": 8817, "pid": 1728719, "tid": 1728719, "ts": 6950381116414.854, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381176429.723, "dur": 215.166 , "args": { "External id": 2420,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8820, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8820, "pid": 0, "tid": 7, "ts": 6950381176429.723, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116415.946, "dur": 1.983 , "args": { "External id": 2420,"cbid": 211, "correlation": 8820 } }, { "ph": "s", "id": 8820, "pid": 1728719, "tid": 1728719, "ts": 6950381116415.946, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381176645.177, "dur": 8.800 , "args": { "External id": 2432,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8833, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8833, "pid": 0, "tid": 7, "ts": 6950381176645.177, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116426.565, "dur": 1.653 , "args": { "External id": 2432,"cbid": 307, "correlation": 8833 } }, { "ph": "s", "id": 8833, "pid": 1728719, "tid": 1728719, "ts": 6950381116426.565, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381176654.233, "dur": 37.183 , "args": { "External id": 2437,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8856, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8856, "pid": 0, "tid": 7, "ts": 6950381176654.233, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116445.240, "dur": 1.653 , "args": { "External id": 2437,"cbid": 307, "correlation": 8856 } }, { "ph": "s", "id": 8856, "pid": 1728719, "tid": 1728719, "ts": 6950381116445.240, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116455.549, "dur": 0.100 , "args": { "External id": 2438,"cbid": 200, "correlation": 8871 } }, { "ph": "f", "id": 8871, "pid": 1728719, "tid": 1728719, "ts": 6950381116455.549, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116455.750, "dur": 0.130 , "args": { "External id": 2438,"cbid": 468, "correlation": 8872 } }, { "ph": "f", "id": 8872, "pid": 1728719, "tid": 1728719, "ts": 6950381116455.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116456.130, "dur": 0.030 , "args": { "External id": 2438,"cbid": 1, "correlation": 8874 } }, { "ph": "f", "id": 8874, "pid": 1728719, "tid": 1728719, "ts": 6950381116456.130, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116456.481, "dur": 0.100 , "args": { "External id": 2438,"cbid": 468, "correlation": 8875 } }, { "ph": "f", "id": 8875, "pid": 1728719, "tid": 1728719, "ts": 6950381116456.481, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116456.812, "dur": 0.020 , "args": { "External id": 2438,"cbid": 1, "correlation": 8877 } }, { "ph": "f", "id": 8877, "pid": 1728719, "tid": 1728719, "ts": 6950381116456.812, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116456.922, "dur": 0.100 , "args": { "External id": 2438,"cbid": 468, "correlation": 8878 } }, { "ph": "f", "id": 8878, "pid": 1728719, "tid": 1728719, "ts": 6950381116456.922, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116457.182, "dur": 0.020 , "args": { "External id": 2438,"cbid": 1, "correlation": 8880 } }, { "ph": "f", "id": 8880, "pid": 1728719, "tid": 1728719, "ts": 6950381116457.182, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116457.293, "dur": 0.410 , "args": { "External id": 2438,"cbid": 273, "correlation": 8881 } }, { "ph": "f", "id": 8881, "pid": 1728719, "tid": 1728719, "ts": 6950381116457.293, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381176691.704, "dur": 228.478 , "args": { "External id": 2438,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8884, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8884, "pid": 0, "tid": 7, "ts": 6950381176691.704, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116458.094, "dur": 1.733 , "args": { "External id": 2438,"cbid": 211, "correlation": 8884 } }, { "ph": "s", "id": 8884, "pid": 1728719, "tid": 1728719, "ts": 6950381116458.094, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381176920.470, "dur": 8.832 , "args": { "External id": 2450,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8893, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8893, "pid": 0, "tid": 7, "ts": 6950381176920.470, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116465.868, "dur": 1.583 , "args": { "External id": 2450,"cbid": 307, "correlation": 8893 } }, { "ph": "s", "id": 8893, "pid": 1728719, "tid": 1728719, "ts": 6950381116465.868, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381176929.558, "dur": 26.015 , "args": { "External id": 2455,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8916, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8916, "pid": 0, "tid": 7, "ts": 6950381176929.558, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116482.109, "dur": 1.563 , "args": { "External id": 2455,"cbid": 307, "correlation": 8916 } }, { "ph": "s", "id": 8916, "pid": 1728719, "tid": 1728719, "ts": 6950381116482.109, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116492.097, "dur": 0.090 , "args": { "External id": 2456,"cbid": 200, "correlation": 8931 } }, { "ph": "f", "id": 8931, "pid": 1728719, "tid": 1728719, "ts": 6950381116492.097, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116492.278, "dur": 0.140 , "args": { "External id": 2456,"cbid": 468, "correlation": 8932 } }, { "ph": "f", "id": 8932, "pid": 1728719, "tid": 1728719, "ts": 6950381116492.278, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116492.618, "dur": 0.030 , "args": { "External id": 2456,"cbid": 1, "correlation": 8934 } }, { "ph": "f", "id": 8934, "pid": 1728719, "tid": 1728719, "ts": 6950381116492.618, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116492.728, "dur": 0.101 , "args": { "External id": 2456,"cbid": 468, "correlation": 8935 } }, { "ph": "f", "id": 8935, "pid": 1728719, "tid": 1728719, "ts": 6950381116492.728, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116492.969, "dur": 0.030 , "args": { "External id": 2456,"cbid": 1, "correlation": 8937 } }, { "ph": "f", "id": 8937, "pid": 1728719, "tid": 1728719, "ts": 6950381116492.969, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116493.099, "dur": 0.100 , "args": { "External id": 2456,"cbid": 468, "correlation": 8938 } }, { "ph": "f", "id": 8938, "pid": 1728719, "tid": 1728719, "ts": 6950381116493.099, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116493.370, "dur": 0.020 , "args": { "External id": 2456,"cbid": 1, "correlation": 8940 } }, { "ph": "f", "id": 8940, "pid": 1728719, "tid": 1728719, "ts": 6950381116493.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116493.480, "dur": 0.371 , "args": { "External id": 2456,"cbid": 273, "correlation": 8941 } }, { "ph": "f", "id": 8941, "pid": 1728719, "tid": 1728719, "ts": 6950381116493.480, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381176955.829, "dur": 230.334 , "args": { "External id": 2456,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8944, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8944, "pid": 0, "tid": 7, "ts": 6950381176955.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116494.221, "dur": 1.713 , "args": { "External id": 2456,"cbid": 211, "correlation": 8944 } }, { "ph": "s", "id": 8944, "pid": 1728719, "tid": 1728719, "ts": 6950381116494.221, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381177186.483, "dur": 59.679 , "args": { "External id": 2468,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8958, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8958, "pid": 0, "tid": 7, "ts": 6950381177186.483, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116503.549, "dur": 1.563 , "args": { "External id": 2468,"cbid": 307, "correlation": 8958 } }, { "ph": "s", "id": 8958, "pid": 1728719, "tid": 1728719, "ts": 6950381116503.549, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381177246.418, "dur": 64.671 , "args": { "External id": 2469,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8968, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8968, "pid": 0, "tid": 7, "ts": 6950381177246.418, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116513.487, "dur": 1.643 , "args": { "External id": 2469,"cbid": 307, "correlation": 8968 } }, { "ph": "s", "id": 8968, "pid": 1728719, "tid": 1728719, "ts": 6950381116513.487, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381177311.377, "dur": 38.496 , "args": { "External id": 2470,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 8977, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 8977, "pid": 0, "tid": 7, "ts": 6950381177311.377, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116522.845, "dur": 1.583 , "args": { "External id": 2470,"cbid": 307, "correlation": 8977 } }, { "ph": "s", "id": 8977, "pid": 1728719, "tid": 1728719, "ts": 6950381116522.845, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116544.345, "dur": 0.180 , "args": { "External id": 2478,"cbid": 200, "correlation": 9012 } }, { "ph": "f", "id": 9012, "pid": 1728719, "tid": 1728719, "ts": 6950381116544.345, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116544.605, "dur": 0.150 , "args": { "External id": 2478,"cbid": 200, "correlation": 9013 } }, { "ph": "f", "id": 9013, "pid": 1728719, "tid": 1728719, "ts": 6950381116544.605, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116544.856, "dur": 0.510 , "args": { "External id": 2478,"cbid": 273, "correlation": 9014 } }, { "ph": "f", "id": 9014, "pid": 1728719, "tid": 1728719, "ts": 6950381116544.856, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381177350.129, "dur": 849.303 , "args": { "External id": 2478,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9016, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9016, "pid": 0, "tid": 7, "ts": 6950381177350.129, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116545.727, "dur": 1.824 , "args": { "External id": 2478,"cbid": 211, "correlation": 9016 } }, { "ph": "s", "id": 9016, "pid": 1728719, "tid": 1728719, "ts": 6950381116545.727, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381178199.720, "dur": 9.440 , "args": { "External id": 2487,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9036, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9036, "pid": 0, "tid": 7, "ts": 6950381178199.720, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116562.398, "dur": 1.583 , "args": { "External id": 2487,"cbid": 652, "correlation": 9036 } }, { "ph": "s", "id": 9036, "pid": 1728719, "tid": 1728719, "ts": 6950381116562.398, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116573.499, "dur": 0.110 , "args": { "External id": 2488,"cbid": 200, "correlation": 9051 } }, { "ph": "f", "id": 9051, "pid": 1728719, "tid": 1728719, "ts": 6950381116573.499, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116573.729, "dur": 0.161 , "args": { "External id": 2488,"cbid": 468, "correlation": 9052 } }, { "ph": "f", "id": 9052, "pid": 1728719, "tid": 1728719, "ts": 6950381116573.729, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116574.180, "dur": 0.030 , "args": { "External id": 2488,"cbid": 1, "correlation": 9054 } }, { "ph": "f", "id": 9054, "pid": 1728719, "tid": 1728719, "ts": 6950381116574.180, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116574.280, "dur": 0.100 , "args": { "External id": 2488,"cbid": 468, "correlation": 9055 } }, { "ph": "f", "id": 9055, "pid": 1728719, "tid": 1728719, "ts": 6950381116574.280, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116574.541, "dur": 0.030 , "args": { "External id": 2488,"cbid": 1, "correlation": 9057 } }, { "ph": "f", "id": 9057, "pid": 1728719, "tid": 1728719, "ts": 6950381116574.541, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116574.681, "dur": 0.100 , "args": { "External id": 2488,"cbid": 468, "correlation": 9058 } }, { "ph": "f", "id": 9058, "pid": 1728719, "tid": 1728719, "ts": 6950381116574.681, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116574.942, "dur": 0.020 , "args": { "External id": 2488,"cbid": 1, "correlation": 9060 } }, { "ph": "f", "id": 9060, "pid": 1728719, "tid": 1728719, "ts": 6950381116574.942, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116575.062, "dur": 0.421 , "args": { "External id": 2488,"cbid": 273, "correlation": 9061 } }, { "ph": "f", "id": 9061, "pid": 1728719, "tid": 1728719, "ts": 6950381116575.062, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381178209.416, "dur": 208.093 , "args": { "External id": 2488,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9064, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9064, "pid": 0, "tid": 7, "ts": 6950381178209.416, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116575.853, "dur": 1.703 , "args": { "External id": 2488,"cbid": 211, "correlation": 9064 } }, { "ph": "s", "id": 9064, "pid": 1728719, "tid": 1728719, "ts": 6950381116575.853, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381178417.797, "dur": 46.592 , "args": { "External id": 2500,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9076, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9076, "pid": 0, "tid": 7, "ts": 6950381178417.797, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116585.702, "dur": 1.643 , "args": { "External id": 2500,"cbid": 307, "correlation": 9076 } }, { "ph": "s", "id": 9076, "pid": 1728719, "tid": 1728719, "ts": 6950381116585.702, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381178464.677, "dur": 12.896 , "args": { "External id": 2501,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9082, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9082, "pid": 0, "tid": 7, "ts": 6950381178464.677, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116593.466, "dur": 1.493 , "args": { "External id": 2501,"cbid": 652, "correlation": 9082 } }, { "ph": "s", "id": 9082, "pid": 1728719, "tid": 1728719, "ts": 6950381116593.466, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381178477.893, "dur": 12.832 , "args": { "External id": 2502,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9088, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9088, "pid": 0, "tid": 7, "ts": 6950381178477.893, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116601.641, "dur": 1.443 , "args": { "External id": 2502,"cbid": 652, "correlation": 9088 } }, { "ph": "s", "id": 9088, "pid": 1728719, "tid": 1728719, "ts": 6950381116601.641, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116613.954, "dur": 0.090 , "args": { "External id": 2503,"cbid": 200, "correlation": 9107 } }, { "ph": "f", "id": 9107, "pid": 1728719, "tid": 1728719, "ts": 6950381116613.954, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116614.145, "dur": 0.120 , "args": { "External id": 2503,"cbid": 468, "correlation": 9108 } }, { "ph": "f", "id": 9108, "pid": 1728719, "tid": 1728719, "ts": 6950381116614.145, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116614.465, "dur": 0.020 , "args": { "External id": 2503,"cbid": 1, "correlation": 9110 } }, { "ph": "f", "id": 9110, "pid": 1728719, "tid": 1728719, "ts": 6950381116614.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116614.625, "dur": 0.101 , "args": { "External id": 2503,"cbid": 468, "correlation": 9111 } }, { "ph": "f", "id": 9111, "pid": 1728719, "tid": 1728719, "ts": 6950381116614.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116614.906, "dur": 0.020 , "args": { "External id": 2503,"cbid": 1, "correlation": 9113 } }, { "ph": "f", "id": 9113, "pid": 1728719, "tid": 1728719, "ts": 6950381116614.906, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116615.016, "dur": 0.100 , "args": { "External id": 2503,"cbid": 468, "correlation": 9114 } }, { "ph": "f", "id": 9114, "pid": 1728719, "tid": 1728719, "ts": 6950381116615.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116615.267, "dur": 0.020 , "args": { "External id": 2503,"cbid": 1, "correlation": 9116 } }, { "ph": "f", "id": 9116, "pid": 1728719, "tid": 1728719, "ts": 6950381116615.267, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116615.377, "dur": 0.391 , "args": { "External id": 2503,"cbid": 273, "correlation": 9117 } }, { "ph": "f", "id": 9117, "pid": 1728719, "tid": 1728719, "ts": 6950381116615.377, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381178491.045, "dur": 664.056 , "args": { "External id": 2503,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9120, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9120, "pid": 0, "tid": 7, "ts": 6950381178491.045, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116616.128, "dur": 1.693 , "args": { "External id": 2503,"cbid": 211, "correlation": 9120 } }, { "ph": "s", "id": 9120, "pid": 1728719, "tid": 1728719, "ts": 6950381116616.128, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116627.870, "dur": 0.080 , "args": { "External id": 2515,"cbid": 200, "correlation": 9144 } }, { "ph": "f", "id": 9144, "pid": 1728719, "tid": 1728719, "ts": 6950381116627.870, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116628.040, "dur": 0.121 , "args": { "External id": 2515,"cbid": 468, "correlation": 9145 } }, { "ph": "f", "id": 9145, "pid": 1728719, "tid": 1728719, "ts": 6950381116628.040, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116628.341, "dur": 0.020 , "args": { "External id": 2515,"cbid": 1, "correlation": 9147 } }, { "ph": "f", "id": 9147, "pid": 1728719, "tid": 1728719, "ts": 6950381116628.341, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116628.431, "dur": 0.100 , "args": { "External id": 2515,"cbid": 468, "correlation": 9148 } }, { "ph": "f", "id": 9148, "pid": 1728719, "tid": 1728719, "ts": 6950381116628.431, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116628.672, "dur": 0.030 , "args": { "External id": 2515,"cbid": 1, "correlation": 9150 } }, { "ph": "f", "id": 9150, "pid": 1728719, "tid": 1728719, "ts": 6950381116628.672, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116628.802, "dur": 0.090 , "args": { "External id": 2515,"cbid": 468, "correlation": 9151 } }, { "ph": "f", "id": 9151, "pid": 1728719, "tid": 1728719, "ts": 6950381116628.802, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116629.052, "dur": 0.020 , "args": { "External id": 2515,"cbid": 1, "correlation": 9153 } }, { "ph": "f", "id": 9153, "pid": 1728719, "tid": 1728719, "ts": 6950381116629.052, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116629.172, "dur": 0.341 , "args": { "External id": 2515,"cbid": 273, "correlation": 9154 } }, { "ph": "f", "id": 9154, "pid": 1728719, "tid": 1728719, "ts": 6950381116629.172, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381179155.389, "dur": 667.129 , "args": { "External id": 2515,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9157, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9157, "pid": 0, "tid": 7, "ts": 6950381179155.389, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116629.854, "dur": 1.793 , "args": { "External id": 2515,"cbid": 211, "correlation": 9157 } }, { "ph": "s", "id": 9157, "pid": 1728719, "tid": 1728719, "ts": 6950381116629.854, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381179822.838, "dur": 156.351 , "args": { "External id": 2527,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9170, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9170, "pid": 0, "tid": 7, "ts": 6950381179822.838, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116639.131, "dur": 1.693 , "args": { "External id": 2527,"cbid": 307, "correlation": 9170 } }, { "ph": "s", "id": 9170, "pid": 1728719, "tid": 1728719, "ts": 6950381116639.131, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381179979.445, "dur": 58.271 , "args": { "External id": 2528,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9176, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9176, "pid": 0, "tid": 7, "ts": 6950381179979.445, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116647.567, "dur": 1.643 , "args": { "External id": 2528,"cbid": 652, "correlation": 9176 } }, { "ph": "s", "id": 9176, "pid": 1728719, "tid": 1728719, "ts": 6950381116647.567, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116659.228, "dur": 0.081 , "args": { "External id": 2529,"cbid": 200, "correlation": 9195 } }, { "ph": "f", "id": 9195, "pid": 1728719, "tid": 1728719, "ts": 6950381116659.228, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116659.409, "dur": 0.110 , "args": { "External id": 2529,"cbid": 468, "correlation": 9196 } }, { "ph": "f", "id": 9196, "pid": 1728719, "tid": 1728719, "ts": 6950381116659.409, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116659.719, "dur": 0.020 , "args": { "External id": 2529,"cbid": 1, "correlation": 9198 } }, { "ph": "f", "id": 9198, "pid": 1728719, "tid": 1728719, "ts": 6950381116659.719, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116659.820, "dur": 0.100 , "args": { "External id": 2529,"cbid": 468, "correlation": 9199 } }, { "ph": "f", "id": 9199, "pid": 1728719, "tid": 1728719, "ts": 6950381116659.820, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116660.070, "dur": 0.020 , "args": { "External id": 2529,"cbid": 1, "correlation": 9201 } }, { "ph": "f", "id": 9201, "pid": 1728719, "tid": 1728719, "ts": 6950381116660.070, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116660.190, "dur": 0.100 , "args": { "External id": 2529,"cbid": 468, "correlation": 9202 } }, { "ph": "f", "id": 9202, "pid": 1728719, "tid": 1728719, "ts": 6950381116660.190, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116660.451, "dur": 0.020 , "args": { "External id": 2529,"cbid": 1, "correlation": 9204 } }, { "ph": "f", "id": 9204, "pid": 1728719, "tid": 1728719, "ts": 6950381116660.451, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116660.551, "dur": 0.361 , "args": { "External id": 2529,"cbid": 273, "correlation": 9205 } }, { "ph": "f", "id": 9205, "pid": 1728719, "tid": 1728719, "ts": 6950381116660.551, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381180037.972, "dur": 677.145 , "args": { "External id": 2529,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9208, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9208, "pid": 0, "tid": 7, "ts": 6950381180037.972, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116661.282, "dur": 1.693 , "args": { "External id": 2529,"cbid": 211, "correlation": 9208 } }, { "ph": "s", "id": 9208, "pid": 1728719, "tid": 1728719, "ts": 6950381116661.282, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381180715.436, "dur": 40.192 , "args": { "External id": 2541,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9219, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9219, "pid": 0, "tid": 7, "ts": 6950381180715.436, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116670.990, "dur": 1.633 , "args": { "External id": 2541,"cbid": 307, "correlation": 9219 } }, { "ph": "s", "id": 9219, "pid": 1728719, "tid": 1728719, "ts": 6950381116670.990, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381180755.916, "dur": 20.640 , "args": { "External id": 2553,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9231, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9231, "pid": 0, "tid": 7, "ts": 6950381180755.916, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116750.609, "dur": 2.524 , "args": { "External id": 2553,"cbid": 307, "correlation": 9231 } }, { "ph": "s", "id": 9231, "pid": 1728719, "tid": 1728719, "ts": 6950381116750.609, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381180776.844, "dur": 12.736 , "args": { "External id": 2554,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9245, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9245, "pid": 0, "tid": 7, "ts": 6950381180776.844, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116764.725, "dur": 1.663 , "args": { "External id": 2554,"cbid": 652, "correlation": 9245 } }, { "ph": "s", "id": 9245, "pid": 1728719, "tid": 1728719, "ts": 6950381116764.725, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381180789.900, "dur": 12.640 , "args": { "External id": 2555,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9259, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9259, "pid": 0, "tid": 7, "ts": 6950381180789.900, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116776.286, "dur": 1.573 , "args": { "External id": 2555,"cbid": 652, "correlation": 9259 } }, { "ph": "s", "id": 9259, "pid": 1728719, "tid": 1728719, "ts": 6950381116776.286, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381180802.828, "dur": 10.880 , "args": { "External id": 2556,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9273, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9273, "pid": 0, "tid": 7, "ts": 6950381180802.828, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116786.846, "dur": 1.623 , "args": { "External id": 2556,"cbid": 652, "correlation": 9273 } }, { "ph": "s", "id": 9273, "pid": 1728719, "tid": 1728719, "ts": 6950381116786.846, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116801.483, "dur": 0.161 , "args": { "External id": 2557,"cbid": 200, "correlation": 9288 } }, { "ph": "f", "id": 9288, "pid": 1728719, "tid": 1728719, "ts": 6950381116801.483, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116801.804, "dur": 0.190 , "args": { "External id": 2557,"cbid": 468, "correlation": 9289 } }, { "ph": "f", "id": 9289, "pid": 1728719, "tid": 1728719, "ts": 6950381116801.804, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116802.355, "dur": 0.030 , "args": { "External id": 2557,"cbid": 1, "correlation": 9291 } }, { "ph": "f", "id": 9291, "pid": 1728719, "tid": 1728719, "ts": 6950381116802.355, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116802.635, "dur": 0.101 , "args": { "External id": 2557,"cbid": 468, "correlation": 9292 } }, { "ph": "f", "id": 9292, "pid": 1728719, "tid": 1728719, "ts": 6950381116802.635, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116802.886, "dur": 0.030 , "args": { "External id": 2557,"cbid": 1, "correlation": 9294 } }, { "ph": "f", "id": 9294, "pid": 1728719, "tid": 1728719, "ts": 6950381116802.886, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116803.006, "dur": 0.100 , "args": { "External id": 2557,"cbid": 468, "correlation": 9295 } }, { "ph": "f", "id": 9295, "pid": 1728719, "tid": 1728719, "ts": 6950381116803.006, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116803.257, "dur": 0.020 , "args": { "External id": 2557,"cbid": 1, "correlation": 9297 } }, { "ph": "f", "id": 9297, "pid": 1728719, "tid": 1728719, "ts": 6950381116803.257, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116803.397, "dur": 0.621 , "args": { "External id": 2557,"cbid": 273, "correlation": 9298 } }, { "ph": "f", "id": 9298, "pid": 1728719, "tid": 1728719, "ts": 6950381116803.397, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381180813.996, "dur": 214.717 , "args": { "External id": 2557,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9301, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9301, "pid": 0, "tid": 7, "ts": 6950381180813.996, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116804.509, "dur": 1.873 , "args": { "External id": 2557,"cbid": 211, "correlation": 9301 } }, { "ph": "s", "id": 9301, "pid": 1728719, "tid": 1728719, "ts": 6950381116804.509, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381181029.001, "dur": 8.864 , "args": { "External id": 2569,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9314, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9314, "pid": 0, "tid": 7, "ts": 6950381181029.001, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116815.369, "dur": 1.663 , "args": { "External id": 2569,"cbid": 307, "correlation": 9314 } }, { "ph": "s", "id": 9314, "pid": 1728719, "tid": 1728719, "ts": 6950381116815.369, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381181038.153, "dur": 37.024 , "args": { "External id": 2574,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9337, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9337, "pid": 0, "tid": 7, "ts": 6950381181038.153, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116834.765, "dur": 1.663 , "args": { "External id": 2574,"cbid": 307, "correlation": 9337 } }, { "ph": "s", "id": 9337, "pid": 1728719, "tid": 1728719, "ts": 6950381116834.765, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116844.924, "dur": 0.140 , "args": { "External id": 2575,"cbid": 200, "correlation": 9352 } }, { "ph": "f", "id": 9352, "pid": 1728719, "tid": 1728719, "ts": 6950381116844.924, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116845.144, "dur": 0.141 , "args": { "External id": 2575,"cbid": 468, "correlation": 9353 } }, { "ph": "f", "id": 9353, "pid": 1728719, "tid": 1728719, "ts": 6950381116845.144, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116845.525, "dur": 0.030 , "args": { "External id": 2575,"cbid": 1, "correlation": 9355 } }, { "ph": "f", "id": 9355, "pid": 1728719, "tid": 1728719, "ts": 6950381116845.525, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116845.625, "dur": 0.101 , "args": { "External id": 2575,"cbid": 468, "correlation": 9356 } }, { "ph": "f", "id": 9356, "pid": 1728719, "tid": 1728719, "ts": 6950381116845.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116845.866, "dur": 0.020 , "args": { "External id": 2575,"cbid": 1, "correlation": 9358 } }, { "ph": "f", "id": 9358, "pid": 1728719, "tid": 1728719, "ts": 6950381116845.866, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116845.986, "dur": 0.090 , "args": { "External id": 2575,"cbid": 468, "correlation": 9359 } }, { "ph": "f", "id": 9359, "pid": 1728719, "tid": 1728719, "ts": 6950381116845.986, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116846.237, "dur": 0.020 , "args": { "External id": 2575,"cbid": 1, "correlation": 9361 } }, { "ph": "f", "id": 9361, "pid": 1728719, "tid": 1728719, "ts": 6950381116846.237, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116846.357, "dur": 0.461 , "args": { "External id": 2575,"cbid": 273, "correlation": 9362 } }, { "ph": "f", "id": 9362, "pid": 1728719, "tid": 1728719, "ts": 6950381116846.357, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381181075.465, "dur": 228.829 , "args": { "External id": 2575,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9365, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9365, "pid": 0, "tid": 7, "ts": 6950381181075.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116847.208, "dur": 1.804 , "args": { "External id": 2575,"cbid": 211, "correlation": 9365 } }, { "ph": "s", "id": 9365, "pid": 1728719, "tid": 1728719, "ts": 6950381116847.208, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381181304.614, "dur": 8.800 , "args": { "External id": 2587,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9374, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9374, "pid": 0, "tid": 7, "ts": 6950381181304.614, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116855.083, "dur": 1.603 , "args": { "External id": 2587,"cbid": 307, "correlation": 9374 } }, { "ph": "s", "id": 9374, "pid": 1728719, "tid": 1728719, "ts": 6950381116855.083, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381181313.702, "dur": 27.488 , "args": { "External id": 2592,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9397, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9397, "pid": 0, "tid": 7, "ts": 6950381181313.702, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116870.882, "dur": 1.643 , "args": { "External id": 2592,"cbid": 307, "correlation": 9397 } }, { "ph": "s", "id": 9397, "pid": 1728719, "tid": 1728719, "ts": 6950381116870.882, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116880.941, "dur": 0.090 , "args": { "External id": 2593,"cbid": 200, "correlation": 9412 } }, { "ph": "f", "id": 9412, "pid": 1728719, "tid": 1728719, "ts": 6950381116880.941, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116881.111, "dur": 0.131 , "args": { "External id": 2593,"cbid": 468, "correlation": 9413 } }, { "ph": "f", "id": 9413, "pid": 1728719, "tid": 1728719, "ts": 6950381116881.111, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116881.452, "dur": 0.030 , "args": { "External id": 2593,"cbid": 1, "correlation": 9415 } }, { "ph": "f", "id": 9415, "pid": 1728719, "tid": 1728719, "ts": 6950381116881.452, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116881.542, "dur": 0.100 , "args": { "External id": 2593,"cbid": 468, "correlation": 9416 } }, { "ph": "f", "id": 9416, "pid": 1728719, "tid": 1728719, "ts": 6950381116881.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116881.793, "dur": 0.020 , "args": { "External id": 2593,"cbid": 1, "correlation": 9418 } }, { "ph": "f", "id": 9418, "pid": 1728719, "tid": 1728719, "ts": 6950381116881.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116881.923, "dur": 0.100 , "args": { "External id": 2593,"cbid": 468, "correlation": 9419 } }, { "ph": "f", "id": 9419, "pid": 1728719, "tid": 1728719, "ts": 6950381116881.923, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116882.173, "dur": 0.020 , "args": { "External id": 2593,"cbid": 1, "correlation": 9421 } }, { "ph": "f", "id": 9421, "pid": 1728719, "tid": 1728719, "ts": 6950381116882.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116882.284, "dur": 0.370 , "args": { "External id": 2593,"cbid": 273, "correlation": 9422 } }, { "ph": "f", "id": 9422, "pid": 1728719, "tid": 1728719, "ts": 6950381116882.284, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381181341.478, "dur": 231.037 , "args": { "External id": 2593,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9425, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9425, "pid": 0, "tid": 7, "ts": 6950381181341.478, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116883.005, "dur": 1.693 , "args": { "External id": 2593,"cbid": 211, "correlation": 9425 } }, { "ph": "s", "id": 9425, "pid": 1728719, "tid": 1728719, "ts": 6950381116883.005, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381181572.803, "dur": 58.528 , "args": { "External id": 2605,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9439, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9439, "pid": 0, "tid": 7, "ts": 6950381181572.803, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116892.292, "dur": 1.513 , "args": { "External id": 2605,"cbid": 307, "correlation": 9439 } }, { "ph": "s", "id": 9439, "pid": 1728719, "tid": 1728719, "ts": 6950381116892.292, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381181631.619, "dur": 67.903 , "args": { "External id": 2606,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9449, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9449, "pid": 0, "tid": 7, "ts": 6950381181631.619, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116902.060, "dur": 1.794 , "args": { "External id": 2606,"cbid": 307, "correlation": 9449 } }, { "ph": "s", "id": 9449, "pid": 1728719, "tid": 1728719, "ts": 6950381116902.060, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381181699.810, "dur": 37.472 , "args": { "External id": 2607,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9458, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9458, "pid": 0, "tid": 7, "ts": 6950381181699.810, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116911.338, "dur": 1.713 , "args": { "External id": 2607,"cbid": 307, "correlation": 9458 } }, { "ph": "s", "id": 9458, "pid": 1728719, "tid": 1728719, "ts": 6950381116911.338, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116926.516, "dur": 0.100 , "args": { "External id": 2615,"cbid": 200, "correlation": 9493 } }, { "ph": "f", "id": 9493, "pid": 1728719, "tid": 1728719, "ts": 6950381116926.516, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116926.706, "dur": 0.070 , "args": { "External id": 2615,"cbid": 200, "correlation": 9494 } }, { "ph": "f", "id": 9494, "pid": 1728719, "tid": 1728719, "ts": 6950381116926.706, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116926.857, "dur": 0.470 , "args": { "External id": 2615,"cbid": 273, "correlation": 9495 } }, { "ph": "f", "id": 9495, "pid": 1728719, "tid": 1728719, "ts": 6950381116926.857, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381181737.537, "dur": 849.175 , "args": { "External id": 2615,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9497, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9497, "pid": 0, "tid": 7, "ts": 6950381181737.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116927.648, "dur": 2.094 , "args": { "External id": 2615,"cbid": 211, "correlation": 9497 } }, { "ph": "s", "id": 9497, "pid": 1728719, "tid": 1728719, "ts": 6950381116927.648, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381182587.000, "dur": 9.504 , "args": { "External id": 2624,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9517, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9517, "pid": 0, "tid": 7, "ts": 6950381182587.000, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116944.209, "dur": 1.623 , "args": { "External id": 2624,"cbid": 652, "correlation": 9517 } }, { "ph": "s", "id": 9517, "pid": 1728719, "tid": 1728719, "ts": 6950381116944.209, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116955.139, "dur": 0.100 , "args": { "External id": 2625,"cbid": 200, "correlation": 9532 } }, { "ph": "f", "id": 9532, "pid": 1728719, "tid": 1728719, "ts": 6950381116955.139, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116955.330, "dur": 0.160 , "args": { "External id": 2625,"cbid": 468, "correlation": 9533 } }, { "ph": "f", "id": 9533, "pid": 1728719, "tid": 1728719, "ts": 6950381116955.330, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116955.740, "dur": 0.020 , "args": { "External id": 2625,"cbid": 1, "correlation": 9535 } }, { "ph": "f", "id": 9535, "pid": 1728719, "tid": 1728719, "ts": 6950381116955.740, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116955.841, "dur": 0.100 , "args": { "External id": 2625,"cbid": 468, "correlation": 9536 } }, { "ph": "f", "id": 9536, "pid": 1728719, "tid": 1728719, "ts": 6950381116955.841, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116956.101, "dur": 0.020 , "args": { "External id": 2625,"cbid": 1, "correlation": 9538 } }, { "ph": "f", "id": 9538, "pid": 1728719, "tid": 1728719, "ts": 6950381116956.101, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116956.221, "dur": 0.100 , "args": { "External id": 2625,"cbid": 468, "correlation": 9539 } }, { "ph": "f", "id": 9539, "pid": 1728719, "tid": 1728719, "ts": 6950381116956.221, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116956.472, "dur": 0.020 , "args": { "External id": 2625,"cbid": 1, "correlation": 9541 } }, { "ph": "f", "id": 9541, "pid": 1728719, "tid": 1728719, "ts": 6950381116956.472, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116956.572, "dur": 0.391 , "args": { "External id": 2625,"cbid": 273, "correlation": 9542 } }, { "ph": "f", "id": 9542, "pid": 1728719, "tid": 1728719, "ts": 6950381116956.572, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381182596.760, "dur": 208.126 , "args": { "External id": 2625,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9545, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9545, "pid": 0, "tid": 7, "ts": 6950381182596.760, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116957.343, "dur": 1.814 , "args": { "External id": 2625,"cbid": 211, "correlation": 9545 } }, { "ph": "s", "id": 9545, "pid": 1728719, "tid": 1728719, "ts": 6950381116957.343, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381182805.174, "dur": 46.239 , "args": { "External id": 2637,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9557, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9557, "pid": 0, "tid": 7, "ts": 6950381182805.174, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116966.981, "dur": 1.703 , "args": { "External id": 2637,"cbid": 307, "correlation": 9557 } }, { "ph": "s", "id": 9557, "pid": 1728719, "tid": 1728719, "ts": 6950381116966.981, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381182851.734, "dur": 12.895 , "args": { "External id": 2638,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9563, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9563, "pid": 0, "tid": 7, "ts": 6950381182851.734, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116975.367, "dur": 1.443 , "args": { "External id": 2638,"cbid": 652, "correlation": 9563 } }, { "ph": "s", "id": 9563, "pid": 1728719, "tid": 1728719, "ts": 6950381116975.367, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381182864.949, "dur": 12.672 , "args": { "External id": 2639,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9569, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9569, "pid": 0, "tid": 7, "ts": 6950381182864.949, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381116983.712, "dur": 1.463 , "args": { "External id": 2639,"cbid": 652, "correlation": 9569 } }, { "ph": "s", "id": 9569, "pid": 1728719, "tid": 1728719, "ts": 6950381116983.712, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116995.334, "dur": 0.090 , "args": { "External id": 2640,"cbid": 200, "correlation": 9588 } }, { "ph": "f", "id": 9588, "pid": 1728719, "tid": 1728719, "ts": 6950381116995.334, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116995.544, "dur": 0.141 , "args": { "External id": 2640,"cbid": 468, "correlation": 9589 } }, { "ph": "f", "id": 9589, "pid": 1728719, "tid": 1728719, "ts": 6950381116995.544, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116995.885, "dur": 0.020 , "args": { "External id": 2640,"cbid": 1, "correlation": 9591 } }, { "ph": "f", "id": 9591, "pid": 1728719, "tid": 1728719, "ts": 6950381116995.885, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116995.995, "dur": 0.090 , "args": { "External id": 2640,"cbid": 468, "correlation": 9592 } }, { "ph": "f", "id": 9592, "pid": 1728719, "tid": 1728719, "ts": 6950381116995.995, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116996.236, "dur": 0.020 , "args": { "External id": 2640,"cbid": 1, "correlation": 9594 } }, { "ph": "f", "id": 9594, "pid": 1728719, "tid": 1728719, "ts": 6950381116996.236, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116996.346, "dur": 0.090 , "args": { "External id": 2640,"cbid": 468, "correlation": 9595 } }, { "ph": "f", "id": 9595, "pid": 1728719, "tid": 1728719, "ts": 6950381116996.346, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381116996.586, "dur": 0.030 , "args": { "External id": 2640,"cbid": 1, "correlation": 9597 } }, { "ph": "f", "id": 9597, "pid": 1728719, "tid": 1728719, "ts": 6950381116996.586, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381116996.707, "dur": 0.360 , "args": { "External id": 2640,"cbid": 273, "correlation": 9598 } }, { "ph": "f", "id": 9598, "pid": 1728719, "tid": 1728719, "ts": 6950381116996.707, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381182877.941, "dur": 657.465 , "args": { "External id": 2640,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9601, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9601, "pid": 0, "tid": 7, "ts": 6950381182877.941, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381116997.438, "dur": 1.703 , "args": { "External id": 2640,"cbid": 211, "correlation": 9601 } }, { "ph": "s", "id": 9601, "pid": 1728719, "tid": 1728719, "ts": 6950381116997.438, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117009.250, "dur": 0.090 , "args": { "External id": 2652,"cbid": 200, "correlation": 9625 } }, { "ph": "f", "id": 9625, "pid": 1728719, "tid": 1728719, "ts": 6950381117009.250, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117009.440, "dur": 0.121 , "args": { "External id": 2652,"cbid": 468, "correlation": 9626 } }, { "ph": "f", "id": 9626, "pid": 1728719, "tid": 1728719, "ts": 6950381117009.440, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117009.741, "dur": 0.030 , "args": { "External id": 2652,"cbid": 1, "correlation": 9628 } }, { "ph": "f", "id": 9628, "pid": 1728719, "tid": 1728719, "ts": 6950381117009.741, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117009.841, "dur": 0.100 , "args": { "External id": 2652,"cbid": 468, "correlation": 9629 } }, { "ph": "f", "id": 9629, "pid": 1728719, "tid": 1728719, "ts": 6950381117009.841, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117010.082, "dur": 0.020 , "args": { "External id": 2652,"cbid": 1, "correlation": 9631 } }, { "ph": "f", "id": 9631, "pid": 1728719, "tid": 1728719, "ts": 6950381117010.082, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117010.192, "dur": 0.100 , "args": { "External id": 2652,"cbid": 468, "correlation": 9632 } }, { "ph": "f", "id": 9632, "pid": 1728719, "tid": 1728719, "ts": 6950381117010.192, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117010.442, "dur": 0.020 , "args": { "External id": 2652,"cbid": 1, "correlation": 9634 } }, { "ph": "f", "id": 9634, "pid": 1728719, "tid": 1728719, "ts": 6950381117010.442, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117010.542, "dur": 0.331 , "args": { "External id": 2652,"cbid": 273, "correlation": 9635 } }, { "ph": "f", "id": 9635, "pid": 1728719, "tid": 1728719, "ts": 6950381117010.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381183535.726, "dur": 674.073 , "args": { "External id": 2652,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9638, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9638, "pid": 0, "tid": 7, "ts": 6950381183535.726, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117011.224, "dur": 1.763 , "args": { "External id": 2652,"cbid": 211, "correlation": 9638 } }, { "ph": "s", "id": 9638, "pid": 1728719, "tid": 1728719, "ts": 6950381117011.224, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381184210.119, "dur": 155.518 , "args": { "External id": 2664,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9651, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9651, "pid": 0, "tid": 7, "ts": 6950381184210.119, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117020.391, "dur": 1.753 , "args": { "External id": 2664,"cbid": 307, "correlation": 9651 } }, { "ph": "s", "id": 9651, "pid": 1728719, "tid": 1728719, "ts": 6950381117020.391, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381184365.893, "dur": 63.327 , "args": { "External id": 2665,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9657, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9657, "pid": 0, "tid": 7, "ts": 6950381184365.893, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117028.957, "dur": 1.573 , "args": { "External id": 2665,"cbid": 652, "correlation": 9657 } }, { "ph": "s", "id": 9657, "pid": 1728719, "tid": 1728719, "ts": 6950381117028.957, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117040.358, "dur": 0.110 , "args": { "External id": 2666,"cbid": 200, "correlation": 9676 } }, { "ph": "f", "id": 9676, "pid": 1728719, "tid": 1728719, "ts": 6950381117040.358, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117040.578, "dur": 0.121 , "args": { "External id": 2666,"cbid": 468, "correlation": 9677 } }, { "ph": "f", "id": 9677, "pid": 1728719, "tid": 1728719, "ts": 6950381117040.578, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117040.899, "dur": 0.030 , "args": { "External id": 2666,"cbid": 1, "correlation": 9679 } }, { "ph": "f", "id": 9679, "pid": 1728719, "tid": 1728719, "ts": 6950381117040.899, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117041.009, "dur": 0.100 , "args": { "External id": 2666,"cbid": 468, "correlation": 9680 } }, { "ph": "f", "id": 9680, "pid": 1728719, "tid": 1728719, "ts": 6950381117041.009, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117041.250, "dur": 0.020 , "args": { "External id": 2666,"cbid": 1, "correlation": 9682 } }, { "ph": "f", "id": 9682, "pid": 1728719, "tid": 1728719, "ts": 6950381117041.250, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117041.370, "dur": 0.090 , "args": { "External id": 2666,"cbid": 468, "correlation": 9683 } }, { "ph": "f", "id": 9683, "pid": 1728719, "tid": 1728719, "ts": 6950381117041.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117041.610, "dur": 0.020 , "args": { "External id": 2666,"cbid": 1, "correlation": 9685 } }, { "ph": "f", "id": 9685, "pid": 1728719, "tid": 1728719, "ts": 6950381117041.610, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117041.730, "dur": 0.361 , "args": { "External id": 2666,"cbid": 273, "correlation": 9686 } }, { "ph": "f", "id": 9686, "pid": 1728719, "tid": 1728719, "ts": 6950381117041.730, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381184429.508, "dur": 679.545 , "args": { "External id": 2666,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9689, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9689, "pid": 0, "tid": 7, "ts": 6950381184429.508, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117042.482, "dur": 1.733 , "args": { "External id": 2666,"cbid": 211, "correlation": 9689 } }, { "ph": "s", "id": 9689, "pid": 1728719, "tid": 1728719, "ts": 6950381117042.482, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381185109.341, "dur": 40.095 , "args": { "External id": 2678,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9700, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9700, "pid": 0, "tid": 7, "ts": 6950381185109.341, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117052.280, "dur": 1.623 , "args": { "External id": 2678,"cbid": 307, "correlation": 9700 } }, { "ph": "s", "id": 9700, "pid": 1728719, "tid": 1728719, "ts": 6950381117052.280, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381185149.692, "dur": 21.024 , "args": { "External id": 2690,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9712, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9712, "pid": 0, "tid": 7, "ts": 6950381185149.692, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117131.347, "dur": 2.445 , "args": { "External id": 2690,"cbid": 307, "correlation": 9712 } }, { "ph": "s", "id": 9712, "pid": 1728719, "tid": 1728719, "ts": 6950381117131.347, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381185171.036, "dur": 12.384 , "args": { "External id": 2691,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9726, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9726, "pid": 0, "tid": 7, "ts": 6950381185171.036, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117145.874, "dur": 1.643 , "args": { "External id": 2691,"cbid": 652, "correlation": 9726 } }, { "ph": "s", "id": 9726, "pid": 1728719, "tid": 1728719, "ts": 6950381117145.874, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381185183.740, "dur": 12.288 , "args": { "External id": 2692,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9740, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9740, "pid": 0, "tid": 7, "ts": 6950381185183.740, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117157.496, "dur": 1.513 , "args": { "External id": 2692,"cbid": 652, "correlation": 9740 } }, { "ph": "s", "id": 9740, "pid": 1728719, "tid": 1728719, "ts": 6950381117157.496, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381185196.316, "dur": 11.008 , "args": { "External id": 2693,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9754, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9754, "pid": 0, "tid": 7, "ts": 6950381185196.316, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117168.076, "dur": 1.532 , "args": { "External id": 2693,"cbid": 652, "correlation": 9754 } }, { "ph": "s", "id": 9754, "pid": 1728719, "tid": 1728719, "ts": 6950381117168.076, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117179.276, "dur": 0.171 , "args": { "External id": 2694,"cbid": 200, "correlation": 9769 } }, { "ph": "f", "id": 9769, "pid": 1728719, "tid": 1728719, "ts": 6950381117179.276, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117179.607, "dur": 0.160 , "args": { "External id": 2694,"cbid": 468, "correlation": 9770 } }, { "ph": "f", "id": 9770, "pid": 1728719, "tid": 1728719, "ts": 6950381117179.607, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117180.108, "dur": 0.030 , "args": { "External id": 2694,"cbid": 1, "correlation": 9772 } }, { "ph": "f", "id": 9772, "pid": 1728719, "tid": 1728719, "ts": 6950381117180.108, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117180.308, "dur": 0.101 , "args": { "External id": 2694,"cbid": 468, "correlation": 9773 } }, { "ph": "f", "id": 9773, "pid": 1728719, "tid": 1728719, "ts": 6950381117180.308, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117180.569, "dur": 0.020 , "args": { "External id": 2694,"cbid": 1, "correlation": 9775 } }, { "ph": "f", "id": 9775, "pid": 1728719, "tid": 1728719, "ts": 6950381117180.569, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117180.689, "dur": 0.100 , "args": { "External id": 2694,"cbid": 468, "correlation": 9776 } }, { "ph": "f", "id": 9776, "pid": 1728719, "tid": 1728719, "ts": 6950381117180.689, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117180.960, "dur": 0.020 , "args": { "External id": 2694,"cbid": 1, "correlation": 9778 } }, { "ph": "f", "id": 9778, "pid": 1728719, "tid": 1728719, "ts": 6950381117180.960, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117181.130, "dur": 0.681 , "args": { "External id": 2694,"cbid": 273, "correlation": 9779 } }, { "ph": "f", "id": 9779, "pid": 1728719, "tid": 1728719, "ts": 6950381117181.130, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381185207.580, "dur": 213.598 , "args": { "External id": 2694,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9782, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9782, "pid": 0, "tid": 7, "ts": 6950381185207.580, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117182.272, "dur": 1.874 , "args": { "External id": 2694,"cbid": 211, "correlation": 9782 } }, { "ph": "s", "id": 9782, "pid": 1728719, "tid": 1728719, "ts": 6950381117182.272, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381185421.466, "dur": 8.863 , "args": { "External id": 2706,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9795, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9795, "pid": 0, "tid": 7, "ts": 6950381185421.466, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117192.711, "dur": 1.613 , "args": { "External id": 2706,"cbid": 307, "correlation": 9795 } }, { "ph": "s", "id": 9795, "pid": 1728719, "tid": 1728719, "ts": 6950381117192.711, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381185430.586, "dur": 36.031 , "args": { "External id": 2711,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9818, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9818, "pid": 0, "tid": 7, "ts": 6950381185430.586, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117211.757, "dur": 1.643 , "args": { "External id": 2711,"cbid": 307, "correlation": 9818 } }, { "ph": "s", "id": 9818, "pid": 1728719, "tid": 1728719, "ts": 6950381117211.757, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117221.816, "dur": 0.100 , "args": { "External id": 2712,"cbid": 200, "correlation": 9833 } }, { "ph": "f", "id": 9833, "pid": 1728719, "tid": 1728719, "ts": 6950381117221.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117222.006, "dur": 0.140 , "args": { "External id": 2712,"cbid": 468, "correlation": 9834 } }, { "ph": "f", "id": 9834, "pid": 1728719, "tid": 1728719, "ts": 6950381117222.006, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117222.417, "dur": 0.050 , "args": { "External id": 2712,"cbid": 1, "correlation": 9836 } }, { "ph": "f", "id": 9836, "pid": 1728719, "tid": 1728719, "ts": 6950381117222.417, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117222.537, "dur": 0.100 , "args": { "External id": 2712,"cbid": 468, "correlation": 9837 } }, { "ph": "f", "id": 9837, "pid": 1728719, "tid": 1728719, "ts": 6950381117222.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117222.787, "dur": 0.020 , "args": { "External id": 2712,"cbid": 1, "correlation": 9839 } }, { "ph": "f", "id": 9839, "pid": 1728719, "tid": 1728719, "ts": 6950381117222.787, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117222.898, "dur": 0.100 , "args": { "External id": 2712,"cbid": 468, "correlation": 9840 } }, { "ph": "f", "id": 9840, "pid": 1728719, "tid": 1728719, "ts": 6950381117222.898, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117223.158, "dur": 0.020 , "args": { "External id": 2712,"cbid": 1, "correlation": 9842 } }, { "ph": "f", "id": 9842, "pid": 1728719, "tid": 1728719, "ts": 6950381117223.158, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117223.268, "dur": 0.391 , "args": { "External id": 2712,"cbid": 273, "correlation": 9843 } }, { "ph": "f", "id": 9843, "pid": 1728719, "tid": 1728719, "ts": 6950381117223.268, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381185466.905, "dur": 227.070 , "args": { "External id": 2712,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9846, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9846, "pid": 0, "tid": 7, "ts": 6950381185466.905, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117224.040, "dur": 1.783 , "args": { "External id": 2712,"cbid": 211, "correlation": 9846 } }, { "ph": "s", "id": 9846, "pid": 1728719, "tid": 1728719, "ts": 6950381117224.040, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381185694.263, "dur": 8.768 , "args": { "External id": 2724,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9855, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9855, "pid": 0, "tid": 7, "ts": 6950381185694.263, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117231.834, "dur": 1.523 , "args": { "External id": 2724,"cbid": 307, "correlation": 9855 } }, { "ph": "s", "id": 9855, "pid": 1728719, "tid": 1728719, "ts": 6950381117231.834, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381185703.286, "dur": 25.664 , "args": { "External id": 2729,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9878, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9878, "pid": 0, "tid": 7, "ts": 6950381185703.286, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117247.754, "dur": 1.603 , "args": { "External id": 2729,"cbid": 307, "correlation": 9878 } }, { "ph": "s", "id": 9878, "pid": 1728719, "tid": 1728719, "ts": 6950381117247.754, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117257.793, "dur": 0.090 , "args": { "External id": 2730,"cbid": 200, "correlation": 9893 } }, { "ph": "f", "id": 9893, "pid": 1728719, "tid": 1728719, "ts": 6950381117257.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117257.963, "dur": 0.140 , "args": { "External id": 2730,"cbid": 468, "correlation": 9894 } }, { "ph": "f", "id": 9894, "pid": 1728719, "tid": 1728719, "ts": 6950381117257.963, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117258.304, "dur": 0.030 , "args": { "External id": 2730,"cbid": 1, "correlation": 9896 } }, { "ph": "f", "id": 9896, "pid": 1728719, "tid": 1728719, "ts": 6950381117258.304, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117258.404, "dur": 0.100 , "args": { "External id": 2730,"cbid": 468, "correlation": 9897 } }, { "ph": "f", "id": 9897, "pid": 1728719, "tid": 1728719, "ts": 6950381117258.404, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117258.654, "dur": 0.020 , "args": { "External id": 2730,"cbid": 1, "correlation": 9899 } }, { "ph": "f", "id": 9899, "pid": 1728719, "tid": 1728719, "ts": 6950381117258.654, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117258.764, "dur": 0.101 , "args": { "External id": 2730,"cbid": 468, "correlation": 9900 } }, { "ph": "f", "id": 9900, "pid": 1728719, "tid": 1728719, "ts": 6950381117258.764, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117259.015, "dur": 0.020 , "args": { "External id": 2730,"cbid": 1, "correlation": 9902 } }, { "ph": "f", "id": 9902, "pid": 1728719, "tid": 1728719, "ts": 6950381117259.015, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117259.125, "dur": 0.361 , "args": { "External id": 2730,"cbid": 273, "correlation": 9903 } }, { "ph": "f", "id": 9903, "pid": 1728719, "tid": 1728719, "ts": 6950381117259.125, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381185729.238, "dur": 231.550 , "args": { "External id": 2730,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9906, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9906, "pid": 0, "tid": 7, "ts": 6950381185729.238, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117259.836, "dur": 1.744 , "args": { "External id": 2730,"cbid": 211, "correlation": 9906 } }, { "ph": "s", "id": 9906, "pid": 1728719, "tid": 1728719, "ts": 6950381117259.836, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381185961.076, "dur": 60.351 , "args": { "External id": 2742,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9920, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9920, "pid": 0, "tid": 7, "ts": 6950381185961.076, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117269.204, "dur": 1.543 , "args": { "External id": 2742,"cbid": 307, "correlation": 9920 } }, { "ph": "s", "id": 9920, "pid": 1728719, "tid": 1728719, "ts": 6950381117269.204, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381186021.715, "dur": 64.863 , "args": { "External id": 2743,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9930, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9930, "pid": 0, "tid": 7, "ts": 6950381186021.715, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117279.253, "dur": 1.863 , "args": { "External id": 2743,"cbid": 307, "correlation": 9930 } }, { "ph": "s", "id": 9930, "pid": 1728719, "tid": 1728719, "ts": 6950381117279.253, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381186086.866, "dur": 39.936 , "args": { "External id": 2744,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9939, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9939, "pid": 0, "tid": 7, "ts": 6950381186086.866, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117289.061, "dur": 1.603 , "args": { "External id": 2744,"cbid": 307, "correlation": 9939 } }, { "ph": "s", "id": 9939, "pid": 1728719, "tid": 1728719, "ts": 6950381117289.061, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117303.878, "dur": 0.111 , "args": { "External id": 2752,"cbid": 200, "correlation": 9974 } }, { "ph": "f", "id": 9974, "pid": 1728719, "tid": 1728719, "ts": 6950381117303.878, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117304.069, "dur": 0.070 , "args": { "External id": 2752,"cbid": 200, "correlation": 9975 } }, { "ph": "f", "id": 9975, "pid": 1728719, "tid": 1728719, "ts": 6950381117304.069, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117304.219, "dur": 0.461 , "args": { "External id": 2752,"cbid": 273, "correlation": 9976 } }, { "ph": "f", "id": 9976, "pid": 1728719, "tid": 1728719, "ts": 6950381117304.219, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381186127.090, "dur": 846.871 , "args": { "External id": 2752,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9978, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9978, "pid": 0, "tid": 7, "ts": 6950381186127.090, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117305.011, "dur": 1.853 , "args": { "External id": 2752,"cbid": 211, "correlation": 9978 } }, { "ph": "s", "id": 9978, "pid": 1728719, "tid": 1728719, "ts": 6950381117305.011, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381186974.217, "dur": 9.152 , "args": { "External id": 2761,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 9998, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 9998, "pid": 0, "tid": 7, "ts": 6950381186974.217, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117321.341, "dur": 1.643 , "args": { "External id": 2761,"cbid": 652, "correlation": 9998 } }, { "ph": "s", "id": 9998, "pid": 1728719, "tid": 1728719, "ts": 6950381117321.341, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117332.331, "dur": 0.111 , "args": { "External id": 2762,"cbid": 200, "correlation": 10013 } }, { "ph": "f", "id": 10013, "pid": 1728719, "tid": 1728719, "ts": 6950381117332.331, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117332.542, "dur": 0.150 , "args": { "External id": 2762,"cbid": 468, "correlation": 10014 } }, { "ph": "f", "id": 10014, "pid": 1728719, "tid": 1728719, "ts": 6950381117332.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117332.942, "dur": 0.041 , "args": { "External id": 2762,"cbid": 1, "correlation": 10016 } }, { "ph": "f", "id": 10016, "pid": 1728719, "tid": 1728719, "ts": 6950381117332.942, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117333.063, "dur": 0.100 , "args": { "External id": 2762,"cbid": 468, "correlation": 10017 } }, { "ph": "f", "id": 10017, "pid": 1728719, "tid": 1728719, "ts": 6950381117333.063, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117333.313, "dur": 0.020 , "args": { "External id": 2762,"cbid": 1, "correlation": 10019 } }, { "ph": "f", "id": 10019, "pid": 1728719, "tid": 1728719, "ts": 6950381117333.313, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117333.433, "dur": 0.091 , "args": { "External id": 2762,"cbid": 468, "correlation": 10020 } }, { "ph": "f", "id": 10020, "pid": 1728719, "tid": 1728719, "ts": 6950381117333.433, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117333.684, "dur": 0.030 , "args": { "External id": 2762,"cbid": 1, "correlation": 10022 } }, { "ph": "f", "id": 10022, "pid": 1728719, "tid": 1728719, "ts": 6950381117333.684, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117333.794, "dur": 0.371 , "args": { "External id": 2762,"cbid": 273, "correlation": 10023 } }, { "ph": "f", "id": 10023, "pid": 1728719, "tid": 1728719, "ts": 6950381117333.794, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381186983.625, "dur": 209.565 , "args": { "External id": 2762,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10026, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10026, "pid": 0, "tid": 7, "ts": 6950381186983.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117334.535, "dur": 1.714 , "args": { "External id": 2762,"cbid": 211, "correlation": 10026 } }, { "ph": "s", "id": 10026, "pid": 1728719, "tid": 1728719, "ts": 6950381117334.535, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381187193.478, "dur": 46.048 , "args": { "External id": 2774,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10038, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10038, "pid": 0, "tid": 7, "ts": 6950381187193.478, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117344.093, "dur": 1.663 , "args": { "External id": 2774,"cbid": 307, "correlation": 10038 } }, { "ph": "s", "id": 10038, "pid": 1728719, "tid": 1728719, "ts": 6950381117344.093, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381187239.846, "dur": 13.120 , "args": { "External id": 2775,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10044, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10044, "pid": 0, "tid": 7, "ts": 6950381187239.846, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117352.088, "dur": 1.523 , "args": { "External id": 2775,"cbid": 652, "correlation": 10044 } }, { "ph": "s", "id": 10044, "pid": 1728719, "tid": 1728719, "ts": 6950381117352.088, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381187253.254, "dur": 12.992 , "args": { "External id": 2776,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10050, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10050, "pid": 0, "tid": 7, "ts": 6950381187253.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117360.303, "dur": 1.423 , "args": { "External id": 2776,"cbid": 652, "correlation": 10050 } }, { "ph": "s", "id": 10050, "pid": 1728719, "tid": 1728719, "ts": 6950381117360.303, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117371.685, "dur": 0.100 , "args": { "External id": 2777,"cbid": 200, "correlation": 10069 } }, { "ph": "f", "id": 10069, "pid": 1728719, "tid": 1728719, "ts": 6950381117371.685, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117371.895, "dur": 0.120 , "args": { "External id": 2777,"cbid": 468, "correlation": 10070 } }, { "ph": "f", "id": 10070, "pid": 1728719, "tid": 1728719, "ts": 6950381117371.895, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117372.226, "dur": 0.020 , "args": { "External id": 2777,"cbid": 1, "correlation": 10072 } }, { "ph": "f", "id": 10072, "pid": 1728719, "tid": 1728719, "ts": 6950381117372.226, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117372.316, "dur": 0.100 , "args": { "External id": 2777,"cbid": 468, "correlation": 10073 } }, { "ph": "f", "id": 10073, "pid": 1728719, "tid": 1728719, "ts": 6950381117372.316, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117372.556, "dur": 0.030 , "args": { "External id": 2777,"cbid": 1, "correlation": 10075 } }, { "ph": "f", "id": 10075, "pid": 1728719, "tid": 1728719, "ts": 6950381117372.556, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117372.666, "dur": 0.101 , "args": { "External id": 2777,"cbid": 468, "correlation": 10076 } }, { "ph": "f", "id": 10076, "pid": 1728719, "tid": 1728719, "ts": 6950381117372.666, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117372.917, "dur": 0.020 , "args": { "External id": 2777,"cbid": 1, "correlation": 10078 } }, { "ph": "f", "id": 10078, "pid": 1728719, "tid": 1728719, "ts": 6950381117372.917, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117373.017, "dur": 0.401 , "args": { "External id": 2777,"cbid": 273, "correlation": 10079 } }, { "ph": "f", "id": 10079, "pid": 1728719, "tid": 1728719, "ts": 6950381117373.017, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381187266.533, "dur": 658.617 , "args": { "External id": 2777,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10082, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10082, "pid": 0, "tid": 7, "ts": 6950381187266.533, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117373.779, "dur": 1.713 , "args": { "External id": 2777,"cbid": 211, "correlation": 10082 } }, { "ph": "s", "id": 10082, "pid": 1728719, "tid": 1728719, "ts": 6950381117373.779, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117385.591, "dur": 0.090 , "args": { "External id": 2789,"cbid": 200, "correlation": 10106 } }, { "ph": "f", "id": 10106, "pid": 1728719, "tid": 1728719, "ts": 6950381117385.591, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117385.781, "dur": 0.110 , "args": { "External id": 2789,"cbid": 468, "correlation": 10107 } }, { "ph": "f", "id": 10107, "pid": 1728719, "tid": 1728719, "ts": 6950381117385.781, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117386.081, "dur": 0.020 , "args": { "External id": 2789,"cbid": 1, "correlation": 10109 } }, { "ph": "f", "id": 10109, "pid": 1728719, "tid": 1728719, "ts": 6950381117386.081, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117386.172, "dur": 0.100 , "args": { "External id": 2789,"cbid": 468, "correlation": 10110 } }, { "ph": "f", "id": 10110, "pid": 1728719, "tid": 1728719, "ts": 6950381117386.172, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117386.422, "dur": 0.020 , "args": { "External id": 2789,"cbid": 1, "correlation": 10112 } }, { "ph": "f", "id": 10112, "pid": 1728719, "tid": 1728719, "ts": 6950381117386.422, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117386.532, "dur": 0.100 , "args": { "External id": 2789,"cbid": 468, "correlation": 10113 } }, { "ph": "f", "id": 10113, "pid": 1728719, "tid": 1728719, "ts": 6950381117386.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117386.783, "dur": 0.020 , "args": { "External id": 2789,"cbid": 1, "correlation": 10115 } }, { "ph": "f", "id": 10115, "pid": 1728719, "tid": 1728719, "ts": 6950381117386.783, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117386.883, "dur": 0.331 , "args": { "External id": 2789,"cbid": 273, "correlation": 10116 } }, { "ph": "f", "id": 10116, "pid": 1728719, "tid": 1728719, "ts": 6950381117386.883, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381187925.438, "dur": 675.321 , "args": { "External id": 2789,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10119, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10119, "pid": 0, "tid": 7, "ts": 6950381187925.438, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117387.554, "dur": 1.663 , "args": { "External id": 2789,"cbid": 211, "correlation": 10119 } }, { "ph": "s", "id": 10119, "pid": 1728719, "tid": 1728719, "ts": 6950381117387.554, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381188601.047, "dur": 152.734 , "args": { "External id": 2801,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10132, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10132, "pid": 0, "tid": 7, "ts": 6950381188601.047, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117396.861, "dur": 1.634 , "args": { "External id": 2801,"cbid": 307, "correlation": 10132 } }, { "ph": "s", "id": 10132, "pid": 1728719, "tid": 1728719, "ts": 6950381117396.861, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381188754.069, "dur": 57.472 , "args": { "External id": 2802,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10138, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10138, "pid": 0, "tid": 7, "ts": 6950381188754.069, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117405.478, "dur": 1.803 , "args": { "External id": 2802,"cbid": 652, "correlation": 10138 } }, { "ph": "s", "id": 10138, "pid": 1728719, "tid": 1728719, "ts": 6950381117405.478, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117416.829, "dur": 0.090 , "args": { "External id": 2803,"cbid": 200, "correlation": 10157 } }, { "ph": "f", "id": 10157, "pid": 1728719, "tid": 1728719, "ts": 6950381117416.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117417.029, "dur": 0.120 , "args": { "External id": 2803,"cbid": 468, "correlation": 10158 } }, { "ph": "f", "id": 10158, "pid": 1728719, "tid": 1728719, "ts": 6950381117417.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117417.360, "dur": 0.020 , "args": { "External id": 2803,"cbid": 1, "correlation": 10160 } }, { "ph": "f", "id": 10160, "pid": 1728719, "tid": 1728719, "ts": 6950381117417.360, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117417.450, "dur": 0.100 , "args": { "External id": 2803,"cbid": 468, "correlation": 10161 } }, { "ph": "f", "id": 10161, "pid": 1728719, "tid": 1728719, "ts": 6950381117417.450, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117417.710, "dur": 0.020 , "args": { "External id": 2803,"cbid": 1, "correlation": 10163 } }, { "ph": "f", "id": 10163, "pid": 1728719, "tid": 1728719, "ts": 6950381117417.710, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117417.820, "dur": 0.091 , "args": { "External id": 2803,"cbid": 468, "correlation": 10164 } }, { "ph": "f", "id": 10164, "pid": 1728719, "tid": 1728719, "ts": 6950381117417.820, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117418.061, "dur": 0.020 , "args": { "External id": 2803,"cbid": 1, "correlation": 10166 } }, { "ph": "f", "id": 10166, "pid": 1728719, "tid": 1728719, "ts": 6950381117418.061, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117418.171, "dur": 0.351 , "args": { "External id": 2803,"cbid": 273, "correlation": 10167 } }, { "ph": "f", "id": 10167, "pid": 1728719, "tid": 1728719, "ts": 6950381117418.171, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381188811.829, "dur": 671.993 , "args": { "External id": 2803,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10170, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10170, "pid": 0, "tid": 7, "ts": 6950381188811.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117418.882, "dur": 1.694 , "args": { "External id": 2803,"cbid": 211, "correlation": 10170 } }, { "ph": "s", "id": 10170, "pid": 1728719, "tid": 1728719, "ts": 6950381117418.882, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381189484.110, "dur": 40.511 , "args": { "External id": 2815,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10181, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10181, "pid": 0, "tid": 7, "ts": 6950381189484.110, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117428.621, "dur": 1.603 , "args": { "External id": 2815,"cbid": 307, "correlation": 10181 } }, { "ph": "s", "id": 10181, "pid": 1728719, "tid": 1728719, "ts": 6950381117428.621, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381189524.909, "dur": 21.088 , "args": { "External id": 2827,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10193, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10193, "pid": 0, "tid": 7, "ts": 6950381189524.909, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117506.085, "dur": 2.474 , "args": { "External id": 2827,"cbid": 307, "correlation": 10193 } }, { "ph": "s", "id": 10193, "pid": 1728719, "tid": 1728719, "ts": 6950381117506.085, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381189546.317, "dur": 12.256 , "args": { "External id": 2828,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10207, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10207, "pid": 0, "tid": 7, "ts": 6950381189546.317, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117520.021, "dur": 1.663 , "args": { "External id": 2828,"cbid": 652, "correlation": 10207 } }, { "ph": "s", "id": 10207, "pid": 1728719, "tid": 1728719, "ts": 6950381117520.021, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381189558.893, "dur": 12.096 , "args": { "External id": 2829,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10221, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10221, "pid": 0, "tid": 7, "ts": 6950381189558.893, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117535.780, "dur": 1.503 , "args": { "External id": 2829,"cbid": 652, "correlation": 10221 } }, { "ph": "s", "id": 10221, "pid": 1728719, "tid": 1728719, "ts": 6950381117535.780, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381189571.277, "dur": 10.847 , "args": { "External id": 2830,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10235, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10235, "pid": 0, "tid": 7, "ts": 6950381189571.277, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117546.189, "dur": 1.553 , "args": { "External id": 2830,"cbid": 652, "correlation": 10235 } }, { "ph": "s", "id": 10235, "pid": 1728719, "tid": 1728719, "ts": 6950381117546.189, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117557.591, "dur": 0.140 , "args": { "External id": 2831,"cbid": 200, "correlation": 10250 } }, { "ph": "f", "id": 10250, "pid": 1728719, "tid": 1728719, "ts": 6950381117557.591, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117557.901, "dur": 0.191 , "args": { "External id": 2831,"cbid": 468, "correlation": 10251 } }, { "ph": "f", "id": 10251, "pid": 1728719, "tid": 1728719, "ts": 6950381117557.901, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117558.432, "dur": 0.030 , "args": { "External id": 2831,"cbid": 1, "correlation": 10253 } }, { "ph": "f", "id": 10253, "pid": 1728719, "tid": 1728719, "ts": 6950381117558.432, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117558.542, "dur": 0.101 , "args": { "External id": 2831,"cbid": 468, "correlation": 10254 } }, { "ph": "f", "id": 10254, "pid": 1728719, "tid": 1728719, "ts": 6950381117558.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117558.793, "dur": 0.020 , "args": { "External id": 2831,"cbid": 1, "correlation": 10256 } }, { "ph": "f", "id": 10256, "pid": 1728719, "tid": 1728719, "ts": 6950381117558.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117558.953, "dur": 0.100 , "args": { "External id": 2831,"cbid": 468, "correlation": 10257 } }, { "ph": "f", "id": 10257, "pid": 1728719, "tid": 1728719, "ts": 6950381117558.953, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117559.224, "dur": 0.020 , "args": { "External id": 2831,"cbid": 1, "correlation": 10259 } }, { "ph": "f", "id": 10259, "pid": 1728719, "tid": 1728719, "ts": 6950381117559.224, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117559.384, "dur": 0.571 , "args": { "External id": 2831,"cbid": 273, "correlation": 10260 } }, { "ph": "f", "id": 10260, "pid": 1728719, "tid": 1728719, "ts": 6950381117559.384, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381189582.412, "dur": 216.350 , "args": { "External id": 2831,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10263, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10263, "pid": 0, "tid": 7, "ts": 6950381189582.412, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117560.426, "dur": 1.853 , "args": { "External id": 2831,"cbid": 211, "correlation": 10263 } }, { "ph": "s", "id": 10263, "pid": 1728719, "tid": 1728719, "ts": 6950381117560.426, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381189799.050, "dur": 8.800 , "args": { "External id": 2843,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10276, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10276, "pid": 0, "tid": 7, "ts": 6950381189799.050, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117571.096, "dur": 1.613 , "args": { "External id": 2843,"cbid": 307, "correlation": 10276 } }, { "ph": "s", "id": 10276, "pid": 1728719, "tid": 1728719, "ts": 6950381117571.096, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381189808.106, "dur": 37.120 , "args": { "External id": 2848,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10299, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10299, "pid": 0, "tid": 7, "ts": 6950381189808.106, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117589.650, "dur": 1.643 , "args": { "External id": 2848,"cbid": 307, "correlation": 10299 } }, { "ph": "s", "id": 10299, "pid": 1728719, "tid": 1728719, "ts": 6950381117589.650, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117599.759, "dur": 0.110 , "args": { "External id": 2849,"cbid": 200, "correlation": 10314 } }, { "ph": "f", "id": 10314, "pid": 1728719, "tid": 1728719, "ts": 6950381117599.759, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117599.980, "dur": 0.130 , "args": { "External id": 2849,"cbid": 468, "correlation": 10315 } }, { "ph": "f", "id": 10315, "pid": 1728719, "tid": 1728719, "ts": 6950381117599.980, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117600.330, "dur": 0.020 , "args": { "External id": 2849,"cbid": 1, "correlation": 10317 } }, { "ph": "f", "id": 10317, "pid": 1728719, "tid": 1728719, "ts": 6950381117600.330, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117600.420, "dur": 0.101 , "args": { "External id": 2849,"cbid": 468, "correlation": 10318 } }, { "ph": "f", "id": 10318, "pid": 1728719, "tid": 1728719, "ts": 6950381117600.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117600.681, "dur": 0.030 , "args": { "External id": 2849,"cbid": 1, "correlation": 10320 } }, { "ph": "f", "id": 10320, "pid": 1728719, "tid": 1728719, "ts": 6950381117600.681, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117600.801, "dur": 0.090 , "args": { "External id": 2849,"cbid": 468, "correlation": 10321 } }, { "ph": "f", "id": 10321, "pid": 1728719, "tid": 1728719, "ts": 6950381117600.801, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117601.052, "dur": 0.020 , "args": { "External id": 2849,"cbid": 1, "correlation": 10323 } }, { "ph": "f", "id": 10323, "pid": 1728719, "tid": 1728719, "ts": 6950381117601.052, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117601.172, "dur": 0.421 , "args": { "External id": 2849,"cbid": 273, "correlation": 10324 } }, { "ph": "f", "id": 10324, "pid": 1728719, "tid": 1728719, "ts": 6950381117601.172, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381189845.514, "dur": 229.117 , "args": { "External id": 2849,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10327, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10327, "pid": 0, "tid": 7, "ts": 6950381189845.514, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117601.963, "dur": 1.864 , "args": { "External id": 2849,"cbid": 211, "correlation": 10327 } }, { "ph": "s", "id": 10327, "pid": 1728719, "tid": 1728719, "ts": 6950381117601.963, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381190074.919, "dur": 8.832 , "args": { "External id": 2861,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10336, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10336, "pid": 0, "tid": 7, "ts": 6950381190074.919, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117609.818, "dur": 1.623 , "args": { "External id": 2861,"cbid": 307, "correlation": 10336 } }, { "ph": "s", "id": 10336, "pid": 1728719, "tid": 1728719, "ts": 6950381117609.818, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381190084.039, "dur": 26.208 , "args": { "External id": 2866,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10359, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10359, "pid": 0, "tid": 7, "ts": 6950381190084.039, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117625.287, "dur": 1.593 , "args": { "External id": 2866,"cbid": 307, "correlation": 10359 } }, { "ph": "s", "id": 10359, "pid": 1728719, "tid": 1728719, "ts": 6950381117625.287, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117635.245, "dur": 0.080 , "args": { "External id": 2867,"cbid": 200, "correlation": 10374 } }, { "ph": "f", "id": 10374, "pid": 1728719, "tid": 1728719, "ts": 6950381117635.245, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117635.426, "dur": 0.120 , "args": { "External id": 2867,"cbid": 468, "correlation": 10375 } }, { "ph": "f", "id": 10375, "pid": 1728719, "tid": 1728719, "ts": 6950381117635.426, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117635.766, "dur": 0.030 , "args": { "External id": 2867,"cbid": 1, "correlation": 10377 } }, { "ph": "f", "id": 10377, "pid": 1728719, "tid": 1728719, "ts": 6950381117635.766, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117635.856, "dur": 0.101 , "args": { "External id": 2867,"cbid": 468, "correlation": 10378 } }, { "ph": "f", "id": 10378, "pid": 1728719, "tid": 1728719, "ts": 6950381117635.856, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117636.117, "dur": 0.020 , "args": { "External id": 2867,"cbid": 1, "correlation": 10380 } }, { "ph": "f", "id": 10380, "pid": 1728719, "tid": 1728719, "ts": 6950381117636.117, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117636.237, "dur": 0.100 , "args": { "External id": 2867,"cbid": 468, "correlation": 10381 } }, { "ph": "f", "id": 10381, "pid": 1728719, "tid": 1728719, "ts": 6950381117636.237, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117636.488, "dur": 0.020 , "args": { "External id": 2867,"cbid": 1, "correlation": 10383 } }, { "ph": "f", "id": 10383, "pid": 1728719, "tid": 1728719, "ts": 6950381117636.488, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117636.598, "dur": 0.370 , "args": { "External id": 2867,"cbid": 273, "correlation": 10384 } }, { "ph": "f", "id": 10384, "pid": 1728719, "tid": 1728719, "ts": 6950381117636.598, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381190110.503, "dur": 231.613 , "args": { "External id": 2867,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10387, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10387, "pid": 0, "tid": 7, "ts": 6950381190110.503, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117637.329, "dur": 1.663 , "args": { "External id": 2867,"cbid": 211, "correlation": 10387 } }, { "ph": "s", "id": 10387, "pid": 1728719, "tid": 1728719, "ts": 6950381117637.329, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381190342.404, "dur": 60.128 , "args": { "External id": 2879,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10401, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10401, "pid": 0, "tid": 7, "ts": 6950381190342.404, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117646.476, "dur": 1.553 , "args": { "External id": 2879,"cbid": 307, "correlation": 10401 } }, { "ph": "s", "id": 10401, "pid": 1728719, "tid": 1728719, "ts": 6950381117646.476, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381190402.820, "dur": 62.655 , "args": { "External id": 2880,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10411, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10411, "pid": 0, "tid": 7, "ts": 6950381190402.820, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117656.415, "dur": 1.603 , "args": { "External id": 2880,"cbid": 307, "correlation": 10411 } }, { "ph": "s", "id": 10411, "pid": 1728719, "tid": 1728719, "ts": 6950381117656.415, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381190465.763, "dur": 42.624 , "args": { "External id": 2881,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10420, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10420, "pid": 0, "tid": 7, "ts": 6950381190465.763, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117665.802, "dur": 1.643 , "args": { "External id": 2881,"cbid": 307, "correlation": 10420 } }, { "ph": "s", "id": 10420, "pid": 1728719, "tid": 1728719, "ts": 6950381117665.802, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117680.670, "dur": 0.120 , "args": { "External id": 2889,"cbid": 200, "correlation": 10455 } }, { "ph": "f", "id": 10455, "pid": 1728719, "tid": 1728719, "ts": 6950381117680.670, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117680.870, "dur": 0.070 , "args": { "External id": 2889,"cbid": 200, "correlation": 10456 } }, { "ph": "f", "id": 10456, "pid": 1728719, "tid": 1728719, "ts": 6950381117680.870, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117681.020, "dur": 0.471 , "args": { "External id": 2889,"cbid": 273, "correlation": 10457 } }, { "ph": "f", "id": 10457, "pid": 1728719, "tid": 1728719, "ts": 6950381117681.020, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381190508.643, "dur": 846.806 , "args": { "External id": 2889,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10459, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10459, "pid": 0, "tid": 7, "ts": 6950381190508.643, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117681.812, "dur": 1.893 , "args": { "External id": 2889,"cbid": 211, "correlation": 10459 } }, { "ph": "s", "id": 10459, "pid": 1728719, "tid": 1728719, "ts": 6950381117681.812, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381191355.737, "dur": 9.344 , "args": { "External id": 2898,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10479, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10479, "pid": 0, "tid": 7, "ts": 6950381191355.737, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117697.732, "dur": 1.512 , "args": { "External id": 2898,"cbid": 652, "correlation": 10479 } }, { "ph": "s", "id": 10479, "pid": 1728719, "tid": 1728719, "ts": 6950381117697.732, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117708.361, "dur": 0.101 , "args": { "External id": 2899,"cbid": 200, "correlation": 10494 } }, { "ph": "f", "id": 10494, "pid": 1728719, "tid": 1728719, "ts": 6950381117708.361, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117708.572, "dur": 0.150 , "args": { "External id": 2899,"cbid": 468, "correlation": 10495 } }, { "ph": "f", "id": 10495, "pid": 1728719, "tid": 1728719, "ts": 6950381117708.572, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117708.972, "dur": 0.031 , "args": { "External id": 2899,"cbid": 1, "correlation": 10497 } }, { "ph": "f", "id": 10497, "pid": 1728719, "tid": 1728719, "ts": 6950381117708.972, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117709.073, "dur": 0.100 , "args": { "External id": 2899,"cbid": 468, "correlation": 10498 } }, { "ph": "f", "id": 10498, "pid": 1728719, "tid": 1728719, "ts": 6950381117709.073, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117709.313, "dur": 0.030 , "args": { "External id": 2899,"cbid": 1, "correlation": 10500 } }, { "ph": "f", "id": 10500, "pid": 1728719, "tid": 1728719, "ts": 6950381117709.313, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117709.453, "dur": 0.091 , "args": { "External id": 2899,"cbid": 468, "correlation": 10501 } }, { "ph": "f", "id": 10501, "pid": 1728719, "tid": 1728719, "ts": 6950381117709.453, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117709.714, "dur": 0.020 , "args": { "External id": 2899,"cbid": 1, "correlation": 10503 } }, { "ph": "f", "id": 10503, "pid": 1728719, "tid": 1728719, "ts": 6950381117709.714, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117709.824, "dur": 0.391 , "args": { "External id": 2899,"cbid": 273, "correlation": 10504 } }, { "ph": "f", "id": 10504, "pid": 1728719, "tid": 1728719, "ts": 6950381117709.824, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381191365.337, "dur": 209.438 , "args": { "External id": 2899,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10507, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10507, "pid": 0, "tid": 7, "ts": 6950381191365.337, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117710.575, "dur": 1.694 , "args": { "External id": 2899,"cbid": 211, "correlation": 10507 } }, { "ph": "s", "id": 10507, "pid": 1728719, "tid": 1728719, "ts": 6950381117710.575, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381191575.095, "dur": 45.823 , "args": { "External id": 2911,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10519, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10519, "pid": 0, "tid": 7, "ts": 6950381191575.095, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117720.253, "dur": 1.623 , "args": { "External id": 2911,"cbid": 307, "correlation": 10519 } }, { "ph": "s", "id": 10519, "pid": 1728719, "tid": 1728719, "ts": 6950381117720.253, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381191621.238, "dur": 13.536 , "args": { "External id": 2912,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10525, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10525, "pid": 0, "tid": 7, "ts": 6950381191621.238, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117728.278, "dur": 1.483 , "args": { "External id": 2912,"cbid": 652, "correlation": 10525 } }, { "ph": "s", "id": 10525, "pid": 1728719, "tid": 1728719, "ts": 6950381117728.278, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381191635.094, "dur": 13.120 , "args": { "External id": 2913,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10531, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10531, "pid": 0, "tid": 7, "ts": 6950381191635.094, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117736.564, "dur": 1.523 , "args": { "External id": 2913,"cbid": 652, "correlation": 10531 } }, { "ph": "s", "id": 10531, "pid": 1728719, "tid": 1728719, "ts": 6950381117736.564, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117747.965, "dur": 0.100 , "args": { "External id": 2914,"cbid": 200, "correlation": 10550 } }, { "ph": "f", "id": 10550, "pid": 1728719, "tid": 1728719, "ts": 6950381117747.965, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117748.226, "dur": 0.130 , "args": { "External id": 2914,"cbid": 468, "correlation": 10551 } }, { "ph": "f", "id": 10551, "pid": 1728719, "tid": 1728719, "ts": 6950381117748.226, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117748.626, "dur": 0.020 , "args": { "External id": 2914,"cbid": 1, "correlation": 10553 } }, { "ph": "f", "id": 10553, "pid": 1728719, "tid": 1728719, "ts": 6950381117748.626, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117748.787, "dur": 0.090 , "args": { "External id": 2914,"cbid": 468, "correlation": 10554 } }, { "ph": "f", "id": 10554, "pid": 1728719, "tid": 1728719, "ts": 6950381117748.787, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117749.047, "dur": 0.020 , "args": { "External id": 2914,"cbid": 1, "correlation": 10556 } }, { "ph": "f", "id": 10556, "pid": 1728719, "tid": 1728719, "ts": 6950381117749.047, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117749.157, "dur": 0.100 , "args": { "External id": 2914,"cbid": 468, "correlation": 10557 } }, { "ph": "f", "id": 10557, "pid": 1728719, "tid": 1728719, "ts": 6950381117749.157, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117749.408, "dur": 0.020 , "args": { "External id": 2914,"cbid": 1, "correlation": 10559 } }, { "ph": "f", "id": 10559, "pid": 1728719, "tid": 1728719, "ts": 6950381117749.408, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117749.518, "dur": 0.381 , "args": { "External id": 2914,"cbid": 273, "correlation": 10560 } }, { "ph": "f", "id": 10560, "pid": 1728719, "tid": 1728719, "ts": 6950381117749.518, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381191648.534, "dur": 663.865 , "args": { "External id": 2914,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10563, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10563, "pid": 0, "tid": 7, "ts": 6950381191648.534, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117750.259, "dur": 1.663 , "args": { "External id": 2914,"cbid": 211, "correlation": 10563 } }, { "ph": "s", "id": 10563, "pid": 1728719, "tid": 1728719, "ts": 6950381117750.259, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117761.961, "dur": 0.090 , "args": { "External id": 2926,"cbid": 200, "correlation": 10587 } }, { "ph": "f", "id": 10587, "pid": 1728719, "tid": 1728719, "ts": 6950381117761.961, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117762.141, "dur": 0.111 , "args": { "External id": 2926,"cbid": 468, "correlation": 10588 } }, { "ph": "f", "id": 10588, "pid": 1728719, "tid": 1728719, "ts": 6950381117762.141, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117762.452, "dur": 0.020 , "args": { "External id": 2926,"cbid": 1, "correlation": 10590 } }, { "ph": "f", "id": 10590, "pid": 1728719, "tid": 1728719, "ts": 6950381117762.452, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117762.542, "dur": 0.090 , "args": { "External id": 2926,"cbid": 468, "correlation": 10591 } }, { "ph": "f", "id": 10591, "pid": 1728719, "tid": 1728719, "ts": 6950381117762.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117762.783, "dur": 0.020 , "args": { "External id": 2926,"cbid": 1, "correlation": 10593 } }, { "ph": "f", "id": 10593, "pid": 1728719, "tid": 1728719, "ts": 6950381117762.783, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117762.893, "dur": 0.100 , "args": { "External id": 2926,"cbid": 468, "correlation": 10594 } }, { "ph": "f", "id": 10594, "pid": 1728719, "tid": 1728719, "ts": 6950381117762.893, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117763.173, "dur": 0.030 , "args": { "External id": 2926,"cbid": 1, "correlation": 10596 } }, { "ph": "f", "id": 10596, "pid": 1728719, "tid": 1728719, "ts": 6950381117763.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117763.284, "dur": 0.330 , "args": { "External id": 2926,"cbid": 273, "correlation": 10597 } }, { "ph": "f", "id": 10597, "pid": 1728719, "tid": 1728719, "ts": 6950381117763.284, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381192312.687, "dur": 671.769 , "args": { "External id": 2926,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10600, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10600, "pid": 0, "tid": 7, "ts": 6950381192312.687, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117763.985, "dur": 1.693 , "args": { "External id": 2926,"cbid": 211, "correlation": 10600 } }, { "ph": "s", "id": 10600, "pid": 1728719, "tid": 1728719, "ts": 6950381117763.985, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381192984.776, "dur": 153.758 , "args": { "External id": 2938,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10613, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10613, "pid": 0, "tid": 7, "ts": 6950381192984.776, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117773.162, "dur": 1.743 , "args": { "External id": 2938,"cbid": 307, "correlation": 10613 } }, { "ph": "s", "id": 10613, "pid": 1728719, "tid": 1728719, "ts": 6950381117773.162, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381193138.822, "dur": 60.287 , "args": { "External id": 2939,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10619, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10619, "pid": 0, "tid": 7, "ts": 6950381193138.822, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117781.407, "dur": 1.593 , "args": { "External id": 2939,"cbid": 652, "correlation": 10619 } }, { "ph": "s", "id": 10619, "pid": 1728719, "tid": 1728719, "ts": 6950381117781.407, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117792.658, "dur": 0.080 , "args": { "External id": 2940,"cbid": 200, "correlation": 10638 } }, { "ph": "f", "id": 10638, "pid": 1728719, "tid": 1728719, "ts": 6950381117792.658, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117792.829, "dur": 0.120 , "args": { "External id": 2940,"cbid": 468, "correlation": 10639 } }, { "ph": "f", "id": 10639, "pid": 1728719, "tid": 1728719, "ts": 6950381117792.829, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117793.169, "dur": 0.020 , "args": { "External id": 2940,"cbid": 1, "correlation": 10641 } }, { "ph": "f", "id": 10641, "pid": 1728719, "tid": 1728719, "ts": 6950381117793.169, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117793.259, "dur": 0.101 , "args": { "External id": 2940,"cbid": 468, "correlation": 10642 } }, { "ph": "f", "id": 10642, "pid": 1728719, "tid": 1728719, "ts": 6950381117793.259, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117793.500, "dur": 0.020 , "args": { "External id": 2940,"cbid": 1, "correlation": 10644 } }, { "ph": "f", "id": 10644, "pid": 1728719, "tid": 1728719, "ts": 6950381117793.500, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117793.610, "dur": 0.100 , "args": { "External id": 2940,"cbid": 468, "correlation": 10645 } }, { "ph": "f", "id": 10645, "pid": 1728719, "tid": 1728719, "ts": 6950381117793.610, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117793.860, "dur": 0.021 , "args": { "External id": 2940,"cbid": 1, "correlation": 10647 } }, { "ph": "f", "id": 10647, "pid": 1728719, "tid": 1728719, "ts": 6950381117793.860, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117793.971, "dur": 0.360 , "args": { "External id": 2940,"cbid": 273, "correlation": 10648 } }, { "ph": "f", "id": 10648, "pid": 1728719, "tid": 1728719, "ts": 6950381117793.971, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381193199.397, "dur": 673.369 , "args": { "External id": 2940,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10651, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10651, "pid": 0, "tid": 7, "ts": 6950381193199.397, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117794.692, "dur": 1.643 , "args": { "External id": 2940,"cbid": 211, "correlation": 10651 } }, { "ph": "s", "id": 10651, "pid": 1728719, "tid": 1728719, "ts": 6950381117794.692, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381193873.086, "dur": 40.480 , "args": { "External id": 2952,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10662, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10662, "pid": 0, "tid": 7, "ts": 6950381193873.086, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117804.200, "dur": 1.603 , "args": { "External id": 2952,"cbid": 307, "correlation": 10662 } }, { "ph": "s", "id": 10662, "pid": 1728719, "tid": 1728719, "ts": 6950381117804.200, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381193913.822, "dur": 20.863 , "args": { "External id": 2964,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10674, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10674, "pid": 0, "tid": 7, "ts": 6950381193913.822, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117881.003, "dur": 2.554 , "args": { "External id": 2964,"cbid": 307, "correlation": 10674 } }, { "ph": "s", "id": 10674, "pid": 1728719, "tid": 1728719, "ts": 6950381117881.003, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381193935.005, "dur": 12.480 , "args": { "External id": 2965,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10688, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10688, "pid": 0, "tid": 7, "ts": 6950381193935.005, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117895.199, "dur": 1.553 , "args": { "External id": 2965,"cbid": 652, "correlation": 10688 } }, { "ph": "s", "id": 10688, "pid": 1728719, "tid": 1728719, "ts": 6950381117895.199, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381193947.805, "dur": 12.032 , "args": { "External id": 2966,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10702, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10702, "pid": 0, "tid": 7, "ts": 6950381193947.805, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117906.460, "dur": 1.573 , "args": { "External id": 2966,"cbid": 652, "correlation": 10702 } }, { "ph": "s", "id": 10702, "pid": 1728719, "tid": 1728719, "ts": 6950381117906.460, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381193960.125, "dur": 10.912 , "args": { "External id": 2967,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10716, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10716, "pid": 0, "tid": 7, "ts": 6950381193960.125, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381117917.400, "dur": 1.463 , "args": { "External id": 2967,"cbid": 652, "correlation": 10716 } }, { "ph": "s", "id": 10716, "pid": 1728719, "tid": 1728719, "ts": 6950381117917.400, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117928.411, "dur": 0.170 , "args": { "External id": 2968,"cbid": 200, "correlation": 10731 } }, { "ph": "f", "id": 10731, "pid": 1728719, "tid": 1728719, "ts": 6950381117928.411, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117928.722, "dur": 0.190 , "args": { "External id": 2968,"cbid": 468, "correlation": 10732 } }, { "ph": "f", "id": 10732, "pid": 1728719, "tid": 1728719, "ts": 6950381117928.722, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117929.222, "dur": 0.030 , "args": { "External id": 2968,"cbid": 1, "correlation": 10734 } }, { "ph": "f", "id": 10734, "pid": 1728719, "tid": 1728719, "ts": 6950381117929.222, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117929.383, "dur": 0.100 , "args": { "External id": 2968,"cbid": 468, "correlation": 10735 } }, { "ph": "f", "id": 10735, "pid": 1728719, "tid": 1728719, "ts": 6950381117929.383, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117929.653, "dur": 0.020 , "args": { "External id": 2968,"cbid": 1, "correlation": 10737 } }, { "ph": "f", "id": 10737, "pid": 1728719, "tid": 1728719, "ts": 6950381117929.653, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117929.773, "dur": 0.111 , "args": { "External id": 2968,"cbid": 468, "correlation": 10738 } }, { "ph": "f", "id": 10738, "pid": 1728719, "tid": 1728719, "ts": 6950381117929.773, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117930.044, "dur": 0.020 , "args": { "External id": 2968,"cbid": 1, "correlation": 10740 } }, { "ph": "f", "id": 10740, "pid": 1728719, "tid": 1728719, "ts": 6950381117930.044, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117930.174, "dur": 0.651 , "args": { "External id": 2968,"cbid": 273, "correlation": 10741 } }, { "ph": "f", "id": 10741, "pid": 1728719, "tid": 1728719, "ts": 6950381117930.174, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381193971.325, "dur": 216.030 , "args": { "External id": 2968,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10744, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10744, "pid": 0, "tid": 7, "ts": 6950381193971.325, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117931.296, "dur": 1.794 , "args": { "External id": 2968,"cbid": 211, "correlation": 10744 } }, { "ph": "s", "id": 10744, "pid": 1728719, "tid": 1728719, "ts": 6950381117931.296, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381194187.642, "dur": 8.801 , "args": { "External id": 2980,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10757, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10757, "pid": 0, "tid": 7, "ts": 6950381194187.642, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117941.565, "dur": 1.583 , "args": { "External id": 2980,"cbid": 307, "correlation": 10757 } }, { "ph": "s", "id": 10757, "pid": 1728719, "tid": 1728719, "ts": 6950381117941.565, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381194196.731, "dur": 37.919 , "args": { "External id": 2985,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10780, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10780, "pid": 0, "tid": 7, "ts": 6950381194196.731, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117960.731, "dur": 1.703 , "args": { "External id": 2985,"cbid": 307, "correlation": 10780 } }, { "ph": "s", "id": 10780, "pid": 1728719, "tid": 1728719, "ts": 6950381117960.731, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117971.000, "dur": 0.100 , "args": { "External id": 2986,"cbid": 200, "correlation": 10795 } }, { "ph": "f", "id": 10795, "pid": 1728719, "tid": 1728719, "ts": 6950381117971.000, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117971.201, "dur": 0.140 , "args": { "External id": 2986,"cbid": 468, "correlation": 10796 } }, { "ph": "f", "id": 10796, "pid": 1728719, "tid": 1728719, "ts": 6950381117971.201, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117971.561, "dur": 0.030 , "args": { "External id": 2986,"cbid": 1, "correlation": 10798 } }, { "ph": "f", "id": 10798, "pid": 1728719, "tid": 1728719, "ts": 6950381117971.561, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117971.661, "dur": 0.101 , "args": { "External id": 2986,"cbid": 468, "correlation": 10799 } }, { "ph": "f", "id": 10799, "pid": 1728719, "tid": 1728719, "ts": 6950381117971.661, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117971.912, "dur": 0.020 , "args": { "External id": 2986,"cbid": 1, "correlation": 10801 } }, { "ph": "f", "id": 10801, "pid": 1728719, "tid": 1728719, "ts": 6950381117971.912, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117972.022, "dur": 0.100 , "args": { "External id": 2986,"cbid": 468, "correlation": 10802 } }, { "ph": "f", "id": 10802, "pid": 1728719, "tid": 1728719, "ts": 6950381117972.022, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381117972.283, "dur": 0.020 , "args": { "External id": 2986,"cbid": 1, "correlation": 10804 } }, { "ph": "f", "id": 10804, "pid": 1728719, "tid": 1728719, "ts": 6950381117972.283, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381117972.413, "dur": 0.441 , "args": { "External id": 2986,"cbid": 273, "correlation": 10805 } }, { "ph": "f", "id": 10805, "pid": 1728719, "tid": 1728719, "ts": 6950381117972.413, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381194234.938, "dur": 226.558 , "args": { "External id": 2986,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10808, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10808, "pid": 0, "tid": 7, "ts": 6950381194234.938, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117973.224, "dur": 1.744 , "args": { "External id": 2986,"cbid": 211, "correlation": 10808 } }, { "ph": "s", "id": 10808, "pid": 1728719, "tid": 1728719, "ts": 6950381117973.224, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381194461.816, "dur": 8.864 , "args": { "External id": 2998,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10817, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10817, "pid": 0, "tid": 7, "ts": 6950381194461.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117980.808, "dur": 1.593 , "args": { "External id": 2998,"cbid": 307, "correlation": 10817 } }, { "ph": "s", "id": 10817, "pid": 1728719, "tid": 1728719, "ts": 6950381117980.808, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381194470.936, "dur": 27.167 , "args": { "External id": 3003,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10840, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10840, "pid": 0, "tid": 7, "ts": 6950381194470.936, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381117996.448, "dur": 1.552 , "args": { "External id": 3003,"cbid": 307, "correlation": 10840 } }, { "ph": "s", "id": 10840, "pid": 1728719, "tid": 1728719, "ts": 6950381117996.448, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118006.406, "dur": 0.100 , "args": { "External id": 3004,"cbid": 200, "correlation": 10855 } }, { "ph": "f", "id": 10855, "pid": 1728719, "tid": 1728719, "ts": 6950381118006.406, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118006.596, "dur": 0.131 , "args": { "External id": 3004,"cbid": 468, "correlation": 10856 } }, { "ph": "f", "id": 10856, "pid": 1728719, "tid": 1728719, "ts": 6950381118006.596, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118006.937, "dur": 0.030 , "args": { "External id": 3004,"cbid": 1, "correlation": 10858 } }, { "ph": "f", "id": 10858, "pid": 1728719, "tid": 1728719, "ts": 6950381118006.937, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118007.037, "dur": 0.090 , "args": { "External id": 3004,"cbid": 468, "correlation": 10859 } }, { "ph": "f", "id": 10859, "pid": 1728719, "tid": 1728719, "ts": 6950381118007.037, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118007.288, "dur": 0.030 , "args": { "External id": 3004,"cbid": 1, "correlation": 10861 } }, { "ph": "f", "id": 10861, "pid": 1728719, "tid": 1728719, "ts": 6950381118007.288, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118007.408, "dur": 0.100 , "args": { "External id": 3004,"cbid": 468, "correlation": 10862 } }, { "ph": "f", "id": 10862, "pid": 1728719, "tid": 1728719, "ts": 6950381118007.408, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118007.658, "dur": 0.020 , "args": { "External id": 3004,"cbid": 1, "correlation": 10864 } }, { "ph": "f", "id": 10864, "pid": 1728719, "tid": 1728719, "ts": 6950381118007.658, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118007.779, "dur": 0.370 , "args": { "External id": 3004,"cbid": 273, "correlation": 10865 } }, { "ph": "f", "id": 10865, "pid": 1728719, "tid": 1728719, "ts": 6950381118007.779, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381194498.359, "dur": 230.238 , "args": { "External id": 3004,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10868, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10868, "pid": 0, "tid": 7, "ts": 6950381194498.359, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118008.540, "dur": 1.703 , "args": { "External id": 3004,"cbid": 211, "correlation": 10868 } }, { "ph": "s", "id": 10868, "pid": 1728719, "tid": 1728719, "ts": 6950381118008.540, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381194728.885, "dur": 59.391 , "args": { "External id": 3016,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10882, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10882, "pid": 0, "tid": 7, "ts": 6950381194728.885, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118017.938, "dur": 1.573 , "args": { "External id": 3016,"cbid": 307, "correlation": 10882 } }, { "ph": "s", "id": 10882, "pid": 1728719, "tid": 1728719, "ts": 6950381118017.938, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381194788.532, "dur": 66.847 , "args": { "External id": 3017,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10892, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10892, "pid": 0, "tid": 7, "ts": 6950381194788.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118027.696, "dur": 1.853 , "args": { "External id": 3017,"cbid": 307, "correlation": 10892 } }, { "ph": "s", "id": 10892, "pid": 1728719, "tid": 1728719, "ts": 6950381118027.696, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381194855.635, "dur": 37.472 , "args": { "External id": 3018,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10901, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10901, "pid": 0, "tid": 7, "ts": 6950381194855.635, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118037.324, "dur": 1.683 , "args": { "External id": 3018,"cbid": 307, "correlation": 10901 } }, { "ph": "s", "id": 10901, "pid": 1728719, "tid": 1728719, "ts": 6950381118037.324, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118052.091, "dur": 0.110 , "args": { "External id": 3026,"cbid": 200, "correlation": 10936 } }, { "ph": "f", "id": 10936, "pid": 1728719, "tid": 1728719, "ts": 6950381118052.091, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118052.292, "dur": 0.070 , "args": { "External id": 3026,"cbid": 200, "correlation": 10937 } }, { "ph": "f", "id": 10937, "pid": 1728719, "tid": 1728719, "ts": 6950381118052.292, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118052.452, "dur": 0.491 , "args": { "External id": 3026,"cbid": 273, "correlation": 10938 } }, { "ph": "f", "id": 10938, "pid": 1728719, "tid": 1728719, "ts": 6950381118052.452, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381194893.395, "dur": 846.391 , "args": { "External id": 3026,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10940, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10940, "pid": 0, "tid": 7, "ts": 6950381194893.395, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118053.263, "dur": 1.884 , "args": { "External id": 3026,"cbid": 211, "correlation": 10940 } }, { "ph": "s", "id": 10940, "pid": 1728719, "tid": 1728719, "ts": 6950381118053.263, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381195740.042, "dur": 9.408 , "args": { "External id": 3035,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10960, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10960, "pid": 0, "tid": 7, "ts": 6950381195740.042, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118069.754, "dur": 1.613 , "args": { "External id": 3035,"cbid": 652, "correlation": 10960 } }, { "ph": "s", "id": 10960, "pid": 1728719, "tid": 1728719, "ts": 6950381118069.754, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118080.514, "dur": 0.100 , "args": { "External id": 3036,"cbid": 200, "correlation": 10975 } }, { "ph": "f", "id": 10975, "pid": 1728719, "tid": 1728719, "ts": 6950381118080.514, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118080.724, "dur": 0.151 , "args": { "External id": 3036,"cbid": 468, "correlation": 10976 } }, { "ph": "f", "id": 10976, "pid": 1728719, "tid": 1728719, "ts": 6950381118080.724, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118081.145, "dur": 0.030 , "args": { "External id": 3036,"cbid": 1, "correlation": 10978 } }, { "ph": "f", "id": 10978, "pid": 1728719, "tid": 1728719, "ts": 6950381118081.145, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118081.245, "dur": 0.101 , "args": { "External id": 3036,"cbid": 468, "correlation": 10979 } }, { "ph": "f", "id": 10979, "pid": 1728719, "tid": 1728719, "ts": 6950381118081.245, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118081.496, "dur": 0.030 , "args": { "External id": 3036,"cbid": 1, "correlation": 10981 } }, { "ph": "f", "id": 10981, "pid": 1728719, "tid": 1728719, "ts": 6950381118081.496, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118081.616, "dur": 0.100 , "args": { "External id": 3036,"cbid": 468, "correlation": 10982 } }, { "ph": "f", "id": 10982, "pid": 1728719, "tid": 1728719, "ts": 6950381118081.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118081.877, "dur": 0.020 , "args": { "External id": 3036,"cbid": 1, "correlation": 10984 } }, { "ph": "f", "id": 10984, "pid": 1728719, "tid": 1728719, "ts": 6950381118081.877, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118082.007, "dur": 0.391 , "args": { "External id": 3036,"cbid": 273, "correlation": 10985 } }, { "ph": "f", "id": 10985, "pid": 1728719, "tid": 1728719, "ts": 6950381118082.007, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381195749.738, "dur": 208.989 , "args": { "External id": 3036,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 10988, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 10988, "pid": 0, "tid": 7, "ts": 6950381195749.738, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118082.758, "dur": 1.814 , "args": { "External id": 3036,"cbid": 211, "correlation": 10988 } }, { "ph": "s", "id": 10988, "pid": 1728719, "tid": 1728719, "ts": 6950381118082.758, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381195958.983, "dur": 46.464 , "args": { "External id": 3048,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11000, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11000, "pid": 0, "tid": 7, "ts": 6950381195958.983, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118092.526, "dur": 1.653 , "args": { "External id": 3048,"cbid": 307, "correlation": 11000 } }, { "ph": "s", "id": 11000, "pid": 1728719, "tid": 1728719, "ts": 6950381118092.526, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381196005.767, "dur": 13.088 , "args": { "External id": 3049,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11006, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11006, "pid": 0, "tid": 7, "ts": 6950381196005.767, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118100.642, "dur": 1.623 , "args": { "External id": 3049,"cbid": 652, "correlation": 11006 } }, { "ph": "s", "id": 11006, "pid": 1728719, "tid": 1728719, "ts": 6950381118100.642, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381196019.175, "dur": 12.576 , "args": { "External id": 3050,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11012, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11012, "pid": 0, "tid": 7, "ts": 6950381196019.175, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118108.947, "dur": 1.503 , "args": { "External id": 3050,"cbid": 652, "correlation": 11012 } }, { "ph": "s", "id": 11012, "pid": 1728719, "tid": 1728719, "ts": 6950381118108.947, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118120.549, "dur": 0.090 , "args": { "External id": 3051,"cbid": 200, "correlation": 11031 } }, { "ph": "f", "id": 11031, "pid": 1728719, "tid": 1728719, "ts": 6950381118120.549, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118120.759, "dur": 0.120 , "args": { "External id": 3051,"cbid": 468, "correlation": 11032 } }, { "ph": "f", "id": 11032, "pid": 1728719, "tid": 1728719, "ts": 6950381118120.759, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118121.080, "dur": 0.020 , "args": { "External id": 3051,"cbid": 1, "correlation": 11034 } }, { "ph": "f", "id": 11034, "pid": 1728719, "tid": 1728719, "ts": 6950381118121.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118121.170, "dur": 0.100 , "args": { "External id": 3051,"cbid": 468, "correlation": 11035 } }, { "ph": "f", "id": 11035, "pid": 1728719, "tid": 1728719, "ts": 6950381118121.170, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118121.410, "dur": 0.030 , "args": { "External id": 3051,"cbid": 1, "correlation": 11037 } }, { "ph": "f", "id": 11037, "pid": 1728719, "tid": 1728719, "ts": 6950381118121.410, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118121.520, "dur": 0.101 , "args": { "External id": 3051,"cbid": 468, "correlation": 11038 } }, { "ph": "f", "id": 11038, "pid": 1728719, "tid": 1728719, "ts": 6950381118121.520, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118121.801, "dur": 0.020 , "args": { "External id": 3051,"cbid": 1, "correlation": 11040 } }, { "ph": "f", "id": 11040, "pid": 1728719, "tid": 1728719, "ts": 6950381118121.801, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118121.901, "dur": 0.381 , "args": { "External id": 3051,"cbid": 273, "correlation": 11041 } }, { "ph": "f", "id": 11041, "pid": 1728719, "tid": 1728719, "ts": 6950381118121.901, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381196032.071, "dur": 658.905 , "args": { "External id": 3051,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11044, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11044, "pid": 0, "tid": 7, "ts": 6950381196032.071, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118122.632, "dur": 1.694 , "args": { "External id": 3051,"cbid": 211, "correlation": 11044 } }, { "ph": "s", "id": 11044, "pid": 1728719, "tid": 1728719, "ts": 6950381118122.632, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118137.530, "dur": 0.090 , "args": { "External id": 3063,"cbid": 200, "correlation": 11068 } }, { "ph": "f", "id": 11068, "pid": 1728719, "tid": 1728719, "ts": 6950381118137.530, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118137.721, "dur": 0.120 , "args": { "External id": 3063,"cbid": 468, "correlation": 11069 } }, { "ph": "f", "id": 11069, "pid": 1728719, "tid": 1728719, "ts": 6950381118137.721, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118138.181, "dur": 0.020 , "args": { "External id": 3063,"cbid": 1, "correlation": 11071 } }, { "ph": "f", "id": 11071, "pid": 1728719, "tid": 1728719, "ts": 6950381118138.181, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118138.382, "dur": 0.100 , "args": { "External id": 3063,"cbid": 468, "correlation": 11072 } }, { "ph": "f", "id": 11072, "pid": 1728719, "tid": 1728719, "ts": 6950381118138.382, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118138.632, "dur": 0.030 , "args": { "External id": 3063,"cbid": 1, "correlation": 11074 } }, { "ph": "f", "id": 11074, "pid": 1728719, "tid": 1728719, "ts": 6950381118138.632, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118138.742, "dur": 0.101 , "args": { "External id": 3063,"cbid": 468, "correlation": 11075 } }, { "ph": "f", "id": 11075, "pid": 1728719, "tid": 1728719, "ts": 6950381118138.742, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118139.013, "dur": 0.020 , "args": { "External id": 3063,"cbid": 1, "correlation": 11077 } }, { "ph": "f", "id": 11077, "pid": 1728719, "tid": 1728719, "ts": 6950381118139.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118139.123, "dur": 0.351 , "args": { "External id": 3063,"cbid": 273, "correlation": 11078 } }, { "ph": "f", "id": 11078, "pid": 1728719, "tid": 1728719, "ts": 6950381118139.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381196691.264, "dur": 676.472 , "args": { "External id": 3063,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11081, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11081, "pid": 0, "tid": 7, "ts": 6950381196691.264, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118139.824, "dur": 1.704 , "args": { "External id": 3063,"cbid": 211, "correlation": 11081 } }, { "ph": "s", "id": 11081, "pid": 1728719, "tid": 1728719, "ts": 6950381118139.824, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381197368.056, "dur": 153.343 , "args": { "External id": 3075,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11094, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11094, "pid": 0, "tid": 7, "ts": 6950381197368.056, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118149.152, "dur": 1.863 , "args": { "External id": 3075,"cbid": 307, "correlation": 11094 } }, { "ph": "s", "id": 11094, "pid": 1728719, "tid": 1728719, "ts": 6950381118149.152, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381197521.686, "dur": 61.408 , "args": { "External id": 3076,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11100, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11100, "pid": 0, "tid": 7, "ts": 6950381197521.686, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118157.728, "dur": 1.533 , "args": { "External id": 3076,"cbid": 652, "correlation": 11100 } }, { "ph": "s", "id": 11100, "pid": 1728719, "tid": 1728719, "ts": 6950381118157.728, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118169.039, "dur": 0.090 , "args": { "External id": 3077,"cbid": 200, "correlation": 11119 } }, { "ph": "f", "id": 11119, "pid": 1728719, "tid": 1728719, "ts": 6950381118169.039, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118169.229, "dur": 0.120 , "args": { "External id": 3077,"cbid": 468, "correlation": 11120 } }, { "ph": "f", "id": 11120, "pid": 1728719, "tid": 1728719, "ts": 6950381118169.229, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118169.560, "dur": 0.030 , "args": { "External id": 3077,"cbid": 1, "correlation": 11122 } }, { "ph": "f", "id": 11122, "pid": 1728719, "tid": 1728719, "ts": 6950381118169.560, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118169.650, "dur": 0.100 , "args": { "External id": 3077,"cbid": 468, "correlation": 11123 } }, { "ph": "f", "id": 11123, "pid": 1728719, "tid": 1728719, "ts": 6950381118169.650, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118169.900, "dur": 0.020 , "args": { "External id": 3077,"cbid": 1, "correlation": 11125 } }, { "ph": "f", "id": 11125, "pid": 1728719, "tid": 1728719, "ts": 6950381118169.900, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118170.011, "dur": 0.100 , "args": { "External id": 3077,"cbid": 468, "correlation": 11126 } }, { "ph": "f", "id": 11126, "pid": 1728719, "tid": 1728719, "ts": 6950381118170.011, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118170.301, "dur": 0.020 , "args": { "External id": 3077,"cbid": 1, "correlation": 11128 } }, { "ph": "f", "id": 11128, "pid": 1728719, "tid": 1728719, "ts": 6950381118170.301, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118170.401, "dur": 0.391 , "args": { "External id": 3077,"cbid": 273, "correlation": 11129 } }, { "ph": "f", "id": 11129, "pid": 1728719, "tid": 1728719, "ts": 6950381118170.401, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381197583.382, "dur": 669.945 , "args": { "External id": 3077,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11132, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11132, "pid": 0, "tid": 7, "ts": 6950381197583.382, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118171.143, "dur": 1.693 , "args": { "External id": 3077,"cbid": 211, "correlation": 11132 } }, { "ph": "s", "id": 11132, "pid": 1728719, "tid": 1728719, "ts": 6950381118171.143, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381198253.615, "dur": 40.415 , "args": { "External id": 3089,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11143, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11143, "pid": 0, "tid": 7, "ts": 6950381198253.615, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118181.061, "dur": 1.703 , "args": { "External id": 3089,"cbid": 307, "correlation": 11143 } }, { "ph": "s", "id": 11143, "pid": 1728719, "tid": 1728719, "ts": 6950381118181.061, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381198294.318, "dur": 20.928 , "args": { "External id": 3101,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11155, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11155, "pid": 0, "tid": 7, "ts": 6950381198294.318, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118259.377, "dur": 2.384 , "args": { "External id": 3101,"cbid": 307, "correlation": 11155 } }, { "ph": "s", "id": 11155, "pid": 1728719, "tid": 1728719, "ts": 6950381118259.377, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381198315.566, "dur": 12.480 , "args": { "External id": 3102,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11169, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11169, "pid": 0, "tid": 7, "ts": 6950381198315.566, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118273.133, "dur": 1.713 , "args": { "External id": 3102,"cbid": 652, "correlation": 11169 } }, { "ph": "s", "id": 11169, "pid": 1728719, "tid": 1728719, "ts": 6950381118273.133, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381198328.366, "dur": 12.256 , "args": { "External id": 3103,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11183, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11183, "pid": 0, "tid": 7, "ts": 6950381198328.366, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118284.734, "dur": 1.503 , "args": { "External id": 3103,"cbid": 652, "correlation": 11183 } }, { "ph": "s", "id": 11183, "pid": 1728719, "tid": 1728719, "ts": 6950381118284.734, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381198340.910, "dur": 10.847 , "args": { "External id": 3104,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11197, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11197, "pid": 0, "tid": 7, "ts": 6950381198340.910, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118295.013, "dur": 1.834 , "args": { "External id": 3104,"cbid": 652, "correlation": 11197 } }, { "ph": "s", "id": 11197, "pid": 1728719, "tid": 1728719, "ts": 6950381118295.013, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118306.765, "dur": 0.150 , "args": { "External id": 3105,"cbid": 200, "correlation": 11212 } }, { "ph": "f", "id": 11212, "pid": 1728719, "tid": 1728719, "ts": 6950381118306.765, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118307.066, "dur": 0.170 , "args": { "External id": 3105,"cbid": 468, "correlation": 11213 } }, { "ph": "f", "id": 11213, "pid": 1728719, "tid": 1728719, "ts": 6950381118307.066, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118307.557, "dur": 0.030 , "args": { "External id": 3105,"cbid": 1, "correlation": 11215 } }, { "ph": "f", "id": 11215, "pid": 1728719, "tid": 1728719, "ts": 6950381118307.557, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118307.657, "dur": 0.170 , "args": { "External id": 3105,"cbid": 468, "correlation": 11216 } }, { "ph": "f", "id": 11216, "pid": 1728719, "tid": 1728719, "ts": 6950381118307.657, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118308.058, "dur": 0.020 , "args": { "External id": 3105,"cbid": 1, "correlation": 11218 } }, { "ph": "f", "id": 11218, "pid": 1728719, "tid": 1728719, "ts": 6950381118308.058, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118308.178, "dur": 0.100 , "args": { "External id": 3105,"cbid": 468, "correlation": 11219 } }, { "ph": "f", "id": 11219, "pid": 1728719, "tid": 1728719, "ts": 6950381118308.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118308.458, "dur": 0.020 , "args": { "External id": 3105,"cbid": 1, "correlation": 11221 } }, { "ph": "f", "id": 11221, "pid": 1728719, "tid": 1728719, "ts": 6950381118308.458, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118308.609, "dur": 0.561 , "args": { "External id": 3105,"cbid": 273, "correlation": 11222 } }, { "ph": "f", "id": 11222, "pid": 1728719, "tid": 1728719, "ts": 6950381118308.609, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381198352.045, "dur": 216.318 , "args": { "External id": 3105,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11225, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11225, "pid": 0, "tid": 7, "ts": 6950381198352.045, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118309.671, "dur": 1.893 , "args": { "External id": 3105,"cbid": 211, "correlation": 11225 } }, { "ph": "s", "id": 11225, "pid": 1728719, "tid": 1728719, "ts": 6950381118309.671, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381198568.651, "dur": 8.864 , "args": { "External id": 3117,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11238, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11238, "pid": 0, "tid": 7, "ts": 6950381198568.651, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118320.190, "dur": 1.603 , "args": { "External id": 3117,"cbid": 307, "correlation": 11238 } }, { "ph": "s", "id": 11238, "pid": 1728719, "tid": 1728719, "ts": 6950381118320.190, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381198577.803, "dur": 36.576 , "args": { "External id": 3122,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11261, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11261, "pid": 0, "tid": 7, "ts": 6950381198577.803, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118339.386, "dur": 1.693 , "args": { "External id": 3122,"cbid": 307, "correlation": 11261 } }, { "ph": "s", "id": 11261, "pid": 1728719, "tid": 1728719, "ts": 6950381118339.386, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118349.455, "dur": 0.100 , "args": { "External id": 3123,"cbid": 200, "correlation": 11276 } }, { "ph": "f", "id": 11276, "pid": 1728719, "tid": 1728719, "ts": 6950381118349.455, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118349.655, "dur": 0.140 , "args": { "External id": 3123,"cbid": 468, "correlation": 11277 } }, { "ph": "f", "id": 11277, "pid": 1728719, "tid": 1728719, "ts": 6950381118349.655, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118350.036, "dur": 0.030 , "args": { "External id": 3123,"cbid": 1, "correlation": 11279 } }, { "ph": "f", "id": 11279, "pid": 1728719, "tid": 1728719, "ts": 6950381118350.036, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118350.216, "dur": 0.090 , "args": { "External id": 3123,"cbid": 468, "correlation": 11280 } }, { "ph": "f", "id": 11280, "pid": 1728719, "tid": 1728719, "ts": 6950381118350.216, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118350.477, "dur": 0.020 , "args": { "External id": 3123,"cbid": 1, "correlation": 11282 } }, { "ph": "f", "id": 11282, "pid": 1728719, "tid": 1728719, "ts": 6950381118350.477, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118350.587, "dur": 0.090 , "args": { "External id": 3123,"cbid": 468, "correlation": 11283 } }, { "ph": "f", "id": 11283, "pid": 1728719, "tid": 1728719, "ts": 6950381118350.587, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118350.837, "dur": 0.020 , "args": { "External id": 3123,"cbid": 1, "correlation": 11285 } }, { "ph": "f", "id": 11285, "pid": 1728719, "tid": 1728719, "ts": 6950381118350.837, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118350.957, "dur": 0.431 , "args": { "External id": 3123,"cbid": 273, "correlation": 11286 } }, { "ph": "f", "id": 11286, "pid": 1728719, "tid": 1728719, "ts": 6950381118350.957, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381198614.667, "dur": 228.125 , "args": { "External id": 3123,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11289, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11289, "pid": 0, "tid": 7, "ts": 6950381198614.667, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118351.809, "dur": 1.643 , "args": { "External id": 3123,"cbid": 211, "correlation": 11289 } }, { "ph": "s", "id": 11289, "pid": 1728719, "tid": 1728719, "ts": 6950381118351.809, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381198843.080, "dur": 8.832 , "args": { "External id": 3135,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11298, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11298, "pid": 0, "tid": 7, "ts": 6950381198843.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118359.684, "dur": 1.523 , "args": { "External id": 3135,"cbid": 307, "correlation": 11298 } }, { "ph": "s", "id": 11298, "pid": 1728719, "tid": 1728719, "ts": 6950381118359.684, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381198852.200, "dur": 26.944 , "args": { "External id": 3140,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11321, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11321, "pid": 0, "tid": 7, "ts": 6950381198852.200, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118375.032, "dur": 1.543 , "args": { "External id": 3140,"cbid": 307, "correlation": 11321 } }, { "ph": "s", "id": 11321, "pid": 1728719, "tid": 1728719, "ts": 6950381118375.032, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118384.780, "dur": 0.091 , "args": { "External id": 3141,"cbid": 200, "correlation": 11336 } }, { "ph": "f", "id": 11336, "pid": 1728719, "tid": 1728719, "ts": 6950381118384.780, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118384.951, "dur": 0.130 , "args": { "External id": 3141,"cbid": 468, "correlation": 11337 } }, { "ph": "f", "id": 11337, "pid": 1728719, "tid": 1728719, "ts": 6950381118384.951, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118385.281, "dur": 0.030 , "args": { "External id": 3141,"cbid": 1, "correlation": 11339 } }, { "ph": "f", "id": 11339, "pid": 1728719, "tid": 1728719, "ts": 6950381118385.281, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118385.382, "dur": 0.100 , "args": { "External id": 3141,"cbid": 468, "correlation": 11340 } }, { "ph": "f", "id": 11340, "pid": 1728719, "tid": 1728719, "ts": 6950381118385.382, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118385.632, "dur": 0.020 , "args": { "External id": 3141,"cbid": 1, "correlation": 11342 } }, { "ph": "f", "id": 11342, "pid": 1728719, "tid": 1728719, "ts": 6950381118385.632, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118385.752, "dur": 0.090 , "args": { "External id": 3141,"cbid": 468, "correlation": 11343 } }, { "ph": "f", "id": 11343, "pid": 1728719, "tid": 1728719, "ts": 6950381118385.752, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118386.013, "dur": 0.020 , "args": { "External id": 3141,"cbid": 1, "correlation": 11345 } }, { "ph": "f", "id": 11345, "pid": 1728719, "tid": 1728719, "ts": 6950381118386.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118386.143, "dur": 0.371 , "args": { "External id": 3141,"cbid": 273, "correlation": 11346 } }, { "ph": "f", "id": 11346, "pid": 1728719, "tid": 1728719, "ts": 6950381118386.143, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381198879.400, "dur": 230.173 , "args": { "External id": 3141,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11349, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11349, "pid": 0, "tid": 7, "ts": 6950381198879.400, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118386.874, "dur": 1.734 , "args": { "External id": 3141,"cbid": 211, "correlation": 11349 } }, { "ph": "s", "id": 11349, "pid": 1728719, "tid": 1728719, "ts": 6950381118386.874, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381199109.861, "dur": 59.359 , "args": { "External id": 3153,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11363, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11363, "pid": 0, "tid": 7, "ts": 6950381199109.861, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118396.202, "dur": 1.543 , "args": { "External id": 3153,"cbid": 307, "correlation": 11363 } }, { "ph": "s", "id": 11363, "pid": 1728719, "tid": 1728719, "ts": 6950381118396.202, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381199169.477, "dur": 65.727 , "args": { "External id": 3154,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11373, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11373, "pid": 0, "tid": 7, "ts": 6950381199169.477, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118406.120, "dur": 1.623 , "args": { "External id": 3154,"cbid": 307, "correlation": 11373 } }, { "ph": "s", "id": 11373, "pid": 1728719, "tid": 1728719, "ts": 6950381118406.120, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381199235.492, "dur": 39.199 , "args": { "External id": 3155,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11382, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11382, "pid": 0, "tid": 7, "ts": 6950381199235.492, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118415.568, "dur": 1.603 , "args": { "External id": 3155,"cbid": 307, "correlation": 11382 } }, { "ph": "s", "id": 11382, "pid": 1728719, "tid": 1728719, "ts": 6950381118415.568, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118430.425, "dur": 0.111 , "args": { "External id": 3163,"cbid": 200, "correlation": 11417 } }, { "ph": "f", "id": 11417, "pid": 1728719, "tid": 1728719, "ts": 6950381118430.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118430.636, "dur": 0.070 , "args": { "External id": 3163,"cbid": 200, "correlation": 11418 } }, { "ph": "f", "id": 11418, "pid": 1728719, "tid": 1728719, "ts": 6950381118430.636, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118430.806, "dur": 0.451 , "args": { "External id": 3163,"cbid": 273, "correlation": 11419 } }, { "ph": "f", "id": 11419, "pid": 1728719, "tid": 1728719, "ts": 6950381118430.806, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381199274.979, "dur": 842.935 , "args": { "External id": 3163,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11421, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11421, "pid": 0, "tid": 7, "ts": 6950381199274.979, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118431.578, "dur": 1.823 , "args": { "External id": 3163,"cbid": 211, "correlation": 11421 } }, { "ph": "s", "id": 11421, "pid": 1728719, "tid": 1728719, "ts": 6950381118431.578, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381200118.202, "dur": 9.344 , "args": { "External id": 3172,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11441, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11441, "pid": 0, "tid": 7, "ts": 6950381200118.202, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118448.058, "dur": 1.623 , "args": { "External id": 3172,"cbid": 652, "correlation": 11441 } }, { "ph": "s", "id": 11441, "pid": 1728719, "tid": 1728719, "ts": 6950381118448.058, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118458.848, "dur": 0.100 , "args": { "External id": 3173,"cbid": 200, "correlation": 11456 } }, { "ph": "f", "id": 11456, "pid": 1728719, "tid": 1728719, "ts": 6950381118458.848, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118459.049, "dur": 0.150 , "args": { "External id": 3173,"cbid": 468, "correlation": 11457 } }, { "ph": "f", "id": 11457, "pid": 1728719, "tid": 1728719, "ts": 6950381118459.049, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118459.429, "dur": 0.020 , "args": { "External id": 3173,"cbid": 1, "correlation": 11459 } }, { "ph": "f", "id": 11459, "pid": 1728719, "tid": 1728719, "ts": 6950381118459.429, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118459.520, "dur": 0.110 , "args": { "External id": 3173,"cbid": 468, "correlation": 11460 } }, { "ph": "f", "id": 11460, "pid": 1728719, "tid": 1728719, "ts": 6950381118459.520, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118459.770, "dur": 0.020 , "args": { "External id": 3173,"cbid": 1, "correlation": 11462 } }, { "ph": "f", "id": 11462, "pid": 1728719, "tid": 1728719, "ts": 6950381118459.770, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118459.890, "dur": 0.090 , "args": { "External id": 3173,"cbid": 468, "correlation": 11463 } }, { "ph": "f", "id": 11463, "pid": 1728719, "tid": 1728719, "ts": 6950381118459.890, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118460.151, "dur": 0.020 , "args": { "External id": 3173,"cbid": 1, "correlation": 11465 } }, { "ph": "f", "id": 11465, "pid": 1728719, "tid": 1728719, "ts": 6950381118460.151, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118460.281, "dur": 0.381 , "args": { "External id": 3173,"cbid": 273, "correlation": 11466 } }, { "ph": "f", "id": 11466, "pid": 1728719, "tid": 1728719, "ts": 6950381118460.281, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381200127.802, "dur": 209.662 , "args": { "External id": 3173,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11469, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11469, "pid": 0, "tid": 7, "ts": 6950381200127.802, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118461.042, "dur": 1.764 , "args": { "External id": 3173,"cbid": 211, "correlation": 11469 } }, { "ph": "s", "id": 11469, "pid": 1728719, "tid": 1728719, "ts": 6950381118461.042, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381200337.752, "dur": 45.568 , "args": { "External id": 3185,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11481, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11481, "pid": 0, "tid": 7, "ts": 6950381200337.752, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118470.781, "dur": 1.623 , "args": { "External id": 3185,"cbid": 307, "correlation": 11481 } }, { "ph": "s", "id": 11481, "pid": 1728719, "tid": 1728719, "ts": 6950381118470.781, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381200383.639, "dur": 13.440 , "args": { "External id": 3186,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11487, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11487, "pid": 0, "tid": 7, "ts": 6950381200383.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118479.056, "dur": 1.463 , "args": { "External id": 3186,"cbid": 652, "correlation": 11487 } }, { "ph": "s", "id": 11487, "pid": 1728719, "tid": 1728719, "ts": 6950381118479.056, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381200397.399, "dur": 13.184 , "args": { "External id": 3187,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11493, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11493, "pid": 0, "tid": 7, "ts": 6950381200397.399, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118487.782, "dur": 1.513 , "args": { "External id": 3187,"cbid": 652, "correlation": 11493 } }, { "ph": "s", "id": 11493, "pid": 1728719, "tid": 1728719, "ts": 6950381118487.782, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118499.314, "dur": 0.100 , "args": { "External id": 3188,"cbid": 200, "correlation": 11512 } }, { "ph": "f", "id": 11512, "pid": 1728719, "tid": 1728719, "ts": 6950381118499.314, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118499.534, "dur": 0.120 , "args": { "External id": 3188,"cbid": 468, "correlation": 11513 } }, { "ph": "f", "id": 11513, "pid": 1728719, "tid": 1728719, "ts": 6950381118499.534, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118499.865, "dur": 0.030 , "args": { "External id": 3188,"cbid": 1, "correlation": 11515 } }, { "ph": "f", "id": 11515, "pid": 1728719, "tid": 1728719, "ts": 6950381118499.865, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118499.955, "dur": 0.110 , "args": { "External id": 3188,"cbid": 468, "correlation": 11516 } }, { "ph": "f", "id": 11516, "pid": 1728719, "tid": 1728719, "ts": 6950381118499.955, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118500.225, "dur": 0.030 , "args": { "External id": 3188,"cbid": 1, "correlation": 11518 } }, { "ph": "f", "id": 11518, "pid": 1728719, "tid": 1728719, "ts": 6950381118500.225, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118500.346, "dur": 0.090 , "args": { "External id": 3188,"cbid": 468, "correlation": 11519 } }, { "ph": "f", "id": 11519, "pid": 1728719, "tid": 1728719, "ts": 6950381118500.346, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118500.616, "dur": 0.020 , "args": { "External id": 3188,"cbid": 1, "correlation": 11521 } }, { "ph": "f", "id": 11521, "pid": 1728719, "tid": 1728719, "ts": 6950381118500.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118500.716, "dur": 0.391 , "args": { "External id": 3188,"cbid": 273, "correlation": 11522 } }, { "ph": "f", "id": 11522, "pid": 1728719, "tid": 1728719, "ts": 6950381118500.716, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381200410.903, "dur": 657.625 , "args": { "External id": 3188,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11525, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11525, "pid": 0, "tid": 7, "ts": 6950381200410.903, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118501.458, "dur": 1.693 , "args": { "External id": 3188,"cbid": 211, "correlation": 11525 } }, { "ph": "s", "id": 11525, "pid": 1728719, "tid": 1728719, "ts": 6950381118501.458, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118513.119, "dur": 0.090 , "args": { "External id": 3200,"cbid": 200, "correlation": 11549 } }, { "ph": "f", "id": 11549, "pid": 1728719, "tid": 1728719, "ts": 6950381118513.119, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118513.310, "dur": 0.120 , "args": { "External id": 3200,"cbid": 468, "correlation": 11550 } }, { "ph": "f", "id": 11550, "pid": 1728719, "tid": 1728719, "ts": 6950381118513.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118513.620, "dur": 0.020 , "args": { "External id": 3200,"cbid": 1, "correlation": 11552 } }, { "ph": "f", "id": 11552, "pid": 1728719, "tid": 1728719, "ts": 6950381118513.620, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118513.700, "dur": 0.111 , "args": { "External id": 3200,"cbid": 468, "correlation": 11553 } }, { "ph": "f", "id": 11553, "pid": 1728719, "tid": 1728719, "ts": 6950381118513.700, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118513.961, "dur": 0.020 , "args": { "External id": 3200,"cbid": 1, "correlation": 11555 } }, { "ph": "f", "id": 11555, "pid": 1728719, "tid": 1728719, "ts": 6950381118513.961, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118514.071, "dur": 0.090 , "args": { "External id": 3200,"cbid": 468, "correlation": 11556 } }, { "ph": "f", "id": 11556, "pid": 1728719, "tid": 1728719, "ts": 6950381118514.071, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118514.362, "dur": 0.030 , "args": { "External id": 3200,"cbid": 1, "correlation": 11558 } }, { "ph": "f", "id": 11558, "pid": 1728719, "tid": 1728719, "ts": 6950381118514.362, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118514.472, "dur": 0.330 , "args": { "External id": 3200,"cbid": 273, "correlation": 11559 } }, { "ph": "f", "id": 11559, "pid": 1728719, "tid": 1728719, "ts": 6950381118514.472, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381201068.848, "dur": 678.233 , "args": { "External id": 3200,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11562, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11562, "pid": 0, "tid": 7, "ts": 6950381201068.848, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118515.173, "dur": 1.693 , "args": { "External id": 3200,"cbid": 211, "correlation": 11562 } }, { "ph": "s", "id": 11562, "pid": 1728719, "tid": 1728719, "ts": 6950381118515.173, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381201747.369, "dur": 153.406 , "args": { "External id": 3212,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11575, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11575, "pid": 0, "tid": 7, "ts": 6950381201747.369, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118524.450, "dur": 1.633 , "args": { "External id": 3212,"cbid": 307, "correlation": 11575 } }, { "ph": "s", "id": 11575, "pid": 1728719, "tid": 1728719, "ts": 6950381118524.450, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381201901.031, "dur": 64.255 , "args": { "External id": 3213,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11581, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11581, "pid": 0, "tid": 7, "ts": 6950381201901.031, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118535.671, "dur": 1.523 , "args": { "External id": 3213,"cbid": 652, "correlation": 11581 } }, { "ph": "s", "id": 11581, "pid": 1728719, "tid": 1728719, "ts": 6950381118535.671, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118547.073, "dur": 0.100 , "args": { "External id": 3214,"cbid": 200, "correlation": 11600 } }, { "ph": "f", "id": 11600, "pid": 1728719, "tid": 1728719, "ts": 6950381118547.073, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118547.283, "dur": 0.120 , "args": { "External id": 3214,"cbid": 468, "correlation": 11601 } }, { "ph": "f", "id": 11601, "pid": 1728719, "tid": 1728719, "ts": 6950381118547.283, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118547.614, "dur": 0.030 , "args": { "External id": 3214,"cbid": 1, "correlation": 11603 } }, { "ph": "f", "id": 11603, "pid": 1728719, "tid": 1728719, "ts": 6950381118547.614, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118547.714, "dur": 0.090 , "args": { "External id": 3214,"cbid": 468, "correlation": 11604 } }, { "ph": "f", "id": 11604, "pid": 1728719, "tid": 1728719, "ts": 6950381118547.714, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118547.944, "dur": 0.020 , "args": { "External id": 3214,"cbid": 1, "correlation": 11606 } }, { "ph": "f", "id": 11606, "pid": 1728719, "tid": 1728719, "ts": 6950381118547.944, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118548.054, "dur": 0.101 , "args": { "External id": 3214,"cbid": 468, "correlation": 11607 } }, { "ph": "f", "id": 11607, "pid": 1728719, "tid": 1728719, "ts": 6950381118548.054, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118548.315, "dur": 0.030 , "args": { "External id": 3214,"cbid": 1, "correlation": 11609 } }, { "ph": "f", "id": 11609, "pid": 1728719, "tid": 1728719, "ts": 6950381118548.315, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118548.435, "dur": 0.391 , "args": { "External id": 3214,"cbid": 273, "correlation": 11610 } }, { "ph": "f", "id": 11610, "pid": 1728719, "tid": 1728719, "ts": 6950381118548.435, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381201965.574, "dur": 673.337 , "args": { "External id": 3214,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11613, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11613, "pid": 0, "tid": 7, "ts": 6950381201965.574, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118549.186, "dur": 1.694 , "args": { "External id": 3214,"cbid": 211, "correlation": 11613 } }, { "ph": "s", "id": 11613, "pid": 1728719, "tid": 1728719, "ts": 6950381118549.186, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381202639.199, "dur": 40.672 , "args": { "External id": 3226,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11624, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11624, "pid": 0, "tid": 7, "ts": 6950381202639.199, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118559.065, "dur": 1.663 , "args": { "External id": 3226,"cbid": 307, "correlation": 11624 } }, { "ph": "s", "id": 11624, "pid": 1728719, "tid": 1728719, "ts": 6950381118559.065, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381202680.159, "dur": 21.343 , "args": { "External id": 3238,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11636, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11636, "pid": 0, "tid": 7, "ts": 6950381202680.159, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118638.493, "dur": 2.474 , "args": { "External id": 3238,"cbid": 307, "correlation": 11636 } }, { "ph": "s", "id": 11636, "pid": 1728719, "tid": 1728719, "ts": 6950381118638.493, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381202701.790, "dur": 12.608 , "args": { "External id": 3239,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11650, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11650, "pid": 0, "tid": 7, "ts": 6950381202701.790, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118652.449, "dur": 1.593 , "args": { "External id": 3239,"cbid": 652, "correlation": 11650 } }, { "ph": "s", "id": 11650, "pid": 1728719, "tid": 1728719, "ts": 6950381118652.449, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381202714.718, "dur": 12.192 , "args": { "External id": 3240,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11664, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11664, "pid": 0, "tid": 7, "ts": 6950381202714.718, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118664.010, "dur": 1.573 , "args": { "External id": 3240,"cbid": 652, "correlation": 11664 } }, { "ph": "s", "id": 11664, "pid": 1728719, "tid": 1728719, "ts": 6950381118664.010, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381202727.198, "dur": 10.688 , "args": { "External id": 3241,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11678, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11678, "pid": 0, "tid": 7, "ts": 6950381202727.198, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118674.359, "dur": 1.523 , "args": { "External id": 3241,"cbid": 652, "correlation": 11678 } }, { "ph": "s", "id": 11678, "pid": 1728719, "tid": 1728719, "ts": 6950381118674.359, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118685.831, "dur": 0.150 , "args": { "External id": 3242,"cbid": 200, "correlation": 11693 } }, { "ph": "f", "id": 11693, "pid": 1728719, "tid": 1728719, "ts": 6950381118685.831, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118686.141, "dur": 0.201 , "args": { "External id": 3242,"cbid": 468, "correlation": 11694 } }, { "ph": "f", "id": 11694, "pid": 1728719, "tid": 1728719, "ts": 6950381118686.141, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118686.662, "dur": 0.040 , "args": { "External id": 3242,"cbid": 1, "correlation": 11696 } }, { "ph": "f", "id": 11696, "pid": 1728719, "tid": 1728719, "ts": 6950381118686.662, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118686.783, "dur": 0.100 , "args": { "External id": 3242,"cbid": 468, "correlation": 11697 } }, { "ph": "f", "id": 11697, "pid": 1728719, "tid": 1728719, "ts": 6950381118686.783, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118687.043, "dur": 0.030 , "args": { "External id": 3242,"cbid": 1, "correlation": 11699 } }, { "ph": "f", "id": 11699, "pid": 1728719, "tid": 1728719, "ts": 6950381118687.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118687.183, "dur": 0.101 , "args": { "External id": 3242,"cbid": 468, "correlation": 11700 } }, { "ph": "f", "id": 11700, "pid": 1728719, "tid": 1728719, "ts": 6950381118687.183, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118687.494, "dur": 0.020 , "args": { "External id": 3242,"cbid": 1, "correlation": 11702 } }, { "ph": "f", "id": 11702, "pid": 1728719, "tid": 1728719, "ts": 6950381118687.494, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118687.664, "dur": 0.621 , "args": { "External id": 3242,"cbid": 273, "correlation": 11703 } }, { "ph": "f", "id": 11703, "pid": 1728719, "tid": 1728719, "ts": 6950381118687.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381202738.174, "dur": 215.422 , "args": { "External id": 3242,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11706, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11706, "pid": 0, "tid": 7, "ts": 6950381202738.174, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118688.726, "dur": 1.844 , "args": { "External id": 3242,"cbid": 211, "correlation": 11706 } }, { "ph": "s", "id": 11706, "pid": 1728719, "tid": 1728719, "ts": 6950381118688.726, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381202953.916, "dur": 8.800 , "args": { "External id": 3254,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11719, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11719, "pid": 0, "tid": 7, "ts": 6950381202953.916, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118699.096, "dur": 1.592 , "args": { "External id": 3254,"cbid": 307, "correlation": 11719 } }, { "ph": "s", "id": 11719, "pid": 1728719, "tid": 1728719, "ts": 6950381118699.096, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381202963.004, "dur": 37.023 , "args": { "External id": 3259,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11742, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11742, "pid": 0, "tid": 7, "ts": 6950381202963.004, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118717.790, "dur": 1.683 , "args": { "External id": 3259,"cbid": 307, "correlation": 11742 } }, { "ph": "s", "id": 11742, "pid": 1728719, "tid": 1728719, "ts": 6950381118717.790, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118728.089, "dur": 0.101 , "args": { "External id": 3260,"cbid": 200, "correlation": 11757 } }, { "ph": "f", "id": 11757, "pid": 1728719, "tid": 1728719, "ts": 6950381118728.089, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118728.300, "dur": 0.140 , "args": { "External id": 3260,"cbid": 468, "correlation": 11758 } }, { "ph": "f", "id": 11758, "pid": 1728719, "tid": 1728719, "ts": 6950381118728.300, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118728.661, "dur": 0.030 , "args": { "External id": 3260,"cbid": 1, "correlation": 11760 } }, { "ph": "f", "id": 11760, "pid": 1728719, "tid": 1728719, "ts": 6950381118728.661, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118728.751, "dur": 0.100 , "args": { "External id": 3260,"cbid": 468, "correlation": 11761 } }, { "ph": "f", "id": 11761, "pid": 1728719, "tid": 1728719, "ts": 6950381118728.751, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118729.001, "dur": 0.020 , "args": { "External id": 3260,"cbid": 1, "correlation": 11763 } }, { "ph": "f", "id": 11763, "pid": 1728719, "tid": 1728719, "ts": 6950381118729.001, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118729.121, "dur": 0.091 , "args": { "External id": 3260,"cbid": 468, "correlation": 11764 } }, { "ph": "f", "id": 11764, "pid": 1728719, "tid": 1728719, "ts": 6950381118729.121, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118729.372, "dur": 0.020 , "args": { "External id": 3260,"cbid": 1, "correlation": 11766 } }, { "ph": "f", "id": 11766, "pid": 1728719, "tid": 1728719, "ts": 6950381118729.372, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118729.482, "dur": 0.421 , "args": { "External id": 3260,"cbid": 273, "correlation": 11767 } }, { "ph": "f", "id": 11767, "pid": 1728719, "tid": 1728719, "ts": 6950381118729.482, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381203000.315, "dur": 227.710 , "args": { "External id": 3260,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11770, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11770, "pid": 0, "tid": 7, "ts": 6950381203000.315, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118730.304, "dur": 1.783 , "args": { "External id": 3260,"cbid": 211, "correlation": 11770 } }, { "ph": "s", "id": 11770, "pid": 1728719, "tid": 1728719, "ts": 6950381118730.304, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381203228.313, "dur": 8.736 , "args": { "External id": 3272,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11779, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11779, "pid": 0, "tid": 7, "ts": 6950381203228.313, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118737.968, "dur": 1.643 , "args": { "External id": 3272,"cbid": 307, "correlation": 11779 } }, { "ph": "s", "id": 11779, "pid": 1728719, "tid": 1728719, "ts": 6950381118737.968, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381203237.337, "dur": 26.751 , "args": { "External id": 3277,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11802, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11802, "pid": 0, "tid": 7, "ts": 6950381203237.337, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118754.449, "dur": 1.562 , "args": { "External id": 3277,"cbid": 307, "correlation": 11802 } }, { "ph": "s", "id": 11802, "pid": 1728719, "tid": 1728719, "ts": 6950381118754.449, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118764.507, "dur": 0.090 , "args": { "External id": 3278,"cbid": 200, "correlation": 11817 } }, { "ph": "f", "id": 11817, "pid": 1728719, "tid": 1728719, "ts": 6950381118764.507, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118764.688, "dur": 0.120 , "args": { "External id": 3278,"cbid": 468, "correlation": 11818 } }, { "ph": "f", "id": 11818, "pid": 1728719, "tid": 1728719, "ts": 6950381118764.688, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118764.998, "dur": 0.020 , "args": { "External id": 3278,"cbid": 1, "correlation": 11820 } }, { "ph": "f", "id": 11820, "pid": 1728719, "tid": 1728719, "ts": 6950381118764.998, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118765.088, "dur": 0.101 , "args": { "External id": 3278,"cbid": 468, "correlation": 11821 } }, { "ph": "f", "id": 11821, "pid": 1728719, "tid": 1728719, "ts": 6950381118765.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118765.349, "dur": 0.020 , "args": { "External id": 3278,"cbid": 1, "correlation": 11823 } }, { "ph": "f", "id": 11823, "pid": 1728719, "tid": 1728719, "ts": 6950381118765.349, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118765.459, "dur": 0.100 , "args": { "External id": 3278,"cbid": 468, "correlation": 11824 } }, { "ph": "f", "id": 11824, "pid": 1728719, "tid": 1728719, "ts": 6950381118765.459, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118765.699, "dur": 0.021 , "args": { "External id": 3278,"cbid": 1, "correlation": 11826 } }, { "ph": "f", "id": 11826, "pid": 1728719, "tid": 1728719, "ts": 6950381118765.699, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118765.810, "dur": 0.390 , "args": { "External id": 3278,"cbid": 273, "correlation": 11827 } }, { "ph": "f", "id": 11827, "pid": 1728719, "tid": 1728719, "ts": 6950381118765.810, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381203264.376, "dur": 230.302 , "args": { "External id": 3278,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11830, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11830, "pid": 0, "tid": 7, "ts": 6950381203264.376, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118766.551, "dur": 1.733 , "args": { "External id": 3278,"cbid": 211, "correlation": 11830 } }, { "ph": "s", "id": 11830, "pid": 1728719, "tid": 1728719, "ts": 6950381118766.551, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381203494.966, "dur": 59.903 , "args": { "External id": 3290,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11844, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11844, "pid": 0, "tid": 7, "ts": 6950381203494.966, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118775.778, "dur": 1.593 , "args": { "External id": 3290,"cbid": 307, "correlation": 11844 } }, { "ph": "s", "id": 11844, "pid": 1728719, "tid": 1728719, "ts": 6950381118775.778, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381203555.157, "dur": 63.072 , "args": { "External id": 3291,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11854, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11854, "pid": 0, "tid": 7, "ts": 6950381203555.157, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118785.677, "dur": 1.643 , "args": { "External id": 3291,"cbid": 307, "correlation": 11854 } }, { "ph": "s", "id": 11854, "pid": 1728719, "tid": 1728719, "ts": 6950381118785.677, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381203618.485, "dur": 42.335 , "args": { "External id": 3292,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11863, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11863, "pid": 0, "tid": 7, "ts": 6950381203618.485, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118795.024, "dur": 1.663 , "args": { "External id": 3292,"cbid": 307, "correlation": 11863 } }, { "ph": "s", "id": 11863, "pid": 1728719, "tid": 1728719, "ts": 6950381118795.024, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118810.022, "dur": 0.110 , "args": { "External id": 3300,"cbid": 200, "correlation": 11898 } }, { "ph": "f", "id": 11898, "pid": 1728719, "tid": 1728719, "ts": 6950381118810.022, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118810.202, "dur": 0.070 , "args": { "External id": 3300,"cbid": 200, "correlation": 11899 } }, { "ph": "f", "id": 11899, "pid": 1728719, "tid": 1728719, "ts": 6950381118810.202, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118810.363, "dur": 0.440 , "args": { "External id": 3300,"cbid": 273, "correlation": 11900 } }, { "ph": "f", "id": 11900, "pid": 1728719, "tid": 1728719, "ts": 6950381118810.363, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381203661.076, "dur": 838.199 , "args": { "External id": 3300,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11902, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11902, "pid": 0, "tid": 7, "ts": 6950381203661.076, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118811.124, "dur": 1.954 , "args": { "External id": 3300,"cbid": 211, "correlation": 11902 } }, { "ph": "s", "id": 11902, "pid": 1728719, "tid": 1728719, "ts": 6950381118811.124, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381204499.531, "dur": 9.216 , "args": { "External id": 3309,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11922, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11922, "pid": 0, "tid": 7, "ts": 6950381204499.531, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118827.324, "dur": 1.603 , "args": { "External id": 3309,"cbid": 652, "correlation": 11922 } }, { "ph": "s", "id": 11922, "pid": 1728719, "tid": 1728719, "ts": 6950381118827.324, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118838.325, "dur": 0.170 , "args": { "External id": 3310,"cbid": 200, "correlation": 11937 } }, { "ph": "f", "id": 11937, "pid": 1728719, "tid": 1728719, "ts": 6950381118838.325, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118838.605, "dur": 0.150 , "args": { "External id": 3310,"cbid": 468, "correlation": 11938 } }, { "ph": "f", "id": 11938, "pid": 1728719, "tid": 1728719, "ts": 6950381118838.605, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118839.056, "dur": 0.020 , "args": { "External id": 3310,"cbid": 1, "correlation": 11940 } }, { "ph": "f", "id": 11940, "pid": 1728719, "tid": 1728719, "ts": 6950381118839.056, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118839.146, "dur": 0.100 , "args": { "External id": 3310,"cbid": 468, "correlation": 11941 } }, { "ph": "f", "id": 11941, "pid": 1728719, "tid": 1728719, "ts": 6950381118839.146, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118839.397, "dur": 0.020 , "args": { "External id": 3310,"cbid": 1, "correlation": 11943 } }, { "ph": "f", "id": 11943, "pid": 1728719, "tid": 1728719, "ts": 6950381118839.397, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118839.517, "dur": 0.100 , "args": { "External id": 3310,"cbid": 468, "correlation": 11944 } }, { "ph": "f", "id": 11944, "pid": 1728719, "tid": 1728719, "ts": 6950381118839.517, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118839.807, "dur": 0.020 , "args": { "External id": 3310,"cbid": 1, "correlation": 11946 } }, { "ph": "f", "id": 11946, "pid": 1728719, "tid": 1728719, "ts": 6950381118839.807, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118839.918, "dur": 0.390 , "args": { "External id": 3310,"cbid": 273, "correlation": 11947 } }, { "ph": "f", "id": 11947, "pid": 1728719, "tid": 1728719, "ts": 6950381118839.918, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381204509.035, "dur": 207.069 , "args": { "External id": 3310,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11950, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11950, "pid": 0, "tid": 7, "ts": 6950381204509.035, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118840.669, "dur": 1.743 , "args": { "External id": 3310,"cbid": 211, "correlation": 11950 } }, { "ph": "s", "id": 11950, "pid": 1728719, "tid": 1728719, "ts": 6950381118840.669, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381204716.424, "dur": 47.232 , "args": { "External id": 3322,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11962, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11962, "pid": 0, "tid": 7, "ts": 6950381204716.424, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118850.437, "dur": 1.673 , "args": { "External id": 3322,"cbid": 307, "correlation": 11962 } }, { "ph": "s", "id": 11962, "pid": 1728719, "tid": 1728719, "ts": 6950381118850.437, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381204763.944, "dur": 13.056 , "args": { "External id": 3323,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11968, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11968, "pid": 0, "tid": 7, "ts": 6950381204763.944, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118858.252, "dur": 1.462 , "args": { "External id": 3323,"cbid": 652, "correlation": 11968 } }, { "ph": "s", "id": 11968, "pid": 1728719, "tid": 1728719, "ts": 6950381118858.252, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381204777.288, "dur": 12.768 , "args": { "External id": 3324,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 11974, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 11974, "pid": 0, "tid": 7, "ts": 6950381204777.288, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118866.307, "dur": 1.442 , "args": { "External id": 3324,"cbid": 652, "correlation": 11974 } }, { "ph": "s", "id": 11974, "pid": 1728719, "tid": 1728719, "ts": 6950381118866.307, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118877.748, "dur": 0.090 , "args": { "External id": 3325,"cbid": 200, "correlation": 11993 } }, { "ph": "f", "id": 11993, "pid": 1728719, "tid": 1728719, "ts": 6950381118877.748, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118877.928, "dur": 0.121 , "args": { "External id": 3325,"cbid": 468, "correlation": 11994 } }, { "ph": "f", "id": 11994, "pid": 1728719, "tid": 1728719, "ts": 6950381118877.928, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118878.269, "dur": 0.020 , "args": { "External id": 3325,"cbid": 1, "correlation": 11996 } }, { "ph": "f", "id": 11996, "pid": 1728719, "tid": 1728719, "ts": 6950381118878.269, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118878.359, "dur": 0.100 , "args": { "External id": 3325,"cbid": 468, "correlation": 11997 } }, { "ph": "f", "id": 11997, "pid": 1728719, "tid": 1728719, "ts": 6950381118878.359, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118878.610, "dur": 0.020 , "args": { "External id": 3325,"cbid": 1, "correlation": 11999 } }, { "ph": "f", "id": 11999, "pid": 1728719, "tid": 1728719, "ts": 6950381118878.610, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118878.720, "dur": 0.100 , "args": { "External id": 3325,"cbid": 468, "correlation": 12000 } }, { "ph": "f", "id": 12000, "pid": 1728719, "tid": 1728719, "ts": 6950381118878.720, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118878.970, "dur": 0.020 , "args": { "External id": 3325,"cbid": 1, "correlation": 12002 } }, { "ph": "f", "id": 12002, "pid": 1728719, "tid": 1728719, "ts": 6950381118878.970, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118879.081, "dur": 0.370 , "args": { "External id": 3325,"cbid": 273, "correlation": 12003 } }, { "ph": "f", "id": 12003, "pid": 1728719, "tid": 1728719, "ts": 6950381118879.081, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381204790.344, "dur": 663.832 , "args": { "External id": 3325,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12006, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12006, "pid": 0, "tid": 7, "ts": 6950381204790.344, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118879.822, "dur": 1.673 , "args": { "External id": 3325,"cbid": 211, "correlation": 12006 } }, { "ph": "s", "id": 12006, "pid": 1728719, "tid": 1728719, "ts": 6950381118879.822, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118891.313, "dur": 0.090 , "args": { "External id": 3337,"cbid": 200, "correlation": 12030 } }, { "ph": "f", "id": 12030, "pid": 1728719, "tid": 1728719, "ts": 6950381118891.313, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118891.494, "dur": 0.110 , "args": { "External id": 3337,"cbid": 468, "correlation": 12031 } }, { "ph": "f", "id": 12031, "pid": 1728719, "tid": 1728719, "ts": 6950381118891.494, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118891.814, "dur": 0.030 , "args": { "External id": 3337,"cbid": 1, "correlation": 12033 } }, { "ph": "f", "id": 12033, "pid": 1728719, "tid": 1728719, "ts": 6950381118891.814, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118891.904, "dur": 0.101 , "args": { "External id": 3337,"cbid": 468, "correlation": 12034 } }, { "ph": "f", "id": 12034, "pid": 1728719, "tid": 1728719, "ts": 6950381118891.904, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118892.155, "dur": 0.020 , "args": { "External id": 3337,"cbid": 1, "correlation": 12036 } }, { "ph": "f", "id": 12036, "pid": 1728719, "tid": 1728719, "ts": 6950381118892.155, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118892.265, "dur": 0.090 , "args": { "External id": 3337,"cbid": 468, "correlation": 12037 } }, { "ph": "f", "id": 12037, "pid": 1728719, "tid": 1728719, "ts": 6950381118892.265, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118892.506, "dur": 0.020 , "args": { "External id": 3337,"cbid": 1, "correlation": 12039 } }, { "ph": "f", "id": 12039, "pid": 1728719, "tid": 1728719, "ts": 6950381118892.506, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118892.616, "dur": 0.330 , "args": { "External id": 3337,"cbid": 273, "correlation": 12040 } }, { "ph": "f", "id": 12040, "pid": 1728719, "tid": 1728719, "ts": 6950381118892.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381205454.464, "dur": 677.753 , "args": { "External id": 3337,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12043, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12043, "pid": 0, "tid": 7, "ts": 6950381205454.464, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118893.297, "dur": 1.673 , "args": { "External id": 3337,"cbid": 211, "correlation": 12043 } }, { "ph": "s", "id": 12043, "pid": 1728719, "tid": 1728719, "ts": 6950381118893.297, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381206132.537, "dur": 153.918 , "args": { "External id": 3349,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12056, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12056, "pid": 0, "tid": 7, "ts": 6950381206132.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118902.574, "dur": 1.583 , "args": { "External id": 3349,"cbid": 307, "correlation": 12056 } }, { "ph": "s", "id": 12056, "pid": 1728719, "tid": 1728719, "ts": 6950381118902.574, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381206286.712, "dur": 61.887 , "args": { "External id": 3350,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12062, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12062, "pid": 0, "tid": 7, "ts": 6950381206286.712, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381118910.870, "dur": 1.583 , "args": { "External id": 3350,"cbid": 652, "correlation": 12062 } }, { "ph": "s", "id": 12062, "pid": 1728719, "tid": 1728719, "ts": 6950381118910.870, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118922.161, "dur": 0.080 , "args": { "External id": 3351,"cbid": 200, "correlation": 12081 } }, { "ph": "f", "id": 12081, "pid": 1728719, "tid": 1728719, "ts": 6950381118922.161, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118922.321, "dur": 0.120 , "args": { "External id": 3351,"cbid": 468, "correlation": 12082 } }, { "ph": "f", "id": 12082, "pid": 1728719, "tid": 1728719, "ts": 6950381118922.321, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118922.662, "dur": 0.030 , "args": { "External id": 3351,"cbid": 1, "correlation": 12084 } }, { "ph": "f", "id": 12084, "pid": 1728719, "tid": 1728719, "ts": 6950381118922.662, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118922.762, "dur": 0.100 , "args": { "External id": 3351,"cbid": 468, "correlation": 12085 } }, { "ph": "f", "id": 12085, "pid": 1728719, "tid": 1728719, "ts": 6950381118922.762, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118923.002, "dur": 0.020 , "args": { "External id": 3351,"cbid": 1, "correlation": 12087 } }, { "ph": "f", "id": 12087, "pid": 1728719, "tid": 1728719, "ts": 6950381118923.002, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118923.112, "dur": 0.101 , "args": { "External id": 3351,"cbid": 468, "correlation": 12088 } }, { "ph": "f", "id": 12088, "pid": 1728719, "tid": 1728719, "ts": 6950381118923.112, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381118923.363, "dur": 0.020 , "args": { "External id": 3351,"cbid": 1, "correlation": 12090 } }, { "ph": "f", "id": 12090, "pid": 1728719, "tid": 1728719, "ts": 6950381118923.363, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381118923.473, "dur": 0.361 , "args": { "External id": 3351,"cbid": 273, "correlation": 12091 } }, { "ph": "f", "id": 12091, "pid": 1728719, "tid": 1728719, "ts": 6950381118923.473, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381206348.855, "dur": 676.313 , "args": { "External id": 3351,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12094, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12094, "pid": 0, "tid": 7, "ts": 6950381206348.855, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118924.195, "dur": 1.663 , "args": { "External id": 3351,"cbid": 211, "correlation": 12094 } }, { "ph": "s", "id": 12094, "pid": 1728719, "tid": 1728719, "ts": 6950381118924.195, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381207025.456, "dur": 40.223 , "args": { "External id": 3363,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12105, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12105, "pid": 0, "tid": 7, "ts": 6950381207025.456, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381118933.963, "dur": 1.703 , "args": { "External id": 3363,"cbid": 307, "correlation": 12105 } }, { "ph": "s", "id": 12105, "pid": 1728719, "tid": 1728719, "ts": 6950381118933.963, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381207065.967, "dur": 21.344 , "args": { "External id": 3375,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12117, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12117, "pid": 0, "tid": 7, "ts": 6950381207065.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119010.565, "dur": 2.455 , "args": { "External id": 3375,"cbid": 307, "correlation": 12117 } }, { "ph": "s", "id": 12117, "pid": 1728719, "tid": 1728719, "ts": 6950381119010.565, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381207087.631, "dur": 12.832 , "args": { "External id": 3376,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12131, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12131, "pid": 0, "tid": 7, "ts": 6950381207087.631, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119024.481, "dur": 1.613 , "args": { "External id": 3376,"cbid": 652, "correlation": 12131 } }, { "ph": "s", "id": 12131, "pid": 1728719, "tid": 1728719, "ts": 6950381119024.481, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381207100.751, "dur": 12.192 , "args": { "External id": 3377,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12145, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12145, "pid": 0, "tid": 7, "ts": 6950381207100.751, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119035.842, "dur": 1.473 , "args": { "External id": 3377,"cbid": 652, "correlation": 12145 } }, { "ph": "s", "id": 12145, "pid": 1728719, "tid": 1728719, "ts": 6950381119035.842, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381207113.231, "dur": 10.655 , "args": { "External id": 3378,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12159, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12159, "pid": 0, "tid": 7, "ts": 6950381207113.231, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119046.492, "dur": 1.513 , "args": { "External id": 3378,"cbid": 652, "correlation": 12159 } }, { "ph": "s", "id": 12159, "pid": 1728719, "tid": 1728719, "ts": 6950381119046.492, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119057.773, "dur": 0.120 , "args": { "External id": 3379,"cbid": 200, "correlation": 12174 } }, { "ph": "f", "id": 12174, "pid": 1728719, "tid": 1728719, "ts": 6950381119057.773, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119058.024, "dur": 0.170 , "args": { "External id": 3379,"cbid": 468, "correlation": 12175 } }, { "ph": "f", "id": 12175, "pid": 1728719, "tid": 1728719, "ts": 6950381119058.024, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119058.545, "dur": 0.030 , "args": { "External id": 3379,"cbid": 1, "correlation": 12177 } }, { "ph": "f", "id": 12177, "pid": 1728719, "tid": 1728719, "ts": 6950381119058.545, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119058.645, "dur": 0.100 , "args": { "External id": 3379,"cbid": 468, "correlation": 12178 } }, { "ph": "f", "id": 12178, "pid": 1728719, "tid": 1728719, "ts": 6950381119058.645, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119058.915, "dur": 0.020 , "args": { "External id": 3379,"cbid": 1, "correlation": 12180 } }, { "ph": "f", "id": 12180, "pid": 1728719, "tid": 1728719, "ts": 6950381119058.915, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119059.036, "dur": 0.090 , "args": { "External id": 3379,"cbid": 468, "correlation": 12181 } }, { "ph": "f", "id": 12181, "pid": 1728719, "tid": 1728719, "ts": 6950381119059.036, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119059.296, "dur": 0.020 , "args": { "External id": 3379,"cbid": 1, "correlation": 12183 } }, { "ph": "f", "id": 12183, "pid": 1728719, "tid": 1728719, "ts": 6950381119059.296, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119059.456, "dur": 0.571 , "args": { "External id": 3379,"cbid": 273, "correlation": 12184 } }, { "ph": "f", "id": 12184, "pid": 1728719, "tid": 1728719, "ts": 6950381119059.456, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381207124.174, "dur": 217.694 , "args": { "External id": 3379,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12187, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12187, "pid": 0, "tid": 7, "ts": 6950381207124.174, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119060.518, "dur": 1.834 , "args": { "External id": 3379,"cbid": 211, "correlation": 12187 } }, { "ph": "s", "id": 12187, "pid": 1728719, "tid": 1728719, "ts": 6950381119060.518, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381207342.156, "dur": 8.736 , "args": { "External id": 3391,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12200, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12200, "pid": 0, "tid": 7, "ts": 6950381207342.156, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119070.918, "dur": 1.613 , "args": { "External id": 3391,"cbid": 307, "correlation": 12200 } }, { "ph": "s", "id": 12200, "pid": 1728719, "tid": 1728719, "ts": 6950381119070.918, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381207351.180, "dur": 37.312 , "args": { "External id": 3396,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12223, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12223, "pid": 0, "tid": 7, "ts": 6950381207351.180, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119089.592, "dur": 1.714 , "args": { "External id": 3396,"cbid": 307, "correlation": 12223 } }, { "ph": "s", "id": 12223, "pid": 1728719, "tid": 1728719, "ts": 6950381119089.592, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119099.862, "dur": 0.100 , "args": { "External id": 3397,"cbid": 200, "correlation": 12238 } }, { "ph": "f", "id": 12238, "pid": 1728719, "tid": 1728719, "ts": 6950381119099.862, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119100.052, "dur": 0.140 , "args": { "External id": 3397,"cbid": 468, "correlation": 12239 } }, { "ph": "f", "id": 12239, "pid": 1728719, "tid": 1728719, "ts": 6950381119100.052, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119100.423, "dur": 0.020 , "args": { "External id": 3397,"cbid": 1, "correlation": 12241 } }, { "ph": "f", "id": 12241, "pid": 1728719, "tid": 1728719, "ts": 6950381119100.423, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119100.533, "dur": 0.100 , "args": { "External id": 3397,"cbid": 468, "correlation": 12242 } }, { "ph": "f", "id": 12242, "pid": 1728719, "tid": 1728719, "ts": 6950381119100.533, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119100.773, "dur": 0.020 , "args": { "External id": 3397,"cbid": 1, "correlation": 12244 } }, { "ph": "f", "id": 12244, "pid": 1728719, "tid": 1728719, "ts": 6950381119100.773, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119100.893, "dur": 0.091 , "args": { "External id": 3397,"cbid": 468, "correlation": 12245 } }, { "ph": "f", "id": 12245, "pid": 1728719, "tid": 1728719, "ts": 6950381119100.893, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119101.134, "dur": 0.020 , "args": { "External id": 3397,"cbid": 1, "correlation": 12247 } }, { "ph": "f", "id": 12247, "pid": 1728719, "tid": 1728719, "ts": 6950381119101.134, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119101.254, "dur": 0.391 , "args": { "External id": 3397,"cbid": 273, "correlation": 12248 } }, { "ph": "f", "id": 12248, "pid": 1728719, "tid": 1728719, "ts": 6950381119101.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381207388.779, "dur": 229.374 , "args": { "External id": 3397,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12251, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12251, "pid": 0, "tid": 7, "ts": 6950381207388.779, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119102.056, "dur": 1.673 , "args": { "External id": 3397,"cbid": 211, "correlation": 12251 } }, { "ph": "s", "id": 12251, "pid": 1728719, "tid": 1728719, "ts": 6950381119102.056, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381207618.473, "dur": 8.832 , "args": { "External id": 3409,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12260, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12260, "pid": 0, "tid": 7, "ts": 6950381207618.473, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119109.600, "dur": 1.803 , "args": { "External id": 3409,"cbid": 307, "correlation": 12260 } }, { "ph": "s", "id": 12260, "pid": 1728719, "tid": 1728719, "ts": 6950381119109.600, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381207627.593, "dur": 25.248 , "args": { "External id": 3414,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12283, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12283, "pid": 0, "tid": 7, "ts": 6950381207627.593, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119125.820, "dur": 1.553 , "args": { "External id": 3414,"cbid": 307, "correlation": 12283 } }, { "ph": "s", "id": 12283, "pid": 1728719, "tid": 1728719, "ts": 6950381119125.820, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119135.758, "dur": 0.101 , "args": { "External id": 3415,"cbid": 200, "correlation": 12298 } }, { "ph": "f", "id": 12298, "pid": 1728719, "tid": 1728719, "ts": 6950381119135.758, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119135.939, "dur": 0.130 , "args": { "External id": 3415,"cbid": 468, "correlation": 12299 } }, { "ph": "f", "id": 12299, "pid": 1728719, "tid": 1728719, "ts": 6950381119135.939, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119136.279, "dur": 0.020 , "args": { "External id": 3415,"cbid": 1, "correlation": 12301 } }, { "ph": "f", "id": 12301, "pid": 1728719, "tid": 1728719, "ts": 6950381119136.279, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119136.369, "dur": 0.101 , "args": { "External id": 3415,"cbid": 468, "correlation": 12302 } }, { "ph": "f", "id": 12302, "pid": 1728719, "tid": 1728719, "ts": 6950381119136.369, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119136.610, "dur": 0.030 , "args": { "External id": 3415,"cbid": 1, "correlation": 12304 } }, { "ph": "f", "id": 12304, "pid": 1728719, "tid": 1728719, "ts": 6950381119136.610, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119136.720, "dur": 0.100 , "args": { "External id": 3415,"cbid": 468, "correlation": 12305 } }, { "ph": "f", "id": 12305, "pid": 1728719, "tid": 1728719, "ts": 6950381119136.720, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119136.981, "dur": 0.020 , "args": { "External id": 3415,"cbid": 1, "correlation": 12307 } }, { "ph": "f", "id": 12307, "pid": 1728719, "tid": 1728719, "ts": 6950381119136.981, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119137.101, "dur": 0.461 , "args": { "External id": 3415,"cbid": 273, "correlation": 12308 } }, { "ph": "f", "id": 12308, "pid": 1728719, "tid": 1728719, "ts": 6950381119137.101, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381207653.129, "dur": 232.253 , "args": { "External id": 3415,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12311, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12311, "pid": 0, "tid": 7, "ts": 6950381207653.129, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119137.912, "dur": 1.754 , "args": { "External id": 3415,"cbid": 211, "correlation": 12311 } }, { "ph": "s", "id": 12311, "pid": 1728719, "tid": 1728719, "ts": 6950381119137.912, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381207885.670, "dur": 57.408 , "args": { "External id": 3427,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12325, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12325, "pid": 0, "tid": 7, "ts": 6950381207885.670, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119147.190, "dur": 1.643 , "args": { "External id": 3427,"cbid": 307, "correlation": 12325 } }, { "ph": "s", "id": 12325, "pid": 1728719, "tid": 1728719, "ts": 6950381119147.190, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381207943.334, "dur": 66.975 , "args": { "External id": 3428,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12335, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12335, "pid": 0, "tid": 7, "ts": 6950381207943.334, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119156.988, "dur": 1.833 , "args": { "External id": 3428,"cbid": 307, "correlation": 12335 } }, { "ph": "s", "id": 12335, "pid": 1728719, "tid": 1728719, "ts": 6950381119156.988, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381208010.565, "dur": 39.743 , "args": { "External id": 3429,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12344, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12344, "pid": 0, "tid": 7, "ts": 6950381208010.565, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119166.245, "dur": 1.593 , "args": { "External id": 3429,"cbid": 307, "correlation": 12344 } }, { "ph": "s", "id": 12344, "pid": 1728719, "tid": 1728719, "ts": 6950381119166.245, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119181.143, "dur": 0.110 , "args": { "External id": 3437,"cbid": 200, "correlation": 12379 } }, { "ph": "f", "id": 12379, "pid": 1728719, "tid": 1728719, "ts": 6950381119181.143, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119181.323, "dur": 0.080 , "args": { "External id": 3437,"cbid": 200, "correlation": 12380 } }, { "ph": "f", "id": 12380, "pid": 1728719, "tid": 1728719, "ts": 6950381119181.323, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119181.493, "dur": 0.471 , "args": { "External id": 3437,"cbid": 273, "correlation": 12381 } }, { "ph": "f", "id": 12381, "pid": 1728719, "tid": 1728719, "ts": 6950381119181.493, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381208050.564, "dur": 835.159 , "args": { "External id": 3437,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12383, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12383, "pid": 0, "tid": 7, "ts": 6950381208050.564, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119182.285, "dur": 1.974 , "args": { "External id": 3437,"cbid": 211, "correlation": 12383 } }, { "ph": "s", "id": 12383, "pid": 1728719, "tid": 1728719, "ts": 6950381119182.285, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381208886.011, "dur": 9.408 , "args": { "External id": 3446,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12403, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12403, "pid": 0, "tid": 7, "ts": 6950381208886.011, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119198.705, "dur": 1.694 , "args": { "External id": 3446,"cbid": 652, "correlation": 12403 } }, { "ph": "s", "id": 12403, "pid": 1728719, "tid": 1728719, "ts": 6950381119198.705, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119209.496, "dur": 0.110 , "args": { "External id": 3447,"cbid": 200, "correlation": 12418 } }, { "ph": "f", "id": 12418, "pid": 1728719, "tid": 1728719, "ts": 6950381119209.496, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119209.716, "dur": 0.140 , "args": { "External id": 3447,"cbid": 468, "correlation": 12419 } }, { "ph": "f", "id": 12419, "pid": 1728719, "tid": 1728719, "ts": 6950381119209.716, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119210.117, "dur": 0.040 , "args": { "External id": 3447,"cbid": 1, "correlation": 12421 } }, { "ph": "f", "id": 12421, "pid": 1728719, "tid": 1728719, "ts": 6950381119210.117, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119210.227, "dur": 0.100 , "args": { "External id": 3447,"cbid": 468, "correlation": 12422 } }, { "ph": "f", "id": 12422, "pid": 1728719, "tid": 1728719, "ts": 6950381119210.227, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119210.477, "dur": 0.020 , "args": { "External id": 3447,"cbid": 1, "correlation": 12424 } }, { "ph": "f", "id": 12424, "pid": 1728719, "tid": 1728719, "ts": 6950381119210.477, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119210.588, "dur": 0.100 , "args": { "External id": 3447,"cbid": 468, "correlation": 12425 } }, { "ph": "f", "id": 12425, "pid": 1728719, "tid": 1728719, "ts": 6950381119210.588, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119210.848, "dur": 0.020 , "args": { "External id": 3447,"cbid": 1, "correlation": 12427 } }, { "ph": "f", "id": 12427, "pid": 1728719, "tid": 1728719, "ts": 6950381119210.848, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119210.968, "dur": 0.381 , "args": { "External id": 3447,"cbid": 273, "correlation": 12428 } }, { "ph": "f", "id": 12428, "pid": 1728719, "tid": 1728719, "ts": 6950381119210.968, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381208895.707, "dur": 207.070 , "args": { "External id": 3447,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12431, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12431, "pid": 0, "tid": 7, "ts": 6950381208895.707, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119211.730, "dur": 1.753 , "args": { "External id": 3447,"cbid": 211, "correlation": 12431 } }, { "ph": "s", "id": 12431, "pid": 1728719, "tid": 1728719, "ts": 6950381119211.730, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381209103.065, "dur": 45.887 , "args": { "External id": 3459,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12443, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12443, "pid": 0, "tid": 7, "ts": 6950381209103.065, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119221.388, "dur": 1.683 , "args": { "External id": 3459,"cbid": 307, "correlation": 12443 } }, { "ph": "s", "id": 12443, "pid": 1728719, "tid": 1728719, "ts": 6950381119221.388, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381209149.241, "dur": 13.471 , "args": { "External id": 3460,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12449, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12449, "pid": 0, "tid": 7, "ts": 6950381209149.241, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119229.653, "dur": 1.583 , "args": { "External id": 3460,"cbid": 652, "correlation": 12449 } }, { "ph": "s", "id": 12449, "pid": 1728719, "tid": 1728719, "ts": 6950381119229.653, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381209163.032, "dur": 12.832 , "args": { "External id": 3461,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12455, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12455, "pid": 0, "tid": 7, "ts": 6950381209163.032, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119238.109, "dur": 1.603 , "args": { "External id": 3461,"cbid": 652, "correlation": 12455 } }, { "ph": "s", "id": 12455, "pid": 1728719, "tid": 1728719, "ts": 6950381119238.109, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119249.680, "dur": 0.101 , "args": { "External id": 3462,"cbid": 200, "correlation": 12474 } }, { "ph": "f", "id": 12474, "pid": 1728719, "tid": 1728719, "ts": 6950381119249.680, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119249.891, "dur": 0.130 , "args": { "External id": 3462,"cbid": 468, "correlation": 12475 } }, { "ph": "f", "id": 12475, "pid": 1728719, "tid": 1728719, "ts": 6950381119249.891, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119250.282, "dur": 0.020 , "args": { "External id": 3462,"cbid": 1, "correlation": 12477 } }, { "ph": "f", "id": 12477, "pid": 1728719, "tid": 1728719, "ts": 6950381119250.282, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119250.392, "dur": 0.100 , "args": { "External id": 3462,"cbid": 468, "correlation": 12478 } }, { "ph": "f", "id": 12478, "pid": 1728719, "tid": 1728719, "ts": 6950381119250.392, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119250.642, "dur": 0.020 , "args": { "External id": 3462,"cbid": 1, "correlation": 12480 } }, { "ph": "f", "id": 12480, "pid": 1728719, "tid": 1728719, "ts": 6950381119250.642, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119250.752, "dur": 0.101 , "args": { "External id": 3462,"cbid": 468, "correlation": 12481 } }, { "ph": "f", "id": 12481, "pid": 1728719, "tid": 1728719, "ts": 6950381119250.752, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119251.013, "dur": 0.020 , "args": { "External id": 3462,"cbid": 1, "correlation": 12483 } }, { "ph": "f", "id": 12483, "pid": 1728719, "tid": 1728719, "ts": 6950381119251.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119251.123, "dur": 0.371 , "args": { "External id": 3462,"cbid": 273, "correlation": 12484 } }, { "ph": "f", "id": 12484, "pid": 1728719, "tid": 1728719, "ts": 6950381119251.123, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381209176.184, "dur": 661.273 , "args": { "External id": 3462,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12487, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12487, "pid": 0, "tid": 7, "ts": 6950381209176.184, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119251.864, "dur": 1.714 , "args": { "External id": 3462,"cbid": 211, "correlation": 12487 } }, { "ph": "s", "id": 12487, "pid": 1728719, "tid": 1728719, "ts": 6950381119251.864, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119263.586, "dur": 0.090 , "args": { "External id": 3474,"cbid": 200, "correlation": 12511 } }, { "ph": "f", "id": 12511, "pid": 1728719, "tid": 1728719, "ts": 6950381119263.586, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119263.767, "dur": 0.130 , "args": { "External id": 3474,"cbid": 468, "correlation": 12512 } }, { "ph": "f", "id": 12512, "pid": 1728719, "tid": 1728719, "ts": 6950381119263.767, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119264.077, "dur": 0.020 , "args": { "External id": 3474,"cbid": 1, "correlation": 12514 } }, { "ph": "f", "id": 12514, "pid": 1728719, "tid": 1728719, "ts": 6950381119264.077, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119264.167, "dur": 0.091 , "args": { "External id": 3474,"cbid": 468, "correlation": 12515 } }, { "ph": "f", "id": 12515, "pid": 1728719, "tid": 1728719, "ts": 6950381119264.167, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119264.408, "dur": 0.020 , "args": { "External id": 3474,"cbid": 1, "correlation": 12517 } }, { "ph": "f", "id": 12517, "pid": 1728719, "tid": 1728719, "ts": 6950381119264.408, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119264.528, "dur": 0.100 , "args": { "External id": 3474,"cbid": 468, "correlation": 12518 } }, { "ph": "f", "id": 12518, "pid": 1728719, "tid": 1728719, "ts": 6950381119264.528, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119264.778, "dur": 0.031 , "args": { "External id": 3474,"cbid": 1, "correlation": 12520 } }, { "ph": "f", "id": 12520, "pid": 1728719, "tid": 1728719, "ts": 6950381119264.778, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119264.889, "dur": 0.320 , "args": { "External id": 3474,"cbid": 273, "correlation": 12521 } }, { "ph": "f", "id": 12521, "pid": 1728719, "tid": 1728719, "ts": 6950381119264.889, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381209837.745, "dur": 682.041 , "args": { "External id": 3474,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12524, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12524, "pid": 0, "tid": 7, "ts": 6950381209837.745, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119265.550, "dur": 1.683 , "args": { "External id": 3474,"cbid": 211, "correlation": 12524 } }, { "ph": "s", "id": 12524, "pid": 1728719, "tid": 1728719, "ts": 6950381119265.550, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381210520.074, "dur": 154.622 , "args": { "External id": 3486,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12537, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12537, "pid": 0, "tid": 7, "ts": 6950381210520.074, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119274.597, "dur": 1.793 , "args": { "External id": 3486,"cbid": 307, "correlation": 12537 } }, { "ph": "s", "id": 12537, "pid": 1728719, "tid": 1728719, "ts": 6950381119274.597, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381210674.984, "dur": 59.583 , "args": { "External id": 3487,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12543, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12543, "pid": 0, "tid": 7, "ts": 6950381210674.984, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119283.133, "dur": 1.522 , "args": { "External id": 3487,"cbid": 652, "correlation": 12543 } }, { "ph": "s", "id": 12543, "pid": 1728719, "tid": 1728719, "ts": 6950381119283.133, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119294.273, "dur": 0.091 , "args": { "External id": 3488,"cbid": 200, "correlation": 12562 } }, { "ph": "f", "id": 12562, "pid": 1728719, "tid": 1728719, "ts": 6950381119294.273, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119294.454, "dur": 0.110 , "args": { "External id": 3488,"cbid": 468, "correlation": 12563 } }, { "ph": "f", "id": 12563, "pid": 1728719, "tid": 1728719, "ts": 6950381119294.454, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119294.774, "dur": 0.030 , "args": { "External id": 3488,"cbid": 1, "correlation": 12565 } }, { "ph": "f", "id": 12565, "pid": 1728719, "tid": 1728719, "ts": 6950381119294.774, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119294.864, "dur": 0.101 , "args": { "External id": 3488,"cbid": 468, "correlation": 12566 } }, { "ph": "f", "id": 12566, "pid": 1728719, "tid": 1728719, "ts": 6950381119294.864, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119295.105, "dur": 0.020 , "args": { "External id": 3488,"cbid": 1, "correlation": 12568 } }, { "ph": "f", "id": 12568, "pid": 1728719, "tid": 1728719, "ts": 6950381119295.105, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119295.235, "dur": 0.090 , "args": { "External id": 3488,"cbid": 468, "correlation": 12569 } }, { "ph": "f", "id": 12569, "pid": 1728719, "tid": 1728719, "ts": 6950381119295.235, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119295.476, "dur": 0.020 , "args": { "External id": 3488,"cbid": 1, "correlation": 12571 } }, { "ph": "f", "id": 12571, "pid": 1728719, "tid": 1728719, "ts": 6950381119295.476, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119295.586, "dur": 0.391 , "args": { "External id": 3488,"cbid": 273, "correlation": 12572 } }, { "ph": "f", "id": 12572, "pid": 1728719, "tid": 1728719, "ts": 6950381119295.586, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381210734.824, "dur": 679.000 , "args": { "External id": 3488,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12575, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12575, "pid": 0, "tid": 7, "ts": 6950381210734.824, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119296.347, "dur": 1.673 , "args": { "External id": 3488,"cbid": 211, "correlation": 12575 } }, { "ph": "s", "id": 12575, "pid": 1728719, "tid": 1728719, "ts": 6950381119296.347, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381211414.112, "dur": 40.288 , "args": { "External id": 3500,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12586, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12586, "pid": 0, "tid": 7, "ts": 6950381211414.112, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119306.115, "dur": 1.664 , "args": { "External id": 3500,"cbid": 307, "correlation": 12586 } }, { "ph": "s", "id": 12586, "pid": 1728719, "tid": 1728719, "ts": 6950381119306.115, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381211454.656, "dur": 21.216 , "args": { "External id": 3512,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12598, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12598, "pid": 0, "tid": 7, "ts": 6950381211454.656, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119384.481, "dur": 2.455 , "args": { "External id": 3512,"cbid": 307, "correlation": 12598 } }, { "ph": "s", "id": 12598, "pid": 1728719, "tid": 1728719, "ts": 6950381119384.481, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381211476.191, "dur": 12.608 , "args": { "External id": 3513,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12612, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12612, "pid": 0, "tid": 7, "ts": 6950381211476.191, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119398.027, "dur": 1.633 , "args": { "External id": 3513,"cbid": 652, "correlation": 12612 } }, { "ph": "s", "id": 12612, "pid": 1728719, "tid": 1728719, "ts": 6950381119398.027, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381211489.087, "dur": 12.064 , "args": { "External id": 3514,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12626, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12626, "pid": 0, "tid": 7, "ts": 6950381211489.087, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119409.358, "dur": 1.522 , "args": { "External id": 3514,"cbid": 652, "correlation": 12626 } }, { "ph": "s", "id": 12626, "pid": 1728719, "tid": 1728719, "ts": 6950381119409.358, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381211501.439, "dur": 10.784 , "args": { "External id": 3515,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12640, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12640, "pid": 0, "tid": 7, "ts": 6950381211501.439, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119420.238, "dur": 1.463 , "args": { "External id": 3515,"cbid": 652, "correlation": 12640 } }, { "ph": "s", "id": 12640, "pid": 1728719, "tid": 1728719, "ts": 6950381119420.238, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119431.298, "dur": 0.191 , "args": { "External id": 3516,"cbid": 200, "correlation": 12655 } }, { "ph": "f", "id": 12655, "pid": 1728719, "tid": 1728719, "ts": 6950381119431.298, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119431.639, "dur": 0.180 , "args": { "External id": 3516,"cbid": 468, "correlation": 12656 } }, { "ph": "f", "id": 12656, "pid": 1728719, "tid": 1728719, "ts": 6950381119431.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119432.150, "dur": 0.030 , "args": { "External id": 3516,"cbid": 1, "correlation": 12658 } }, { "ph": "f", "id": 12658, "pid": 1728719, "tid": 1728719, "ts": 6950381119432.150, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119432.250, "dur": 0.110 , "args": { "External id": 3516,"cbid": 468, "correlation": 12659 } }, { "ph": "f", "id": 12659, "pid": 1728719, "tid": 1728719, "ts": 6950381119432.250, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119432.531, "dur": 0.020 , "args": { "External id": 3516,"cbid": 1, "correlation": 12661 } }, { "ph": "f", "id": 12661, "pid": 1728719, "tid": 1728719, "ts": 6950381119432.531, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119432.651, "dur": 0.100 , "args": { "External id": 3516,"cbid": 468, "correlation": 12662 } }, { "ph": "f", "id": 12662, "pid": 1728719, "tid": 1728719, "ts": 6950381119432.651, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119432.911, "dur": 0.020 , "args": { "External id": 3516,"cbid": 1, "correlation": 12664 } }, { "ph": "f", "id": 12664, "pid": 1728719, "tid": 1728719, "ts": 6950381119432.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119433.072, "dur": 0.661 , "args": { "External id": 3516,"cbid": 273, "correlation": 12665 } }, { "ph": "f", "id": 12665, "pid": 1728719, "tid": 1728719, "ts": 6950381119433.072, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381211512.511, "dur": 217.533 , "args": { "External id": 3516,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12668, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12668, "pid": 0, "tid": 7, "ts": 6950381211512.511, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119434.214, "dur": 2.134 , "args": { "External id": 3516,"cbid": 211, "correlation": 12668 } }, { "ph": "s", "id": 12668, "pid": 1728719, "tid": 1728719, "ts": 6950381119434.214, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381211730.365, "dur": 8.767 , "args": { "External id": 3528,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12681, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12681, "pid": 0, "tid": 7, "ts": 6950381211730.365, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119445.074, "dur": 1.733 , "args": { "External id": 3528,"cbid": 307, "correlation": 12681 } }, { "ph": "s", "id": 12681, "pid": 1728719, "tid": 1728719, "ts": 6950381119445.074, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381211739.420, "dur": 37.504 , "args": { "External id": 3533,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12704, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12704, "pid": 0, "tid": 7, "ts": 6950381211739.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119463.819, "dur": 1.683 , "args": { "External id": 3533,"cbid": 307, "correlation": 12704 } }, { "ph": "s", "id": 12704, "pid": 1728719, "tid": 1728719, "ts": 6950381119463.819, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119474.028, "dur": 0.110 , "args": { "External id": 3534,"cbid": 200, "correlation": 12719 } }, { "ph": "f", "id": 12719, "pid": 1728719, "tid": 1728719, "ts": 6950381119474.028, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119474.238, "dur": 0.141 , "args": { "External id": 3534,"cbid": 468, "correlation": 12720 } }, { "ph": "f", "id": 12720, "pid": 1728719, "tid": 1728719, "ts": 6950381119474.238, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119474.589, "dur": 0.030 , "args": { "External id": 3534,"cbid": 1, "correlation": 12722 } }, { "ph": "f", "id": 12722, "pid": 1728719, "tid": 1728719, "ts": 6950381119474.589, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119474.689, "dur": 0.100 , "args": { "External id": 3534,"cbid": 468, "correlation": 12723 } }, { "ph": "f", "id": 12723, "pid": 1728719, "tid": 1728719, "ts": 6950381119474.689, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119474.950, "dur": 0.020 , "args": { "External id": 3534,"cbid": 1, "correlation": 12725 } }, { "ph": "f", "id": 12725, "pid": 1728719, "tid": 1728719, "ts": 6950381119474.950, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119475.070, "dur": 0.090 , "args": { "External id": 3534,"cbid": 468, "correlation": 12726 } }, { "ph": "f", "id": 12726, "pid": 1728719, "tid": 1728719, "ts": 6950381119475.070, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119475.310, "dur": 0.020 , "args": { "External id": 3534,"cbid": 1, "correlation": 12728 } }, { "ph": "f", "id": 12728, "pid": 1728719, "tid": 1728719, "ts": 6950381119475.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119475.431, "dur": 0.420 , "args": { "External id": 3534,"cbid": 273, "correlation": 12729 } }, { "ph": "f", "id": 12729, "pid": 1728719, "tid": 1728719, "ts": 6950381119475.431, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381211777.212, "dur": 228.573 , "args": { "External id": 3534,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12732, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12732, "pid": 0, "tid": 7, "ts": 6950381211777.212, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119476.232, "dur": 1.763 , "args": { "External id": 3534,"cbid": 211, "correlation": 12732 } }, { "ph": "s", "id": 12732, "pid": 1728719, "tid": 1728719, "ts": 6950381119476.232, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381212006.106, "dur": 8.895 , "args": { "External id": 3546,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12741, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12741, "pid": 0, "tid": 7, "ts": 6950381212006.106, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119483.816, "dur": 1.623 , "args": { "External id": 3546,"cbid": 307, "correlation": 12741 } }, { "ph": "s", "id": 12741, "pid": 1728719, "tid": 1728719, "ts": 6950381119483.816, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381212015.289, "dur": 30.176 , "args": { "External id": 3551,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12764, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12764, "pid": 0, "tid": 7, "ts": 6950381212015.289, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119500.658, "dur": 1.572 , "args": { "External id": 3551,"cbid": 307, "correlation": 12764 } }, { "ph": "s", "id": 12764, "pid": 1728719, "tid": 1728719, "ts": 6950381119500.658, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119510.676, "dur": 0.100 , "args": { "External id": 3552,"cbid": 200, "correlation": 12779 } }, { "ph": "f", "id": 12779, "pid": 1728719, "tid": 1728719, "ts": 6950381119510.676, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119510.857, "dur": 0.130 , "args": { "External id": 3552,"cbid": 468, "correlation": 12780 } }, { "ph": "f", "id": 12780, "pid": 1728719, "tid": 1728719, "ts": 6950381119510.857, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119511.187, "dur": 0.030 , "args": { "External id": 3552,"cbid": 1, "correlation": 12782 } }, { "ph": "f", "id": 12782, "pid": 1728719, "tid": 1728719, "ts": 6950381119511.187, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119511.287, "dur": 0.101 , "args": { "External id": 3552,"cbid": 468, "correlation": 12783 } }, { "ph": "f", "id": 12783, "pid": 1728719, "tid": 1728719, "ts": 6950381119511.287, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119511.548, "dur": 0.020 , "args": { "External id": 3552,"cbid": 1, "correlation": 12785 } }, { "ph": "f", "id": 12785, "pid": 1728719, "tid": 1728719, "ts": 6950381119511.548, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119511.668, "dur": 0.090 , "args": { "External id": 3552,"cbid": 468, "correlation": 12786 } }, { "ph": "f", "id": 12786, "pid": 1728719, "tid": 1728719, "ts": 6950381119511.668, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119511.929, "dur": 0.020 , "args": { "External id": 3552,"cbid": 1, "correlation": 12788 } }, { "ph": "f", "id": 12788, "pid": 1728719, "tid": 1728719, "ts": 6950381119511.929, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119512.029, "dur": 0.431 , "args": { "External id": 3552,"cbid": 273, "correlation": 12789 } }, { "ph": "f", "id": 12789, "pid": 1728719, "tid": 1728719, "ts": 6950381119512.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381212045.753, "dur": 228.925 , "args": { "External id": 3552,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12792, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12792, "pid": 0, "tid": 7, "ts": 6950381212045.753, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119512.820, "dur": 1.713 , "args": { "External id": 3552,"cbid": 211, "correlation": 12792 } }, { "ph": "s", "id": 12792, "pid": 1728719, "tid": 1728719, "ts": 6950381119512.820, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381212274.967, "dur": 59.775 , "args": { "External id": 3564,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12806, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12806, "pid": 0, "tid": 7, "ts": 6950381212274.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119522.138, "dur": 1.693 , "args": { "External id": 3564,"cbid": 307, "correlation": 12806 } }, { "ph": "s", "id": 12806, "pid": 1728719, "tid": 1728719, "ts": 6950381119522.138, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381212334.998, "dur": 65.343 , "args": { "External id": 3565,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12816, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12816, "pid": 0, "tid": 7, "ts": 6950381212334.998, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119532.747, "dur": 1.744 , "args": { "External id": 3565,"cbid": 307, "correlation": 12816 } }, { "ph": "s", "id": 12816, "pid": 1728719, "tid": 1728719, "ts": 6950381119532.747, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381212400.629, "dur": 38.112 , "args": { "External id": 3566,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12825, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12825, "pid": 0, "tid": 7, "ts": 6950381212400.629, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119546.092, "dur": 1.653 , "args": { "External id": 3566,"cbid": 307, "correlation": 12825 } }, { "ph": "s", "id": 12825, "pid": 1728719, "tid": 1728719, "ts": 6950381119546.092, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119561.330, "dur": 0.121 , "args": { "External id": 3574,"cbid": 200, "correlation": 12860 } }, { "ph": "f", "id": 12860, "pid": 1728719, "tid": 1728719, "ts": 6950381119561.330, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119561.531, "dur": 0.080 , "args": { "External id": 3574,"cbid": 200, "correlation": 12861 } }, { "ph": "f", "id": 12861, "pid": 1728719, "tid": 1728719, "ts": 6950381119561.531, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119561.721, "dur": 0.471 , "args": { "External id": 3574,"cbid": 273, "correlation": 12862 } }, { "ph": "f", "id": 12862, "pid": 1728719, "tid": 1728719, "ts": 6950381119561.721, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381212439.029, "dur": 835.383 , "args": { "External id": 3574,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12864, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12864, "pid": 0, "tid": 7, "ts": 6950381212439.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119562.513, "dur": 1.953 , "args": { "External id": 3574,"cbid": 211, "correlation": 12864 } }, { "ph": "s", "id": 12864, "pid": 1728719, "tid": 1728719, "ts": 6950381119562.513, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381213274.700, "dur": 9.344 , "args": { "External id": 3583,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12884, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12884, "pid": 0, "tid": 7, "ts": 6950381213274.700, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119579.865, "dur": 1.733 , "args": { "External id": 3583,"cbid": 652, "correlation": 12884 } }, { "ph": "s", "id": 12884, "pid": 1728719, "tid": 1728719, "ts": 6950381119579.865, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119594.602, "dur": 0.101 , "args": { "External id": 3584,"cbid": 200, "correlation": 12899 } }, { "ph": "f", "id": 12899, "pid": 1728719, "tid": 1728719, "ts": 6950381119594.602, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119594.823, "dur": 0.150 , "args": { "External id": 3584,"cbid": 468, "correlation": 12900 } }, { "ph": "f", "id": 12900, "pid": 1728719, "tid": 1728719, "ts": 6950381119594.823, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119595.244, "dur": 0.030 , "args": { "External id": 3584,"cbid": 1, "correlation": 12902 } }, { "ph": "f", "id": 12902, "pid": 1728719, "tid": 1728719, "ts": 6950381119595.244, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119595.344, "dur": 0.100 , "args": { "External id": 3584,"cbid": 468, "correlation": 12903 } }, { "ph": "f", "id": 12903, "pid": 1728719, "tid": 1728719, "ts": 6950381119595.344, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119595.594, "dur": 0.020 , "args": { "External id": 3584,"cbid": 1, "correlation": 12905 } }, { "ph": "f", "id": 12905, "pid": 1728719, "tid": 1728719, "ts": 6950381119595.594, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119595.715, "dur": 0.090 , "args": { "External id": 3584,"cbid": 468, "correlation": 12906 } }, { "ph": "f", "id": 12906, "pid": 1728719, "tid": 1728719, "ts": 6950381119595.715, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119595.985, "dur": 0.020 , "args": { "External id": 3584,"cbid": 1, "correlation": 12908 } }, { "ph": "f", "id": 12908, "pid": 1728719, "tid": 1728719, "ts": 6950381119595.985, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119596.105, "dur": 0.471 , "args": { "External id": 3584,"cbid": 273, "correlation": 12909 } }, { "ph": "f", "id": 12909, "pid": 1728719, "tid": 1728719, "ts": 6950381119596.105, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381213284.300, "dur": 206.525 , "args": { "External id": 3584,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12912, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12912, "pid": 0, "tid": 7, "ts": 6950381213284.300, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119596.977, "dur": 1.863 , "args": { "External id": 3584,"cbid": 211, "correlation": 12912 } }, { "ph": "s", "id": 12912, "pid": 1728719, "tid": 1728719, "ts": 6950381119596.977, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381213491.113, "dur": 46.560 , "args": { "External id": 3596,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12924, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12924, "pid": 0, "tid": 7, "ts": 6950381213491.113, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119606.965, "dur": 1.684 , "args": { "External id": 3596,"cbid": 307, "correlation": 12924 } }, { "ph": "s", "id": 12924, "pid": 1728719, "tid": 1728719, "ts": 6950381119606.965, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381213537.993, "dur": 12.896 , "args": { "External id": 3597,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12930, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12930, "pid": 0, "tid": 7, "ts": 6950381213537.993, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119615.501, "dur": 1.463 , "args": { "External id": 3597,"cbid": 652, "correlation": 12930 } }, { "ph": "s", "id": 12930, "pid": 1728719, "tid": 1728719, "ts": 6950381119615.501, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381213551.177, "dur": 12.448 , "args": { "External id": 3598,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12936, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12936, "pid": 0, "tid": 7, "ts": 6950381213551.177, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119623.877, "dur": 1.493 , "args": { "External id": 3598,"cbid": 652, "correlation": 12936 } }, { "ph": "s", "id": 12936, "pid": 1728719, "tid": 1728719, "ts": 6950381119623.877, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119635.328, "dur": 0.171 , "args": { "External id": 3599,"cbid": 200, "correlation": 12955 } }, { "ph": "f", "id": 12955, "pid": 1728719, "tid": 1728719, "ts": 6950381119635.328, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119635.609, "dur": 0.120 , "args": { "External id": 3599,"cbid": 468, "correlation": 12956 } }, { "ph": "f", "id": 12956, "pid": 1728719, "tid": 1728719, "ts": 6950381119635.609, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119635.959, "dur": 0.020 , "args": { "External id": 3599,"cbid": 1, "correlation": 12958 } }, { "ph": "f", "id": 12958, "pid": 1728719, "tid": 1728719, "ts": 6950381119635.959, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119636.050, "dur": 0.100 , "args": { "External id": 3599,"cbid": 468, "correlation": 12959 } }, { "ph": "f", "id": 12959, "pid": 1728719, "tid": 1728719, "ts": 6950381119636.050, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119636.310, "dur": 0.020 , "args": { "External id": 3599,"cbid": 1, "correlation": 12961 } }, { "ph": "f", "id": 12961, "pid": 1728719, "tid": 1728719, "ts": 6950381119636.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119636.420, "dur": 0.100 , "args": { "External id": 3599,"cbid": 468, "correlation": 12962 } }, { "ph": "f", "id": 12962, "pid": 1728719, "tid": 1728719, "ts": 6950381119636.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119636.681, "dur": 0.030 , "args": { "External id": 3599,"cbid": 1, "correlation": 12964 } }, { "ph": "f", "id": 12964, "pid": 1728719, "tid": 1728719, "ts": 6950381119636.681, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119636.791, "dur": 0.391 , "args": { "External id": 3599,"cbid": 273, "correlation": 12965 } }, { "ph": "f", "id": 12965, "pid": 1728719, "tid": 1728719, "ts": 6950381119636.791, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381213563.913, "dur": 662.713 , "args": { "External id": 3599,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 12968, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 12968, "pid": 0, "tid": 7, "ts": 6950381213563.913, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119637.562, "dur": 1.694 , "args": { "External id": 3599,"cbid": 211, "correlation": 12968 } }, { "ph": "s", "id": 12968, "pid": 1728719, "tid": 1728719, "ts": 6950381119637.562, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119649.294, "dur": 0.090 , "args": { "External id": 3611,"cbid": 200, "correlation": 12992 } }, { "ph": "f", "id": 12992, "pid": 1728719, "tid": 1728719, "ts": 6950381119649.294, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119649.485, "dur": 0.110 , "args": { "External id": 3611,"cbid": 468, "correlation": 12993 } }, { "ph": "f", "id": 12993, "pid": 1728719, "tid": 1728719, "ts": 6950381119649.485, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119649.785, "dur": 0.020 , "args": { "External id": 3611,"cbid": 1, "correlation": 12995 } }, { "ph": "f", "id": 12995, "pid": 1728719, "tid": 1728719, "ts": 6950381119649.785, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119649.865, "dur": 0.100 , "args": { "External id": 3611,"cbid": 468, "correlation": 12996 } }, { "ph": "f", "id": 12996, "pid": 1728719, "tid": 1728719, "ts": 6950381119649.865, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119650.116, "dur": 0.020 , "args": { "External id": 3611,"cbid": 1, "correlation": 12998 } }, { "ph": "f", "id": 12998, "pid": 1728719, "tid": 1728719, "ts": 6950381119650.116, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119650.226, "dur": 0.100 , "args": { "External id": 3611,"cbid": 468, "correlation": 12999 } }, { "ph": "f", "id": 12999, "pid": 1728719, "tid": 1728719, "ts": 6950381119650.226, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119650.476, "dur": 0.020 , "args": { "External id": 3611,"cbid": 1, "correlation": 13001 } }, { "ph": "f", "id": 13001, "pid": 1728719, "tid": 1728719, "ts": 6950381119650.476, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119650.577, "dur": 0.350 , "args": { "External id": 3611,"cbid": 273, "correlation": 13002 } }, { "ph": "f", "id": 13002, "pid": 1728719, "tid": 1728719, "ts": 6950381119650.577, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381214226.914, "dur": 675.800 , "args": { "External id": 3611,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13005, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13005, "pid": 0, "tid": 7, "ts": 6950381214226.914, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119651.268, "dur": 1.653 , "args": { "External id": 3611,"cbid": 211, "correlation": 13005 } }, { "ph": "s", "id": 13005, "pid": 1728719, "tid": 1728719, "ts": 6950381119651.268, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381214903.002, "dur": 153.534 , "args": { "External id": 3623,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13018, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13018, "pid": 0, "tid": 7, "ts": 6950381214903.002, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119660.495, "dur": 1.783 , "args": { "External id": 3623,"cbid": 307, "correlation": 13018 } }, { "ph": "s", "id": 13018, "pid": 1728719, "tid": 1728719, "ts": 6950381119660.495, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381215056.793, "dur": 59.551 , "args": { "External id": 3624,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13024, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13024, "pid": 0, "tid": 7, "ts": 6950381215056.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119669.081, "dur": 1.553 , "args": { "External id": 3624,"cbid": 652, "correlation": 13024 } }, { "ph": "s", "id": 13024, "pid": 1728719, "tid": 1728719, "ts": 6950381119669.081, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119680.362, "dur": 0.080 , "args": { "External id": 3625,"cbid": 200, "correlation": 13043 } }, { "ph": "f", "id": 13043, "pid": 1728719, "tid": 1728719, "ts": 6950381119680.362, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119680.542, "dur": 0.111 , "args": { "External id": 3625,"cbid": 468, "correlation": 13044 } }, { "ph": "f", "id": 13044, "pid": 1728719, "tid": 1728719, "ts": 6950381119680.542, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119680.843, "dur": 0.020 , "args": { "External id": 3625,"cbid": 1, "correlation": 13046 } }, { "ph": "f", "id": 13046, "pid": 1728719, "tid": 1728719, "ts": 6950381119680.843, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119680.933, "dur": 0.100 , "args": { "External id": 3625,"cbid": 468, "correlation": 13047 } }, { "ph": "f", "id": 13047, "pid": 1728719, "tid": 1728719, "ts": 6950381119680.933, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119681.184, "dur": 0.020 , "args": { "External id": 3625,"cbid": 1, "correlation": 13049 } }, { "ph": "f", "id": 13049, "pid": 1728719, "tid": 1728719, "ts": 6950381119681.184, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119681.294, "dur": 0.090 , "args": { "External id": 3625,"cbid": 468, "correlation": 13050 } }, { "ph": "f", "id": 13050, "pid": 1728719, "tid": 1728719, "ts": 6950381119681.294, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119681.534, "dur": 0.020 , "args": { "External id": 3625,"cbid": 1, "correlation": 13052 } }, { "ph": "f", "id": 13052, "pid": 1728719, "tid": 1728719, "ts": 6950381119681.534, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119681.654, "dur": 0.371 , "args": { "External id": 3625,"cbid": 273, "correlation": 13053 } }, { "ph": "f", "id": 13053, "pid": 1728719, "tid": 1728719, "ts": 6950381119681.654, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381215116.632, "dur": 671.545 , "args": { "External id": 3625,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13056, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13056, "pid": 0, "tid": 7, "ts": 6950381215116.632, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119682.366, "dur": 1.753 , "args": { "External id": 3625,"cbid": 211, "correlation": 13056 } }, { "ph": "s", "id": 13056, "pid": 1728719, "tid": 1728719, "ts": 6950381119682.366, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381215788.465, "dur": 39.839 , "args": { "External id": 3637,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13067, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13067, "pid": 0, "tid": 7, "ts": 6950381215788.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119692.164, "dur": 1.683 , "args": { "External id": 3637,"cbid": 307, "correlation": 13067 } }, { "ph": "s", "id": 13067, "pid": 1728719, "tid": 1728719, "ts": 6950381119692.164, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381215828.560, "dur": 21.184 , "args": { "External id": 3649,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13079, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13079, "pid": 0, "tid": 7, "ts": 6950381215828.560, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119771.892, "dur": 2.485 , "args": { "External id": 3649,"cbid": 307, "correlation": 13079 } }, { "ph": "s", "id": 13079, "pid": 1728719, "tid": 1728719, "ts": 6950381119771.892, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381215850.064, "dur": 12.544 , "args": { "External id": 3650,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13093, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13093, "pid": 0, "tid": 7, "ts": 6950381215850.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119786.760, "dur": 1.633 , "args": { "External id": 3650,"cbid": 652, "correlation": 13093 } }, { "ph": "s", "id": 13093, "pid": 1728719, "tid": 1728719, "ts": 6950381119786.760, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381215862.928, "dur": 12.224 , "args": { "External id": 3651,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13107, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13107, "pid": 0, "tid": 7, "ts": 6950381215862.928, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119798.462, "dur": 1.623 , "args": { "External id": 3651,"cbid": 652, "correlation": 13107 } }, { "ph": "s", "id": 13107, "pid": 1728719, "tid": 1728719, "ts": 6950381119798.462, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381215875.472, "dur": 11.007 , "args": { "External id": 3652,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13121, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13121, "pid": 0, "tid": 7, "ts": 6950381215875.472, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119809.653, "dur": 1.823 , "args": { "External id": 3652,"cbid": 652, "correlation": 13121 } }, { "ph": "s", "id": 13121, "pid": 1728719, "tid": 1728719, "ts": 6950381119809.653, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119821.395, "dur": 0.150 , "args": { "External id": 3653,"cbid": 200, "correlation": 13136 } }, { "ph": "f", "id": 13136, "pid": 1728719, "tid": 1728719, "ts": 6950381119821.395, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119821.685, "dur": 0.190 , "args": { "External id": 3653,"cbid": 468, "correlation": 13137 } }, { "ph": "f", "id": 13137, "pid": 1728719, "tid": 1728719, "ts": 6950381119821.685, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119822.216, "dur": 0.030 , "args": { "External id": 3653,"cbid": 1, "correlation": 13139 } }, { "ph": "f", "id": 13139, "pid": 1728719, "tid": 1728719, "ts": 6950381119822.216, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119822.316, "dur": 0.100 , "args": { "External id": 3653,"cbid": 468, "correlation": 13140 } }, { "ph": "f", "id": 13140, "pid": 1728719, "tid": 1728719, "ts": 6950381119822.316, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119822.597, "dur": 0.020 , "args": { "External id": 3653,"cbid": 1, "correlation": 13142 } }, { "ph": "f", "id": 13142, "pid": 1728719, "tid": 1728719, "ts": 6950381119822.597, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119822.737, "dur": 0.090 , "args": { "External id": 3653,"cbid": 468, "correlation": 13143 } }, { "ph": "f", "id": 13143, "pid": 1728719, "tid": 1728719, "ts": 6950381119822.737, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119822.998, "dur": 0.020 , "args": { "External id": 3653,"cbid": 1, "correlation": 13145 } }, { "ph": "f", "id": 13145, "pid": 1728719, "tid": 1728719, "ts": 6950381119822.998, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119823.148, "dur": 0.611 , "args": { "External id": 3653,"cbid": 273, "correlation": 13146 } }, { "ph": "f", "id": 13146, "pid": 1728719, "tid": 1728719, "ts": 6950381119823.148, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381215886.767, "dur": 216.958 , "args": { "External id": 3653,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13149, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13149, "pid": 0, "tid": 7, "ts": 6950381215886.767, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119824.230, "dur": 1.813 , "args": { "External id": 3653,"cbid": 211, "correlation": 13149 } }, { "ph": "s", "id": 13149, "pid": 1728719, "tid": 1728719, "ts": 6950381119824.230, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381216104.013, "dur": 8.832 , "args": { "External id": 3665,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13162, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13162, "pid": 0, "tid": 7, "ts": 6950381216104.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119834.619, "dur": 1.924 , "args": { "External id": 3665,"cbid": 307, "correlation": 13162 } }, { "ph": "s", "id": 13162, "pid": 1728719, "tid": 1728719, "ts": 6950381119834.619, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381216113.133, "dur": 37.312 , "args": { "External id": 3670,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13185, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13185, "pid": 0, "tid": 7, "ts": 6950381216113.133, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119853.695, "dur": 1.713 , "args": { "External id": 3670,"cbid": 307, "correlation": 13185 } }, { "ph": "s", "id": 13185, "pid": 1728719, "tid": 1728719, "ts": 6950381119853.695, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119863.894, "dur": 0.100 , "args": { "External id": 3671,"cbid": 200, "correlation": 13200 } }, { "ph": "f", "id": 13200, "pid": 1728719, "tid": 1728719, "ts": 6950381119863.894, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119864.094, "dur": 0.140 , "args": { "External id": 3671,"cbid": 468, "correlation": 13201 } }, { "ph": "f", "id": 13201, "pid": 1728719, "tid": 1728719, "ts": 6950381119864.094, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119864.465, "dur": 0.040 , "args": { "External id": 3671,"cbid": 1, "correlation": 13203 } }, { "ph": "f", "id": 13203, "pid": 1728719, "tid": 1728719, "ts": 6950381119864.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119864.575, "dur": 0.090 , "args": { "External id": 3671,"cbid": 468, "correlation": 13204 } }, { "ph": "f", "id": 13204, "pid": 1728719, "tid": 1728719, "ts": 6950381119864.575, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119864.835, "dur": 0.020 , "args": { "External id": 3671,"cbid": 1, "correlation": 13206 } }, { "ph": "f", "id": 13206, "pid": 1728719, "tid": 1728719, "ts": 6950381119864.835, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119864.946, "dur": 0.100 , "args": { "External id": 3671,"cbid": 468, "correlation": 13207 } }, { "ph": "f", "id": 13207, "pid": 1728719, "tid": 1728719, "ts": 6950381119864.946, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119865.206, "dur": 0.020 , "args": { "External id": 3671,"cbid": 1, "correlation": 13209 } }, { "ph": "f", "id": 13209, "pid": 1728719, "tid": 1728719, "ts": 6950381119865.206, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119865.326, "dur": 0.441 , "args": { "External id": 3671,"cbid": 273, "correlation": 13210 } }, { "ph": "f", "id": 13210, "pid": 1728719, "tid": 1728719, "ts": 6950381119865.326, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381216150.733, "dur": 229.821 , "args": { "External id": 3671,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13213, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13213, "pid": 0, "tid": 7, "ts": 6950381216150.733, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119866.158, "dur": 1.743 , "args": { "External id": 3671,"cbid": 211, "correlation": 13213 } }, { "ph": "s", "id": 13213, "pid": 1728719, "tid": 1728719, "ts": 6950381119866.158, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381216380.874, "dur": 8.832 , "args": { "External id": 3683,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13222, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13222, "pid": 0, "tid": 7, "ts": 6950381216380.874, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119873.782, "dur": 1.603 , "args": { "External id": 3683,"cbid": 307, "correlation": 13222 } }, { "ph": "s", "id": 13222, "pid": 1728719, "tid": 1728719, "ts": 6950381119873.782, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381216389.994, "dur": 26.752 , "args": { "External id": 3688,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13245, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13245, "pid": 0, "tid": 7, "ts": 6950381216389.994, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119889.471, "dur": 1.613 , "args": { "External id": 3688,"cbid": 307, "correlation": 13245 } }, { "ph": "s", "id": 13245, "pid": 1728719, "tid": 1728719, "ts": 6950381119889.471, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119899.490, "dur": 0.090 , "args": { "External id": 3689,"cbid": 200, "correlation": 13260 } }, { "ph": "f", "id": 13260, "pid": 1728719, "tid": 1728719, "ts": 6950381119899.490, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119899.670, "dur": 0.131 , "args": { "External id": 3689,"cbid": 468, "correlation": 13261 } }, { "ph": "f", "id": 13261, "pid": 1728719, "tid": 1728719, "ts": 6950381119899.670, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119900.001, "dur": 0.030 , "args": { "External id": 3689,"cbid": 1, "correlation": 13263 } }, { "ph": "f", "id": 13263, "pid": 1728719, "tid": 1728719, "ts": 6950381119900.001, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119900.111, "dur": 0.100 , "args": { "External id": 3689,"cbid": 468, "correlation": 13264 } }, { "ph": "f", "id": 13264, "pid": 1728719, "tid": 1728719, "ts": 6950381119900.111, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119900.352, "dur": 0.020 , "args": { "External id": 3689,"cbid": 1, "correlation": 13266 } }, { "ph": "f", "id": 13266, "pid": 1728719, "tid": 1728719, "ts": 6950381119900.352, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119900.462, "dur": 0.100 , "args": { "External id": 3689,"cbid": 468, "correlation": 13267 } }, { "ph": "f", "id": 13267, "pid": 1728719, "tid": 1728719, "ts": 6950381119900.462, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119900.722, "dur": 0.020 , "args": { "External id": 3689,"cbid": 1, "correlation": 13269 } }, { "ph": "f", "id": 13269, "pid": 1728719, "tid": 1728719, "ts": 6950381119900.722, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119900.832, "dur": 0.391 , "args": { "External id": 3689,"cbid": 273, "correlation": 13270 } }, { "ph": "f", "id": 13270, "pid": 1728719, "tid": 1728719, "ts": 6950381119900.832, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381216417.034, "dur": 232.317 , "args": { "External id": 3689,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13273, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13273, "pid": 0, "tid": 7, "ts": 6950381216417.034, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119901.614, "dur": 1.773 , "args": { "External id": 3689,"cbid": 211, "correlation": 13273 } }, { "ph": "s", "id": 13273, "pid": 1728719, "tid": 1728719, "ts": 6950381119901.614, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381216649.639, "dur": 58.976 , "args": { "External id": 3701,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13287, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13287, "pid": 0, "tid": 7, "ts": 6950381216649.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119910.841, "dur": 1.573 , "args": { "External id": 3701,"cbid": 307, "correlation": 13287 } }, { "ph": "s", "id": 13287, "pid": 1728719, "tid": 1728719, "ts": 6950381119910.841, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381216708.871, "dur": 66.559 , "args": { "External id": 3702,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13297, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13297, "pid": 0, "tid": 7, "ts": 6950381216708.871, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119921.070, "dur": 1.663 , "args": { "External id": 3702,"cbid": 307, "correlation": 13297 } }, { "ph": "s", "id": 13297, "pid": 1728719, "tid": 1728719, "ts": 6950381119921.070, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381216775.718, "dur": 37.535 , "args": { "External id": 3703,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13306, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13306, "pid": 0, "tid": 7, "ts": 6950381216775.718, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119930.538, "dur": 1.823 , "args": { "External id": 3703,"cbid": 307, "correlation": 13306 } }, { "ph": "s", "id": 13306, "pid": 1728719, "tid": 1728719, "ts": 6950381119930.538, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119945.696, "dur": 0.110 , "args": { "External id": 3711,"cbid": 200, "correlation": 13341 } }, { "ph": "f", "id": 13341, "pid": 1728719, "tid": 1728719, "ts": 6950381119945.696, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119945.886, "dur": 0.080 , "args": { "External id": 3711,"cbid": 200, "correlation": 13342 } }, { "ph": "f", "id": 13342, "pid": 1728719, "tid": 1728719, "ts": 6950381119945.886, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119946.057, "dur": 0.460 , "args": { "External id": 3711,"cbid": 273, "correlation": 13343 } }, { "ph": "f", "id": 13343, "pid": 1728719, "tid": 1728719, "ts": 6950381119946.057, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381216813.510, "dur": 837.718 , "args": { "External id": 3711,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13345, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13345, "pid": 0, "tid": 7, "ts": 6950381216813.510, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119946.838, "dur": 1.793 , "args": { "External id": 3711,"cbid": 211, "correlation": 13345 } }, { "ph": "s", "id": 13345, "pid": 1728719, "tid": 1728719, "ts": 6950381119946.838, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381217651.516, "dur": 9.344 , "args": { "External id": 3720,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13365, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13365, "pid": 0, "tid": 7, "ts": 6950381217651.516, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119963.349, "dur": 1.613 , "args": { "External id": 3720,"cbid": 652, "correlation": 13365 } }, { "ph": "s", "id": 13365, "pid": 1728719, "tid": 1728719, "ts": 6950381119963.349, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119974.359, "dur": 0.121 , "args": { "External id": 3721,"cbid": 200, "correlation": 13380 } }, { "ph": "f", "id": 13380, "pid": 1728719, "tid": 1728719, "ts": 6950381119974.359, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119974.570, "dur": 0.140 , "args": { "External id": 3721,"cbid": 468, "correlation": 13381 } }, { "ph": "f", "id": 13381, "pid": 1728719, "tid": 1728719, "ts": 6950381119974.570, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119975.021, "dur": 0.040 , "args": { "External id": 3721,"cbid": 1, "correlation": 13383 } }, { "ph": "f", "id": 13383, "pid": 1728719, "tid": 1728719, "ts": 6950381119975.021, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119975.211, "dur": 0.090 , "args": { "External id": 3721,"cbid": 468, "correlation": 13384 } }, { "ph": "f", "id": 13384, "pid": 1728719, "tid": 1728719, "ts": 6950381119975.211, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119975.451, "dur": 0.020 , "args": { "External id": 3721,"cbid": 1, "correlation": 13386 } }, { "ph": "f", "id": 13386, "pid": 1728719, "tid": 1728719, "ts": 6950381119975.451, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119975.572, "dur": 0.100 , "args": { "External id": 3721,"cbid": 468, "correlation": 13387 } }, { "ph": "f", "id": 13387, "pid": 1728719, "tid": 1728719, "ts": 6950381119975.572, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381119975.842, "dur": 0.020 , "args": { "External id": 3721,"cbid": 1, "correlation": 13389 } }, { "ph": "f", "id": 13389, "pid": 1728719, "tid": 1728719, "ts": 6950381119975.842, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381119975.942, "dur": 0.391 , "args": { "External id": 3721,"cbid": 273, "correlation": 13390 } }, { "ph": "f", "id": 13390, "pid": 1728719, "tid": 1728719, "ts": 6950381119975.942, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381217661.116, "dur": 206.558 , "args": { "External id": 3721,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13393, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13393, "pid": 0, "tid": 7, "ts": 6950381217661.116, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119976.734, "dur": 1.753 , "args": { "External id": 3721,"cbid": 211, "correlation": 13393 } }, { "ph": "s", "id": 13393, "pid": 1728719, "tid": 1728719, "ts": 6950381119976.734, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381217867.962, "dur": 45.760 , "args": { "External id": 3733,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13405, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13405, "pid": 0, "tid": 7, "ts": 6950381217867.962, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381119986.552, "dur": 1.733 , "args": { "External id": 3733,"cbid": 307, "correlation": 13405 } }, { "ph": "s", "id": 13405, "pid": 1728719, "tid": 1728719, "ts": 6950381119986.552, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381217914.010, "dur": 13.375 , "args": { "External id": 3734,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13411, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13411, "pid": 0, "tid": 7, "ts": 6950381217914.010, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381119994.677, "dur": 1.563 , "args": { "External id": 3734,"cbid": 652, "correlation": 13411 } }, { "ph": "s", "id": 13411, "pid": 1728719, "tid": 1728719, "ts": 6950381119994.677, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381217927.706, "dur": 12.639 , "args": { "External id": 3735,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13417, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13417, "pid": 0, "tid": 7, "ts": 6950381217927.706, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120002.922, "dur": 1.543 , "args": { "External id": 3735,"cbid": 652, "correlation": 13417 } }, { "ph": "s", "id": 13417, "pid": 1728719, "tid": 1728719, "ts": 6950381120002.922, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120014.394, "dur": 0.090 , "args": { "External id": 3736,"cbid": 200, "correlation": 13436 } }, { "ph": "f", "id": 13436, "pid": 1728719, "tid": 1728719, "ts": 6950381120014.394, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120014.594, "dur": 0.130 , "args": { "External id": 3736,"cbid": 468, "correlation": 13437 } }, { "ph": "f", "id": 13437, "pid": 1728719, "tid": 1728719, "ts": 6950381120014.594, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120014.915, "dur": 0.020 , "args": { "External id": 3736,"cbid": 1, "correlation": 13439 } }, { "ph": "f", "id": 13439, "pid": 1728719, "tid": 1728719, "ts": 6950381120014.915, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120015.005, "dur": 0.100 , "args": { "External id": 3736,"cbid": 468, "correlation": 13440 } }, { "ph": "f", "id": 13440, "pid": 1728719, "tid": 1728719, "ts": 6950381120015.005, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120015.245, "dur": 0.020 , "args": { "External id": 3736,"cbid": 1, "correlation": 13442 } }, { "ph": "f", "id": 13442, "pid": 1728719, "tid": 1728719, "ts": 6950381120015.245, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120015.366, "dur": 0.100 , "args": { "External id": 3736,"cbid": 468, "correlation": 13443 } }, { "ph": "f", "id": 13443, "pid": 1728719, "tid": 1728719, "ts": 6950381120015.366, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120015.616, "dur": 0.020 , "args": { "External id": 3736,"cbid": 1, "correlation": 13445 } }, { "ph": "f", "id": 13445, "pid": 1728719, "tid": 1728719, "ts": 6950381120015.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120015.726, "dur": 0.371 , "args": { "External id": 3736,"cbid": 273, "correlation": 13446 } }, { "ph": "f", "id": 13446, "pid": 1728719, "tid": 1728719, "ts": 6950381120015.726, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381217940.665, "dur": 661.753 , "args": { "External id": 3736,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13449, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13449, "pid": 0, "tid": 7, "ts": 6950381217940.665, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120016.448, "dur": 1.713 , "args": { "External id": 3736,"cbid": 211, "correlation": 13449 } }, { "ph": "s", "id": 13449, "pid": 1728719, "tid": 1728719, "ts": 6950381120016.448, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120028.280, "dur": 0.090 , "args": { "External id": 3748,"cbid": 200, "correlation": 13473 } }, { "ph": "f", "id": 13473, "pid": 1728719, "tid": 1728719, "ts": 6950381120028.280, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120028.470, "dur": 0.120 , "args": { "External id": 3748,"cbid": 468, "correlation": 13474 } }, { "ph": "f", "id": 13474, "pid": 1728719, "tid": 1728719, "ts": 6950381120028.470, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120028.771, "dur": 0.020 , "args": { "External id": 3748,"cbid": 1, "correlation": 13476 } }, { "ph": "f", "id": 13476, "pid": 1728719, "tid": 1728719, "ts": 6950381120028.771, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120028.861, "dur": 0.100 , "args": { "External id": 3748,"cbid": 468, "correlation": 13477 } }, { "ph": "f", "id": 13477, "pid": 1728719, "tid": 1728719, "ts": 6950381120028.861, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120029.131, "dur": 0.020 , "args": { "External id": 3748,"cbid": 1, "correlation": 13479 } }, { "ph": "f", "id": 13479, "pid": 1728719, "tid": 1728719, "ts": 6950381120029.131, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120029.252, "dur": 0.090 , "args": { "External id": 3748,"cbid": 468, "correlation": 13480 } }, { "ph": "f", "id": 13480, "pid": 1728719, "tid": 1728719, "ts": 6950381120029.252, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120029.492, "dur": 0.020 , "args": { "External id": 3748,"cbid": 1, "correlation": 13482 } }, { "ph": "f", "id": 13482, "pid": 1728719, "tid": 1728719, "ts": 6950381120029.492, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120029.592, "dur": 0.331 , "args": { "External id": 3748,"cbid": 273, "correlation": 13483 } }, { "ph": "f", "id": 13483, "pid": 1728719, "tid": 1728719, "ts": 6950381120029.592, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381218602.706, "dur": 674.937 , "args": { "External id": 3748,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13486, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13486, "pid": 0, "tid": 7, "ts": 6950381218602.706, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120030.263, "dur": 1.724 , "args": { "External id": 3748,"cbid": 211, "correlation": 13486 } }, { "ph": "s", "id": 13486, "pid": 1728719, "tid": 1728719, "ts": 6950381120030.263, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381219277.963, "dur": 153.822 , "args": { "External id": 3760,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13499, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13499, "pid": 0, "tid": 7, "ts": 6950381219277.963, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120039.581, "dur": 1.613 , "args": { "External id": 3760,"cbid": 307, "correlation": 13499 } }, { "ph": "s", "id": 13499, "pid": 1728719, "tid": 1728719, "ts": 6950381120039.581, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381219432.073, "dur": 58.719 , "args": { "External id": 3761,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13505, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13505, "pid": 0, "tid": 7, "ts": 6950381219432.073, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120047.686, "dur": 1.553 , "args": { "External id": 3761,"cbid": 652, "correlation": 13505 } }, { "ph": "s", "id": 13505, "pid": 1728719, "tid": 1728719, "ts": 6950381120047.686, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120058.957, "dur": 0.080 , "args": { "External id": 3762,"cbid": 200, "correlation": 13524 } }, { "ph": "f", "id": 13524, "pid": 1728719, "tid": 1728719, "ts": 6950381120058.957, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120059.147, "dur": 0.110 , "args": { "External id": 3762,"cbid": 468, "correlation": 13525 } }, { "ph": "f", "id": 13525, "pid": 1728719, "tid": 1728719, "ts": 6950381120059.147, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120059.448, "dur": 0.020 , "args": { "External id": 3762,"cbid": 1, "correlation": 13527 } }, { "ph": "f", "id": 13527, "pid": 1728719, "tid": 1728719, "ts": 6950381120059.448, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120059.558, "dur": 0.110 , "args": { "External id": 3762,"cbid": 468, "correlation": 13528 } }, { "ph": "f", "id": 13528, "pid": 1728719, "tid": 1728719, "ts": 6950381120059.558, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120059.808, "dur": 0.020 , "args": { "External id": 3762,"cbid": 1, "correlation": 13530 } }, { "ph": "f", "id": 13530, "pid": 1728719, "tid": 1728719, "ts": 6950381120059.808, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120059.929, "dur": 0.090 , "args": { "External id": 3762,"cbid": 468, "correlation": 13531 } }, { "ph": "f", "id": 13531, "pid": 1728719, "tid": 1728719, "ts": 6950381120059.929, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120060.169, "dur": 0.020 , "args": { "External id": 3762,"cbid": 1, "correlation": 13533 } }, { "ph": "f", "id": 13533, "pid": 1728719, "tid": 1728719, "ts": 6950381120060.169, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120060.279, "dur": 0.371 , "args": { "External id": 3762,"cbid": 273, "correlation": 13534 } }, { "ph": "f", "id": 13534, "pid": 1728719, "tid": 1728719, "ts": 6950381120060.279, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381219491.080, "dur": 675.993 , "args": { "External id": 3762,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13537, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13537, "pid": 0, "tid": 7, "ts": 6950381219491.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120061.001, "dur": 1.703 , "args": { "External id": 3762,"cbid": 211, "correlation": 13537 } }, { "ph": "s", "id": 13537, "pid": 1728719, "tid": 1728719, "ts": 6950381120061.001, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381220167.361, "dur": 40.832 , "args": { "External id": 3774,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13548, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13548, "pid": 0, "tid": 7, "ts": 6950381220167.361, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120070.769, "dur": 1.703 , "args": { "External id": 3774,"cbid": 307, "correlation": 13548 } }, { "ph": "s", "id": 13548, "pid": 1728719, "tid": 1728719, "ts": 6950381120070.769, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381220208.449, "dur": 21.055 , "args": { "External id": 3786,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13560, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13560, "pid": 0, "tid": 7, "ts": 6950381220208.449, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120148.804, "dur": 2.685 , "args": { "External id": 3786,"cbid": 307, "correlation": 13560 } }, { "ph": "s", "id": 13560, "pid": 1728719, "tid": 1728719, "ts": 6950381120148.804, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381220229.824, "dur": 12.352 , "args": { "External id": 3787,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13574, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13574, "pid": 0, "tid": 7, "ts": 6950381220229.824, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120163.151, "dur": 1.693 , "args": { "External id": 3787,"cbid": 652, "correlation": 13574 } }, { "ph": "s", "id": 13574, "pid": 1728719, "tid": 1728719, "ts": 6950381120163.151, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381220242.496, "dur": 11.968 , "args": { "External id": 3788,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13588, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13588, "pid": 0, "tid": 7, "ts": 6950381220242.496, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120174.662, "dur": 1.603 , "args": { "External id": 3788,"cbid": 652, "correlation": 13588 } }, { "ph": "s", "id": 13588, "pid": 1728719, "tid": 1728719, "ts": 6950381120174.662, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381220254.752, "dur": 11.040 , "args": { "External id": 3789,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13602, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13602, "pid": 0, "tid": 7, "ts": 6950381220254.752, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120185.252, "dur": 1.543 , "args": { "External id": 3789,"cbid": 652, "correlation": 13602 } }, { "ph": "s", "id": 13602, "pid": 1728719, "tid": 1728719, "ts": 6950381120185.252, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120196.683, "dur": 0.130 , "args": { "External id": 3790,"cbid": 200, "correlation": 13617 } }, { "ph": "f", "id": 13617, "pid": 1728719, "tid": 1728719, "ts": 6950381120196.683, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120196.984, "dur": 0.180 , "args": { "External id": 3790,"cbid": 468, "correlation": 13618 } }, { "ph": "f", "id": 13618, "pid": 1728719, "tid": 1728719, "ts": 6950381120196.984, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120197.525, "dur": 0.020 , "args": { "External id": 3790,"cbid": 1, "correlation": 13620 } }, { "ph": "f", "id": 13620, "pid": 1728719, "tid": 1728719, "ts": 6950381120197.525, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120197.625, "dur": 0.100 , "args": { "External id": 3790,"cbid": 468, "correlation": 13621 } }, { "ph": "f", "id": 13621, "pid": 1728719, "tid": 1728719, "ts": 6950381120197.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120197.875, "dur": 0.030 , "args": { "External id": 3790,"cbid": 1, "correlation": 13623 } }, { "ph": "f", "id": 13623, "pid": 1728719, "tid": 1728719, "ts": 6950381120197.875, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120198.016, "dur": 0.100 , "args": { "External id": 3790,"cbid": 468, "correlation": 13624 } }, { "ph": "f", "id": 13624, "pid": 1728719, "tid": 1728719, "ts": 6950381120198.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120198.276, "dur": 0.020 , "args": { "External id": 3790,"cbid": 1, "correlation": 13626 } }, { "ph": "f", "id": 13626, "pid": 1728719, "tid": 1728719, "ts": 6950381120198.276, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120198.436, "dur": 0.712 , "args": { "External id": 3790,"cbid": 273, "correlation": 13627 } }, { "ph": "f", "id": 13627, "pid": 1728719, "tid": 1728719, "ts": 6950381120198.436, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381220266.080, "dur": 216.830 , "args": { "External id": 3790,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13630, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13630, "pid": 0, "tid": 7, "ts": 6950381220266.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120199.569, "dur": 1.903 , "args": { "External id": 3790,"cbid": 211, "correlation": 13630 } }, { "ph": "s", "id": 13630, "pid": 1728719, "tid": 1728719, "ts": 6950381120199.569, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381220483.198, "dur": 8.768 , "args": { "External id": 3802,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13643, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13643, "pid": 0, "tid": 7, "ts": 6950381220483.198, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120210.048, "dur": 1.633 , "args": { "External id": 3802,"cbid": 307, "correlation": 13643 } }, { "ph": "s", "id": 13643, "pid": 1728719, "tid": 1728719, "ts": 6950381120210.048, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381220492.254, "dur": 35.871 , "args": { "External id": 3807,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13666, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13666, "pid": 0, "tid": 7, "ts": 6950381220492.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120229.003, "dur": 1.714 , "args": { "External id": 3807,"cbid": 307, "correlation": 13666 } }, { "ph": "s", "id": 13666, "pid": 1728719, "tid": 1728719, "ts": 6950381120229.003, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120239.232, "dur": 0.101 , "args": { "External id": 3808,"cbid": 200, "correlation": 13681 } }, { "ph": "f", "id": 13681, "pid": 1728719, "tid": 1728719, "ts": 6950381120239.232, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120239.433, "dur": 0.130 , "args": { "External id": 3808,"cbid": 468, "correlation": 13682 } }, { "ph": "f", "id": 13682, "pid": 1728719, "tid": 1728719, "ts": 6950381120239.433, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120239.783, "dur": 0.030 , "args": { "External id": 3808,"cbid": 1, "correlation": 13684 } }, { "ph": "f", "id": 13684, "pid": 1728719, "tid": 1728719, "ts": 6950381120239.783, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120239.894, "dur": 0.100 , "args": { "External id": 3808,"cbid": 468, "correlation": 13685 } }, { "ph": "f", "id": 13685, "pid": 1728719, "tid": 1728719, "ts": 6950381120239.894, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120240.144, "dur": 0.020 , "args": { "External id": 3808,"cbid": 1, "correlation": 13687 } }, { "ph": "f", "id": 13687, "pid": 1728719, "tid": 1728719, "ts": 6950381120240.144, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120240.254, "dur": 0.100 , "args": { "External id": 3808,"cbid": 468, "correlation": 13688 } }, { "ph": "f", "id": 13688, "pid": 1728719, "tid": 1728719, "ts": 6950381120240.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120240.515, "dur": 0.020 , "args": { "External id": 3808,"cbid": 1, "correlation": 13690 } }, { "ph": "f", "id": 13690, "pid": 1728719, "tid": 1728719, "ts": 6950381120240.515, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120240.625, "dur": 0.481 , "args": { "External id": 3808,"cbid": 273, "correlation": 13691 } }, { "ph": "f", "id": 13691, "pid": 1728719, "tid": 1728719, "ts": 6950381120240.625, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381220528.413, "dur": 229.598 , "args": { "External id": 3808,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13694, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13694, "pid": 0, "tid": 7, "ts": 6950381220528.413, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120241.487, "dur": 2.013 , "args": { "External id": 3808,"cbid": 211, "correlation": 13694 } }, { "ph": "s", "id": 13694, "pid": 1728719, "tid": 1728719, "ts": 6950381120241.487, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381220758.299, "dur": 8.800 , "args": { "External id": 3820,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13703, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13703, "pid": 0, "tid": 7, "ts": 6950381220758.299, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120249.562, "dur": 1.512 , "args": { "External id": 3820,"cbid": 307, "correlation": 13703 } }, { "ph": "s", "id": 13703, "pid": 1728719, "tid": 1728719, "ts": 6950381120249.562, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381220767.387, "dur": 25.599 , "args": { "External id": 3825,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13726, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13726, "pid": 0, "tid": 7, "ts": 6950381220767.387, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120265.521, "dur": 1.563 , "args": { "External id": 3825,"cbid": 307, "correlation": 13726 } }, { "ph": "s", "id": 13726, "pid": 1728719, "tid": 1728719, "ts": 6950381120265.521, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120275.550, "dur": 0.080 , "args": { "External id": 3826,"cbid": 200, "correlation": 13741 } }, { "ph": "f", "id": 13741, "pid": 1728719, "tid": 1728719, "ts": 6950381120275.550, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120275.720, "dur": 0.121 , "args": { "External id": 3826,"cbid": 468, "correlation": 13742 } }, { "ph": "f", "id": 13742, "pid": 1728719, "tid": 1728719, "ts": 6950381120275.720, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120276.041, "dur": 0.020 , "args": { "External id": 3826,"cbid": 1, "correlation": 13744 } }, { "ph": "f", "id": 13744, "pid": 1728719, "tid": 1728719, "ts": 6950381120276.041, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120276.131, "dur": 0.100 , "args": { "External id": 3826,"cbid": 468, "correlation": 13745 } }, { "ph": "f", "id": 13745, "pid": 1728719, "tid": 1728719, "ts": 6950381120276.131, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120276.392, "dur": 0.030 , "args": { "External id": 3826,"cbid": 1, "correlation": 13747 } }, { "ph": "f", "id": 13747, "pid": 1728719, "tid": 1728719, "ts": 6950381120276.392, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120276.512, "dur": 0.100 , "args": { "External id": 3826,"cbid": 468, "correlation": 13748 } }, { "ph": "f", "id": 13748, "pid": 1728719, "tid": 1728719, "ts": 6950381120276.512, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120276.762, "dur": 0.020 , "args": { "External id": 3826,"cbid": 1, "correlation": 13750 } }, { "ph": "f", "id": 13750, "pid": 1728719, "tid": 1728719, "ts": 6950381120276.762, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120276.862, "dur": 0.351 , "args": { "External id": 3826,"cbid": 273, "correlation": 13751 } }, { "ph": "f", "id": 13751, "pid": 1728719, "tid": 1728719, "ts": 6950381120276.862, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381220793.242, "dur": 232.830 , "args": { "External id": 3826,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13754, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13754, "pid": 0, "tid": 7, "ts": 6950381220793.242, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120277.574, "dur": 1.743 , "args": { "External id": 3826,"cbid": 211, "correlation": 13754 } }, { "ph": "s", "id": 13754, "pid": 1728719, "tid": 1728719, "ts": 6950381120277.574, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381221026.360, "dur": 60.831 , "args": { "External id": 3838,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13768, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13768, "pid": 0, "tid": 7, "ts": 6950381221026.360, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120287.803, "dur": 1.513 , "args": { "External id": 3838,"cbid": 307, "correlation": 13768 } }, { "ph": "s", "id": 13768, "pid": 1728719, "tid": 1728719, "ts": 6950381120287.803, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381221087.479, "dur": 63.808 , "args": { "External id": 3839,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13778, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13778, "pid": 0, "tid": 7, "ts": 6950381221087.479, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120297.761, "dur": 1.653 , "args": { "External id": 3839,"cbid": 307, "correlation": 13778 } }, { "ph": "s", "id": 13778, "pid": 1728719, "tid": 1728719, "ts": 6950381120297.761, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381221151.575, "dur": 39.775 , "args": { "External id": 3840,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13787, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13787, "pid": 0, "tid": 7, "ts": 6950381221151.575, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120307.309, "dur": 1.643 , "args": { "External id": 3840,"cbid": 307, "correlation": 13787 } }, { "ph": "s", "id": 13787, "pid": 1728719, "tid": 1728719, "ts": 6950381120307.309, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120322.537, "dur": 0.121 , "args": { "External id": 3848,"cbid": 200, "correlation": 13822 } }, { "ph": "f", "id": 13822, "pid": 1728719, "tid": 1728719, "ts": 6950381120322.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120322.748, "dur": 0.070 , "args": { "External id": 3848,"cbid": 200, "correlation": 13823 } }, { "ph": "f", "id": 13823, "pid": 1728719, "tid": 1728719, "ts": 6950381120322.748, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120322.898, "dur": 0.521 , "args": { "External id": 3848,"cbid": 273, "correlation": 13824 } }, { "ph": "f", "id": 13824, "pid": 1728719, "tid": 1728719, "ts": 6950381120322.898, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381221191.606, "dur": 838.135 , "args": { "External id": 3848,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13826, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13826, "pid": 0, "tid": 7, "ts": 6950381221191.606, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120323.740, "dur": 1.783 , "args": { "External id": 3848,"cbid": 211, "correlation": 13826 } }, { "ph": "s", "id": 13826, "pid": 1728719, "tid": 1728719, "ts": 6950381120323.740, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381222029.997, "dur": 9.280 , "args": { "External id": 3857,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13846, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13846, "pid": 0, "tid": 7, "ts": 6950381222029.997, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120339.950, "dur": 1.683 , "args": { "External id": 3857,"cbid": 652, "correlation": 13846 } }, { "ph": "s", "id": 13846, "pid": 1728719, "tid": 1728719, "ts": 6950381120339.950, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120350.810, "dur": 0.110 , "args": { "External id": 3858,"cbid": 200, "correlation": 13861 } }, { "ph": "f", "id": 13861, "pid": 1728719, "tid": 1728719, "ts": 6950381120350.810, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120351.010, "dur": 0.151 , "args": { "External id": 3858,"cbid": 468, "correlation": 13862 } }, { "ph": "f", "id": 13862, "pid": 1728719, "tid": 1728719, "ts": 6950381120351.010, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120351.421, "dur": 0.030 , "args": { "External id": 3858,"cbid": 1, "correlation": 13864 } }, { "ph": "f", "id": 13864, "pid": 1728719, "tid": 1728719, "ts": 6950381120351.421, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120351.511, "dur": 0.101 , "args": { "External id": 3858,"cbid": 468, "correlation": 13865 } }, { "ph": "f", "id": 13865, "pid": 1728719, "tid": 1728719, "ts": 6950381120351.511, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120351.762, "dur": 0.020 , "args": { "External id": 3858,"cbid": 1, "correlation": 13867 } }, { "ph": "f", "id": 13867, "pid": 1728719, "tid": 1728719, "ts": 6950381120351.762, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120351.872, "dur": 0.100 , "args": { "External id": 3858,"cbid": 468, "correlation": 13868 } }, { "ph": "f", "id": 13868, "pid": 1728719, "tid": 1728719, "ts": 6950381120351.872, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120352.133, "dur": 0.020 , "args": { "External id": 3858,"cbid": 1, "correlation": 13870 } }, { "ph": "f", "id": 13870, "pid": 1728719, "tid": 1728719, "ts": 6950381120352.133, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120352.233, "dur": 0.400 , "args": { "External id": 3858,"cbid": 273, "correlation": 13871 } }, { "ph": "f", "id": 13871, "pid": 1728719, "tid": 1728719, "ts": 6950381120352.233, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381222039.565, "dur": 209.406 , "args": { "External id": 3858,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13874, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13874, "pid": 0, "tid": 7, "ts": 6950381222039.565, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120352.994, "dur": 1.713 , "args": { "External id": 3858,"cbid": 211, "correlation": 13874 } }, { "ph": "s", "id": 13874, "pid": 1728719, "tid": 1728719, "ts": 6950381120352.994, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381222249.227, "dur": 46.047 , "args": { "External id": 3870,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13886, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13886, "pid": 0, "tid": 7, "ts": 6950381222249.227, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120362.642, "dur": 1.673 , "args": { "External id": 3870,"cbid": 307, "correlation": 13886 } }, { "ph": "s", "id": 13886, "pid": 1728719, "tid": 1728719, "ts": 6950381120362.642, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381222295.594, "dur": 12.992 , "args": { "External id": 3871,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13892, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13892, "pid": 0, "tid": 7, "ts": 6950381222295.594, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120370.667, "dur": 1.523 , "args": { "External id": 3871,"cbid": 652, "correlation": 13892 } }, { "ph": "s", "id": 13892, "pid": 1728719, "tid": 1728719, "ts": 6950381120370.667, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381222308.874, "dur": 12.512 , "args": { "External id": 3872,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13898, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13898, "pid": 0, "tid": 7, "ts": 6950381222308.874, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120379.063, "dur": 1.553 , "args": { "External id": 3872,"cbid": 652, "correlation": 13898 } }, { "ph": "s", "id": 13898, "pid": 1728719, "tid": 1728719, "ts": 6950381120379.063, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120390.694, "dur": 0.111 , "args": { "External id": 3873,"cbid": 200, "correlation": 13917 } }, { "ph": "f", "id": 13917, "pid": 1728719, "tid": 1728719, "ts": 6950381120390.694, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120390.915, "dur": 0.120 , "args": { "External id": 3873,"cbid": 468, "correlation": 13918 } }, { "ph": "f", "id": 13918, "pid": 1728719, "tid": 1728719, "ts": 6950381120390.915, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120391.235, "dur": 0.020 , "args": { "External id": 3873,"cbid": 1, "correlation": 13920 } }, { "ph": "f", "id": 13920, "pid": 1728719, "tid": 1728719, "ts": 6950381120391.235, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120391.626, "dur": 0.090 , "args": { "External id": 3873,"cbid": 468, "correlation": 13921 } }, { "ph": "f", "id": 13921, "pid": 1728719, "tid": 1728719, "ts": 6950381120391.626, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120391.866, "dur": 0.021 , "args": { "External id": 3873,"cbid": 1, "correlation": 13923 } }, { "ph": "f", "id": 13923, "pid": 1728719, "tid": 1728719, "ts": 6950381120391.866, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120391.977, "dur": 0.100 , "args": { "External id": 3873,"cbid": 468, "correlation": 13924 } }, { "ph": "f", "id": 13924, "pid": 1728719, "tid": 1728719, "ts": 6950381120391.977, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120392.227, "dur": 0.020 , "args": { "External id": 3873,"cbid": 1, "correlation": 13926 } }, { "ph": "f", "id": 13926, "pid": 1728719, "tid": 1728719, "ts": 6950381120392.227, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120392.327, "dur": 0.391 , "args": { "External id": 3873,"cbid": 273, "correlation": 13927 } }, { "ph": "f", "id": 13927, "pid": 1728719, "tid": 1728719, "ts": 6950381120392.327, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381222321.706, "dur": 664.729 , "args": { "External id": 3873,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13930, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13930, "pid": 0, "tid": 7, "ts": 6950381222321.706, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120393.079, "dur": 1.703 , "args": { "External id": 3873,"cbid": 211, "correlation": 13930 } }, { "ph": "s", "id": 13930, "pid": 1728719, "tid": 1728719, "ts": 6950381120393.079, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120404.861, "dur": 0.090 , "args": { "External id": 3885,"cbid": 200, "correlation": 13954 } }, { "ph": "f", "id": 13954, "pid": 1728719, "tid": 1728719, "ts": 6950381120404.861, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120405.051, "dur": 0.120 , "args": { "External id": 3885,"cbid": 468, "correlation": 13955 } }, { "ph": "f", "id": 13955, "pid": 1728719, "tid": 1728719, "ts": 6950381120405.051, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120405.352, "dur": 0.020 , "args": { "External id": 3885,"cbid": 1, "correlation": 13957 } }, { "ph": "f", "id": 13957, "pid": 1728719, "tid": 1728719, "ts": 6950381120405.352, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120405.442, "dur": 0.100 , "args": { "External id": 3885,"cbid": 468, "correlation": 13958 } }, { "ph": "f", "id": 13958, "pid": 1728719, "tid": 1728719, "ts": 6950381120405.442, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120405.702, "dur": 0.020 , "args": { "External id": 3885,"cbid": 1, "correlation": 13960 } }, { "ph": "f", "id": 13960, "pid": 1728719, "tid": 1728719, "ts": 6950381120405.702, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120405.822, "dur": 0.101 , "args": { "External id": 3885,"cbid": 468, "correlation": 13961 } }, { "ph": "f", "id": 13961, "pid": 1728719, "tid": 1728719, "ts": 6950381120405.822, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120406.073, "dur": 0.020 , "args": { "External id": 3885,"cbid": 1, "correlation": 13963 } }, { "ph": "f", "id": 13963, "pid": 1728719, "tid": 1728719, "ts": 6950381120406.073, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120406.173, "dur": 0.331 , "args": { "External id": 3885,"cbid": 273, "correlation": 13964 } }, { "ph": "f", "id": 13964, "pid": 1728719, "tid": 1728719, "ts": 6950381120406.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381222986.723, "dur": 676.504 , "args": { "External id": 3885,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13967, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13967, "pid": 0, "tid": 7, "ts": 6950381222986.723, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120406.864, "dur": 1.653 , "args": { "External id": 3885,"cbid": 211, "correlation": 13967 } }, { "ph": "s", "id": 13967, "pid": 1728719, "tid": 1728719, "ts": 6950381120406.864, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381223663.515, "dur": 152.799 , "args": { "External id": 3897,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13980, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13980, "pid": 0, "tid": 7, "ts": 6950381223663.515, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120416.182, "dur": 1.723 , "args": { "External id": 3897,"cbid": 307, "correlation": 13980 } }, { "ph": "s", "id": 13980, "pid": 1728719, "tid": 1728719, "ts": 6950381120416.182, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381223816.570, "dur": 59.487 , "args": { "External id": 3898,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 13986, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 13986, "pid": 0, "tid": 7, "ts": 6950381223816.570, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120424.507, "dur": 1.623 , "args": { "External id": 3898,"cbid": 652, "correlation": 13986 } }, { "ph": "s", "id": 13986, "pid": 1728719, "tid": 1728719, "ts": 6950381120424.507, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120435.818, "dur": 0.100 , "args": { "External id": 3899,"cbid": 200, "correlation": 14005 } }, { "ph": "f", "id": 14005, "pid": 1728719, "tid": 1728719, "ts": 6950381120435.818, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120436.029, "dur": 0.110 , "args": { "External id": 3899,"cbid": 468, "correlation": 14006 } }, { "ph": "f", "id": 14006, "pid": 1728719, "tid": 1728719, "ts": 6950381120436.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120436.339, "dur": 0.030 , "args": { "External id": 3899,"cbid": 1, "correlation": 14008 } }, { "ph": "f", "id": 14008, "pid": 1728719, "tid": 1728719, "ts": 6950381120436.339, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120436.449, "dur": 0.101 , "args": { "External id": 3899,"cbid": 468, "correlation": 14009 } }, { "ph": "f", "id": 14009, "pid": 1728719, "tid": 1728719, "ts": 6950381120436.449, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120436.700, "dur": 0.030 , "args": { "External id": 3899,"cbid": 1, "correlation": 14011 } }, { "ph": "f", "id": 14011, "pid": 1728719, "tid": 1728719, "ts": 6950381120436.700, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120436.820, "dur": 0.090 , "args": { "External id": 3899,"cbid": 468, "correlation": 14012 } }, { "ph": "f", "id": 14012, "pid": 1728719, "tid": 1728719, "ts": 6950381120436.820, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120437.061, "dur": 0.020 , "args": { "External id": 3899,"cbid": 1, "correlation": 14014 } }, { "ph": "f", "id": 14014, "pid": 1728719, "tid": 1728719, "ts": 6950381120437.061, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120437.171, "dur": 0.371 , "args": { "External id": 3899,"cbid": 273, "correlation": 14015 } }, { "ph": "f", "id": 14015, "pid": 1728719, "tid": 1728719, "ts": 6950381120437.171, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381223876.345, "dur": 700.601 , "args": { "External id": 3899,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14018, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14018, "pid": 0, "tid": 7, "ts": 6950381223876.345, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120437.892, "dur": 1.743 , "args": { "External id": 3899,"cbid": 211, "correlation": 14018 } }, { "ph": "s", "id": 14018, "pid": 1728719, "tid": 1728719, "ts": 6950381120437.892, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381224577.234, "dur": 40.383 , "args": { "External id": 3911,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14029, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14029, "pid": 0, "tid": 7, "ts": 6950381224577.234, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120447.590, "dur": 1.683 , "args": { "External id": 3911,"cbid": 307, "correlation": 14029 } }, { "ph": "s", "id": 14029, "pid": 1728719, "tid": 1728719, "ts": 6950381120447.590, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381224617.905, "dur": 21.280 , "args": { "External id": 3923,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14041, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14041, "pid": 0, "tid": 7, "ts": 6950381224617.905, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120526.016, "dur": 3.306 , "args": { "External id": 3923,"cbid": 307, "correlation": 14041 } }, { "ph": "s", "id": 14041, "pid": 1728719, "tid": 1728719, "ts": 6950381120526.016, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381224639.505, "dur": 12.960 , "args": { "External id": 3924,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14055, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14055, "pid": 0, "tid": 7, "ts": 6950381224639.505, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120544.591, "dur": 1.693 , "args": { "External id": 3924,"cbid": 652, "correlation": 14055 } }, { "ph": "s", "id": 14055, "pid": 1728719, "tid": 1728719, "ts": 6950381120544.591, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381224652.785, "dur": 12.639 , "args": { "External id": 3925,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14069, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14069, "pid": 0, "tid": 7, "ts": 6950381224652.785, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120556.413, "dur": 1.573 , "args": { "External id": 3925,"cbid": 652, "correlation": 14069 } }, { "ph": "s", "id": 14069, "pid": 1728719, "tid": 1728719, "ts": 6950381120556.413, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381224665.713, "dur": 10.783 , "args": { "External id": 3926,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14083, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14083, "pid": 0, "tid": 7, "ts": 6950381224665.713, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120567.353, "dur": 1.543 , "args": { "External id": 3926,"cbid": 652, "correlation": 14083 } }, { "ph": "s", "id": 14083, "pid": 1728719, "tid": 1728719, "ts": 6950381120567.353, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120579.015, "dur": 0.190 , "args": { "External id": 3927,"cbid": 200, "correlation": 14098 } }, { "ph": "f", "id": 14098, "pid": 1728719, "tid": 1728719, "ts": 6950381120579.015, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120579.345, "dur": 0.191 , "args": { "External id": 3927,"cbid": 468, "correlation": 14099 } }, { "ph": "f", "id": 14099, "pid": 1728719, "tid": 1728719, "ts": 6950381120579.345, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120579.866, "dur": 0.030 , "args": { "External id": 3927,"cbid": 1, "correlation": 14101 } }, { "ph": "f", "id": 14101, "pid": 1728719, "tid": 1728719, "ts": 6950381120579.866, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120579.967, "dur": 0.110 , "args": { "External id": 3927,"cbid": 468, "correlation": 14102 } }, { "ph": "f", "id": 14102, "pid": 1728719, "tid": 1728719, "ts": 6950381120579.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120580.237, "dur": 0.020 , "args": { "External id": 3927,"cbid": 1, "correlation": 14104 } }, { "ph": "f", "id": 14104, "pid": 1728719, "tid": 1728719, "ts": 6950381120580.237, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120580.367, "dur": 0.101 , "args": { "External id": 3927,"cbid": 468, "correlation": 14105 } }, { "ph": "f", "id": 14105, "pid": 1728719, "tid": 1728719, "ts": 6950381120580.367, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120580.658, "dur": 0.020 , "args": { "External id": 3927,"cbid": 1, "correlation": 14107 } }, { "ph": "f", "id": 14107, "pid": 1728719, "tid": 1728719, "ts": 6950381120580.658, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120580.808, "dur": 0.671 , "args": { "External id": 3927,"cbid": 273, "correlation": 14108 } }, { "ph": "f", "id": 14108, "pid": 1728719, "tid": 1728719, "ts": 6950381120580.808, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381224676.784, "dur": 220.254 , "args": { "External id": 3927,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14111, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14111, "pid": 0, "tid": 7, "ts": 6950381224676.784, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120581.990, "dur": 1.824 , "args": { "External id": 3927,"cbid": 211, "correlation": 14111 } }, { "ph": "s", "id": 14111, "pid": 1728719, "tid": 1728719, "ts": 6950381120581.990, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381224897.326, "dur": 8.800 , "args": { "External id": 3939,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14124, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14124, "pid": 0, "tid": 7, "ts": 6950381224897.326, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120592.680, "dur": 1.603 , "args": { "External id": 3939,"cbid": 307, "correlation": 14124 } }, { "ph": "s", "id": 14124, "pid": 1728719, "tid": 1728719, "ts": 6950381120592.680, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381224906.414, "dur": 37.023 , "args": { "External id": 3944,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14147, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14147, "pid": 0, "tid": 7, "ts": 6950381224906.414, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120611.205, "dur": 1.673 , "args": { "External id": 3944,"cbid": 307, "correlation": 14147 } }, { "ph": "s", "id": 14147, "pid": 1728719, "tid": 1728719, "ts": 6950381120611.205, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120621.344, "dur": 0.110 , "args": { "External id": 3945,"cbid": 200, "correlation": 14162 } }, { "ph": "f", "id": 14162, "pid": 1728719, "tid": 1728719, "ts": 6950381120621.344, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120621.564, "dur": 0.140 , "args": { "External id": 3945,"cbid": 468, "correlation": 14163 } }, { "ph": "f", "id": 14163, "pid": 1728719, "tid": 1728719, "ts": 6950381120621.564, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120621.925, "dur": 0.020 , "args": { "External id": 3945,"cbid": 1, "correlation": 14165 } }, { "ph": "f", "id": 14165, "pid": 1728719, "tid": 1728719, "ts": 6950381120621.925, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120622.015, "dur": 0.100 , "args": { "External id": 3945,"cbid": 468, "correlation": 14166 } }, { "ph": "f", "id": 14166, "pid": 1728719, "tid": 1728719, "ts": 6950381120622.015, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120622.255, "dur": 0.020 , "args": { "External id": 3945,"cbid": 1, "correlation": 14168 } }, { "ph": "f", "id": 14168, "pid": 1728719, "tid": 1728719, "ts": 6950381120622.255, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120622.376, "dur": 0.090 , "args": { "External id": 3945,"cbid": 468, "correlation": 14169 } }, { "ph": "f", "id": 14169, "pid": 1728719, "tid": 1728719, "ts": 6950381120622.376, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120622.626, "dur": 0.020 , "args": { "External id": 3945,"cbid": 1, "correlation": 14171 } }, { "ph": "f", "id": 14171, "pid": 1728719, "tid": 1728719, "ts": 6950381120622.626, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120622.736, "dur": 0.401 , "args": { "External id": 3945,"cbid": 273, "correlation": 14172 } }, { "ph": "f", "id": 14172, "pid": 1728719, "tid": 1728719, "ts": 6950381120622.736, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381224943.725, "dur": 233.886 , "args": { "External id": 3945,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14175, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14175, "pid": 0, "tid": 7, "ts": 6950381224943.725, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120623.528, "dur": 1.973 , "args": { "External id": 3945,"cbid": 211, "correlation": 14175 } }, { "ph": "s", "id": 14175, "pid": 1728719, "tid": 1728719, "ts": 6950381120623.528, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381225177.931, "dur": 8.896 , "args": { "External id": 3957,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14184, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14184, "pid": 0, "tid": 7, "ts": 6950381225177.931, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120631.152, "dur": 1.573 , "args": { "External id": 3957,"cbid": 307, "correlation": 14184 } }, { "ph": "s", "id": 14184, "pid": 1728719, "tid": 1728719, "ts": 6950381120631.152, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381225187.083, "dur": 26.272 , "args": { "External id": 3962,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14207, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14207, "pid": 0, "tid": 7, "ts": 6950381225187.083, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120646.741, "dur": 1.683 , "args": { "External id": 3962,"cbid": 307, "correlation": 14207 } }, { "ph": "s", "id": 14207, "pid": 1728719, "tid": 1728719, "ts": 6950381120646.741, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120656.820, "dur": 0.090 , "args": { "External id": 3963,"cbid": 200, "correlation": 14222 } }, { "ph": "f", "id": 14222, "pid": 1728719, "tid": 1728719, "ts": 6950381120656.820, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120656.990, "dur": 0.130 , "args": { "External id": 3963,"cbid": 468, "correlation": 14223 } }, { "ph": "f", "id": 14223, "pid": 1728719, "tid": 1728719, "ts": 6950381120656.990, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120657.321, "dur": 0.020 , "args": { "External id": 3963,"cbid": 1, "correlation": 14225 } }, { "ph": "f", "id": 14225, "pid": 1728719, "tid": 1728719, "ts": 6950381120657.321, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120657.411, "dur": 0.100 , "args": { "External id": 3963,"cbid": 468, "correlation": 14226 } }, { "ph": "f", "id": 14226, "pid": 1728719, "tid": 1728719, "ts": 6950381120657.411, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120657.661, "dur": 0.020 , "args": { "External id": 3963,"cbid": 1, "correlation": 14228 } }, { "ph": "f", "id": 14228, "pid": 1728719, "tid": 1728719, "ts": 6950381120657.661, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120657.771, "dur": 0.101 , "args": { "External id": 3963,"cbid": 468, "correlation": 14229 } }, { "ph": "f", "id": 14229, "pid": 1728719, "tid": 1728719, "ts": 6950381120657.771, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120658.022, "dur": 0.020 , "args": { "External id": 3963,"cbid": 1, "correlation": 14231 } }, { "ph": "f", "id": 14231, "pid": 1728719, "tid": 1728719, "ts": 6950381120658.022, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120658.122, "dur": 0.381 , "args": { "External id": 3963,"cbid": 273, "correlation": 14232 } }, { "ph": "f", "id": 14232, "pid": 1728719, "tid": 1728719, "ts": 6950381120658.122, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381225213.611, "dur": 237.693 , "args": { "External id": 3963,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14235, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14235, "pid": 0, "tid": 7, "ts": 6950381225213.611, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120658.874, "dur": 1.793 , "args": { "External id": 3963,"cbid": 211, "correlation": 14235 } }, { "ph": "s", "id": 14235, "pid": 1728719, "tid": 1728719, "ts": 6950381120658.874, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381225451.624, "dur": 58.815 , "args": { "External id": 3975,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14249, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14249, "pid": 0, "tid": 7, "ts": 6950381225451.624, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120667.950, "dur": 1.583 , "args": { "External id": 3975,"cbid": 307, "correlation": 14249 } }, { "ph": "s", "id": 14249, "pid": 1728719, "tid": 1728719, "ts": 6950381120667.950, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381225510.727, "dur": 66.271 , "args": { "External id": 3976,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14259, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14259, "pid": 0, "tid": 7, "ts": 6950381225510.727, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120678.189, "dur": 1.704 , "args": { "External id": 3976,"cbid": 307, "correlation": 14259 } }, { "ph": "s", "id": 14259, "pid": 1728719, "tid": 1728719, "ts": 6950381120678.189, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381225577.255, "dur": 40.223 , "args": { "External id": 3977,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14268, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14268, "pid": 0, "tid": 7, "ts": 6950381225577.255, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120687.898, "dur": 1.623 , "args": { "External id": 3977,"cbid": 307, "correlation": 14268 } }, { "ph": "s", "id": 14268, "pid": 1728719, "tid": 1728719, "ts": 6950381120687.898, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120703.557, "dur": 0.120 , "args": { "External id": 3985,"cbid": 200, "correlation": 14303 } }, { "ph": "f", "id": 14303, "pid": 1728719, "tid": 1728719, "ts": 6950381120703.557, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120703.747, "dur": 0.080 , "args": { "External id": 3985,"cbid": 200, "correlation": 14304 } }, { "ph": "f", "id": 14304, "pid": 1728719, "tid": 1728719, "ts": 6950381120703.747, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120703.917, "dur": 0.461 , "args": { "External id": 3985,"cbid": 273, "correlation": 14305 } }, { "ph": "f", "id": 14305, "pid": 1728719, "tid": 1728719, "ts": 6950381120703.917, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381225617.734, "dur": 853.559 , "args": { "External id": 3985,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14307, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14307, "pid": 0, "tid": 7, "ts": 6950381225617.734, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120704.789, "dur": 1.863 , "args": { "External id": 3985,"cbid": 211, "correlation": 14307 } }, { "ph": "s", "id": 14307, "pid": 1728719, "tid": 1728719, "ts": 6950381120704.789, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381226471.581, "dur": 9.472 , "args": { "External id": 3994,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14327, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14327, "pid": 0, "tid": 7, "ts": 6950381226471.581, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120721.089, "dur": 1.573 , "args": { "External id": 3994,"cbid": 652, "correlation": 14327 } }, { "ph": "s", "id": 14327, "pid": 1728719, "tid": 1728719, "ts": 6950381120721.089, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120732.110, "dur": 0.100 , "args": { "External id": 3995,"cbid": 200, "correlation": 14342 } }, { "ph": "f", "id": 14342, "pid": 1728719, "tid": 1728719, "ts": 6950381120732.110, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120732.310, "dur": 0.160 , "args": { "External id": 3995,"cbid": 468, "correlation": 14343 } }, { "ph": "f", "id": 14343, "pid": 1728719, "tid": 1728719, "ts": 6950381120732.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120732.731, "dur": 0.020 , "args": { "External id": 3995,"cbid": 1, "correlation": 14345 } }, { "ph": "f", "id": 14345, "pid": 1728719, "tid": 1728719, "ts": 6950381120732.731, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120732.821, "dur": 0.100 , "args": { "External id": 3995,"cbid": 468, "correlation": 14346 } }, { "ph": "f", "id": 14346, "pid": 1728719, "tid": 1728719, "ts": 6950381120732.821, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120733.092, "dur": 0.020 , "args": { "External id": 3995,"cbid": 1, "correlation": 14348 } }, { "ph": "f", "id": 14348, "pid": 1728719, "tid": 1728719, "ts": 6950381120733.092, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120733.212, "dur": 0.100 , "args": { "External id": 3995,"cbid": 468, "correlation": 14349 } }, { "ph": "f", "id": 14349, "pid": 1728719, "tid": 1728719, "ts": 6950381120733.212, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120733.462, "dur": 0.020 , "args": { "External id": 3995,"cbid": 1, "correlation": 14351 } }, { "ph": "f", "id": 14351, "pid": 1728719, "tid": 1728719, "ts": 6950381120733.462, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120733.563, "dur": 0.370 , "args": { "External id": 3995,"cbid": 273, "correlation": 14352 } }, { "ph": "f", "id": 14352, "pid": 1728719, "tid": 1728719, "ts": 6950381120733.563, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381226481.341, "dur": 211.933 , "args": { "External id": 3995,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14355, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14355, "pid": 0, "tid": 7, "ts": 6950381226481.341, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120734.324, "dur": 1.703 , "args": { "External id": 3995,"cbid": 211, "correlation": 14355 } }, { "ph": "s", "id": 14355, "pid": 1728719, "tid": 1728719, "ts": 6950381120734.324, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381226693.562, "dur": 47.392 , "args": { "External id": 4007,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14367, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14367, "pid": 0, "tid": 7, "ts": 6950381226693.562, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120743.842, "dur": 1.763 , "args": { "External id": 4007,"cbid": 307, "correlation": 14367 } }, { "ph": "s", "id": 14367, "pid": 1728719, "tid": 1728719, "ts": 6950381120743.842, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381226741.242, "dur": 13.376 , "args": { "External id": 4008,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14373, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14373, "pid": 0, "tid": 7, "ts": 6950381226741.242, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120751.937, "dur": 1.523 , "args": { "External id": 4008,"cbid": 652, "correlation": 14373 } }, { "ph": "s", "id": 14373, "pid": 1728719, "tid": 1728719, "ts": 6950381120751.937, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381226754.938, "dur": 13.056 , "args": { "External id": 4009,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14379, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14379, "pid": 0, "tid": 7, "ts": 6950381226754.938, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120760.613, "dur": 1.553 , "args": { "External id": 4009,"cbid": 652, "correlation": 14379 } }, { "ph": "s", "id": 14379, "pid": 1728719, "tid": 1728719, "ts": 6950381120760.613, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120772.275, "dur": 0.100 , "args": { "External id": 4010,"cbid": 200, "correlation": 14398 } }, { "ph": "f", "id": 14398, "pid": 1728719, "tid": 1728719, "ts": 6950381120772.275, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120772.465, "dur": 0.130 , "args": { "External id": 4010,"cbid": 468, "correlation": 14399 } }, { "ph": "f", "id": 14399, "pid": 1728719, "tid": 1728719, "ts": 6950381120772.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120772.956, "dur": 0.020 , "args": { "External id": 4010,"cbid": 1, "correlation": 14401 } }, { "ph": "f", "id": 14401, "pid": 1728719, "tid": 1728719, "ts": 6950381120772.956, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120773.126, "dur": 0.100 , "args": { "External id": 4010,"cbid": 468, "correlation": 14402 } }, { "ph": "f", "id": 14402, "pid": 1728719, "tid": 1728719, "ts": 6950381120773.126, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120773.387, "dur": 0.030 , "args": { "External id": 4010,"cbid": 1, "correlation": 14404 } }, { "ph": "f", "id": 14404, "pid": 1728719, "tid": 1728719, "ts": 6950381120773.387, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120773.517, "dur": 0.100 , "args": { "External id": 4010,"cbid": 468, "correlation": 14405 } }, { "ph": "f", "id": 14405, "pid": 1728719, "tid": 1728719, "ts": 6950381120773.517, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120773.767, "dur": 0.020 , "args": { "External id": 4010,"cbid": 1, "correlation": 14407 } }, { "ph": "f", "id": 14407, "pid": 1728719, "tid": 1728719, "ts": 6950381120773.767, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120773.878, "dur": 0.370 , "args": { "External id": 4010,"cbid": 273, "correlation": 14408 } }, { "ph": "f", "id": 14408, "pid": 1728719, "tid": 1728719, "ts": 6950381120773.878, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381226768.282, "dur": 677.176 , "args": { "External id": 4010,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14411, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14411, "pid": 0, "tid": 7, "ts": 6950381226768.282, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120774.619, "dur": 1.713 , "args": { "External id": 4010,"cbid": 211, "correlation": 14411 } }, { "ph": "s", "id": 14411, "pid": 1728719, "tid": 1728719, "ts": 6950381120774.619, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120786.431, "dur": 0.090 , "args": { "External id": 4022,"cbid": 200, "correlation": 14435 } }, { "ph": "f", "id": 14435, "pid": 1728719, "tid": 1728719, "ts": 6950381120786.431, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120786.611, "dur": 0.110 , "args": { "External id": 4022,"cbid": 468, "correlation": 14436 } }, { "ph": "f", "id": 14436, "pid": 1728719, "tid": 1728719, "ts": 6950381120786.611, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120786.922, "dur": 0.020 , "args": { "External id": 4022,"cbid": 1, "correlation": 14438 } }, { "ph": "f", "id": 14438, "pid": 1728719, "tid": 1728719, "ts": 6950381120786.922, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120787.012, "dur": 0.120 , "args": { "External id": 4022,"cbid": 468, "correlation": 14439 } }, { "ph": "f", "id": 14439, "pid": 1728719, "tid": 1728719, "ts": 6950381120787.012, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120787.283, "dur": 0.020 , "args": { "External id": 4022,"cbid": 1, "correlation": 14441 } }, { "ph": "f", "id": 14441, "pid": 1728719, "tid": 1728719, "ts": 6950381120787.283, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120787.383, "dur": 0.100 , "args": { "External id": 4022,"cbid": 468, "correlation": 14442 } }, { "ph": "f", "id": 14442, "pid": 1728719, "tid": 1728719, "ts": 6950381120787.383, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120787.643, "dur": 0.020 , "args": { "External id": 4022,"cbid": 1, "correlation": 14444 } }, { "ph": "f", "id": 14444, "pid": 1728719, "tid": 1728719, "ts": 6950381120787.643, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120787.743, "dur": 0.321 , "args": { "External id": 4022,"cbid": 273, "correlation": 14445 } }, { "ph": "f", "id": 14445, "pid": 1728719, "tid": 1728719, "ts": 6950381120787.743, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381227445.778, "dur": 692.121 , "args": { "External id": 4022,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14448, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14448, "pid": 0, "tid": 7, "ts": 6950381227445.778, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120788.415, "dur": 1.643 , "args": { "External id": 4022,"cbid": 211, "correlation": 14448 } }, { "ph": "s", "id": 14448, "pid": 1728719, "tid": 1728719, "ts": 6950381120788.415, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381228138.219, "dur": 153.342 , "args": { "External id": 4034,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14461, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14461, "pid": 0, "tid": 7, "ts": 6950381228138.219, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120797.552, "dur": 1.603 , "args": { "External id": 4034,"cbid": 307, "correlation": 14461 } }, { "ph": "s", "id": 14461, "pid": 1728719, "tid": 1728719, "ts": 6950381120797.552, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381228291.817, "dur": 57.951 , "args": { "External id": 4035,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14467, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14467, "pid": 0, "tid": 7, "ts": 6950381228291.817, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120805.937, "dur": 1.583 , "args": { "External id": 4035,"cbid": 652, "correlation": 14467 } }, { "ph": "s", "id": 14467, "pid": 1728719, "tid": 1728719, "ts": 6950381120805.937, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120817.158, "dur": 0.080 , "args": { "External id": 4036,"cbid": 200, "correlation": 14486 } }, { "ph": "f", "id": 14486, "pid": 1728719, "tid": 1728719, "ts": 6950381120817.158, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120817.328, "dur": 0.111 , "args": { "External id": 4036,"cbid": 468, "correlation": 14487 } }, { "ph": "f", "id": 14487, "pid": 1728719, "tid": 1728719, "ts": 6950381120817.328, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120817.639, "dur": 0.020 , "args": { "External id": 4036,"cbid": 1, "correlation": 14489 } }, { "ph": "f", "id": 14489, "pid": 1728719, "tid": 1728719, "ts": 6950381120817.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120817.729, "dur": 0.090 , "args": { "External id": 4036,"cbid": 468, "correlation": 14490 } }, { "ph": "f", "id": 14490, "pid": 1728719, "tid": 1728719, "ts": 6950381120817.729, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120817.970, "dur": 0.020 , "args": { "External id": 4036,"cbid": 1, "correlation": 14492 } }, { "ph": "f", "id": 14492, "pid": 1728719, "tid": 1728719, "ts": 6950381120817.970, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120818.070, "dur": 0.100 , "args": { "External id": 4036,"cbid": 468, "correlation": 14493 } }, { "ph": "f", "id": 14493, "pid": 1728719, "tid": 1728719, "ts": 6950381120818.070, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120818.320, "dur": 0.020 , "args": { "External id": 4036,"cbid": 1, "correlation": 14495 } }, { "ph": "f", "id": 14495, "pid": 1728719, "tid": 1728719, "ts": 6950381120818.320, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120818.420, "dur": 0.381 , "args": { "External id": 4036,"cbid": 273, "correlation": 14496 } }, { "ph": "f", "id": 14496, "pid": 1728719, "tid": 1728719, "ts": 6950381120818.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381228350.057, "dur": 690.936 , "args": { "External id": 4036,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14499, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14499, "pid": 0, "tid": 7, "ts": 6950381228350.057, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120819.172, "dur": 1.733 , "args": { "External id": 4036,"cbid": 211, "correlation": 14499 } }, { "ph": "s", "id": 14499, "pid": 1728719, "tid": 1728719, "ts": 6950381120819.172, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381229041.281, "dur": 40.160 , "args": { "External id": 4048,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14510, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14510, "pid": 0, "tid": 7, "ts": 6950381229041.281, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120828.970, "dur": 1.864 , "args": { "External id": 4048,"cbid": 307, "correlation": 14510 } }, { "ph": "s", "id": 14510, "pid": 1728719, "tid": 1728719, "ts": 6950381120828.970, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381229081.697, "dur": 21.823 , "args": { "External id": 4060,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14522, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14522, "pid": 0, "tid": 7, "ts": 6950381229081.697, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120907.476, "dur": 2.495 , "args": { "External id": 4060,"cbid": 307, "correlation": 14522 } }, { "ph": "s", "id": 14522, "pid": 1728719, "tid": 1728719, "ts": 6950381120907.476, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381229103.840, "dur": 13.344 , "args": { "External id": 4061,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14536, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14536, "pid": 0, "tid": 7, "ts": 6950381229103.840, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120921.643, "dur": 1.643 , "args": { "External id": 4061,"cbid": 652, "correlation": 14536 } }, { "ph": "s", "id": 14536, "pid": 1728719, "tid": 1728719, "ts": 6950381120921.643, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381229117.504, "dur": 12.288 , "args": { "External id": 4062,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14550, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14550, "pid": 0, "tid": 7, "ts": 6950381229117.504, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120933.364, "dur": 1.583 , "args": { "External id": 4062,"cbid": 652, "correlation": 14550 } }, { "ph": "s", "id": 14550, "pid": 1728719, "tid": 1728719, "ts": 6950381120933.364, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381229130.080, "dur": 10.720 , "args": { "External id": 4063,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14564, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14564, "pid": 0, "tid": 7, "ts": 6950381229130.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381120944.385, "dur": 1.533 , "args": { "External id": 4063,"cbid": 652, "correlation": 14564 } }, { "ph": "s", "id": 14564, "pid": 1728719, "tid": 1728719, "ts": 6950381120944.385, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120955.766, "dur": 0.160 , "args": { "External id": 4064,"cbid": 200, "correlation": 14579 } }, { "ph": "f", "id": 14579, "pid": 1728719, "tid": 1728719, "ts": 6950381120955.766, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120956.077, "dur": 0.200 , "args": { "External id": 4064,"cbid": 468, "correlation": 14580 } }, { "ph": "f", "id": 14580, "pid": 1728719, "tid": 1728719, "ts": 6950381120956.077, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120956.608, "dur": 0.030 , "args": { "External id": 4064,"cbid": 1, "correlation": 14582 } }, { "ph": "f", "id": 14582, "pid": 1728719, "tid": 1728719, "ts": 6950381120956.608, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120956.718, "dur": 0.090 , "args": { "External id": 4064,"cbid": 468, "correlation": 14583 } }, { "ph": "f", "id": 14583, "pid": 1728719, "tid": 1728719, "ts": 6950381120956.718, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120956.958, "dur": 0.030 , "args": { "External id": 4064,"cbid": 1, "correlation": 14585 } }, { "ph": "f", "id": 14585, "pid": 1728719, "tid": 1728719, "ts": 6950381120956.958, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120957.089, "dur": 0.090 , "args": { "External id": 4064,"cbid": 468, "correlation": 14586 } }, { "ph": "f", "id": 14586, "pid": 1728719, "tid": 1728719, "ts": 6950381120957.089, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120957.349, "dur": 0.020 , "args": { "External id": 4064,"cbid": 1, "correlation": 14588 } }, { "ph": "f", "id": 14588, "pid": 1728719, "tid": 1728719, "ts": 6950381120957.349, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120957.509, "dur": 0.672 , "args": { "External id": 4064,"cbid": 273, "correlation": 14589 } }, { "ph": "f", "id": 14589, "pid": 1728719, "tid": 1728719, "ts": 6950381120957.509, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381229141.088, "dur": 223.038 , "args": { "External id": 4064,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14592, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14592, "pid": 0, "tid": 7, "ts": 6950381229141.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120958.642, "dur": 1.873 , "args": { "External id": 4064,"cbid": 211, "correlation": 14592 } }, { "ph": "s", "id": 14592, "pid": 1728719, "tid": 1728719, "ts": 6950381120958.642, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381229364.414, "dur": 8.799 , "args": { "External id": 4076,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14605, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14605, "pid": 0, "tid": 7, "ts": 6950381229364.414, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120969.341, "dur": 1.633 , "args": { "External id": 4076,"cbid": 307, "correlation": 14605 } }, { "ph": "s", "id": 14605, "pid": 1728719, "tid": 1728719, "ts": 6950381120969.341, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381229373.501, "dur": 36.672 , "args": { "External id": 4081,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14628, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14628, "pid": 0, "tid": 7, "ts": 6950381229373.501, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381120988.407, "dur": 1.663 , "args": { "External id": 4081,"cbid": 307, "correlation": 14628 } }, { "ph": "s", "id": 14628, "pid": 1728719, "tid": 1728719, "ts": 6950381120988.407, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381120998.616, "dur": 0.100 , "args": { "External id": 4082,"cbid": 200, "correlation": 14643 } }, { "ph": "f", "id": 14643, "pid": 1728719, "tid": 1728719, "ts": 6950381120998.616, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120998.816, "dur": 0.131 , "args": { "External id": 4082,"cbid": 468, "correlation": 14644 } }, { "ph": "f", "id": 14644, "pid": 1728719, "tid": 1728719, "ts": 6950381120998.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120999.177, "dur": 0.020 , "args": { "External id": 4082,"cbid": 1, "correlation": 14646 } }, { "ph": "f", "id": 14646, "pid": 1728719, "tid": 1728719, "ts": 6950381120999.177, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120999.277, "dur": 0.100 , "args": { "External id": 4082,"cbid": 468, "correlation": 14647 } }, { "ph": "f", "id": 14647, "pid": 1728719, "tid": 1728719, "ts": 6950381120999.277, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120999.538, "dur": 0.020 , "args": { "External id": 4082,"cbid": 1, "correlation": 14649 } }, { "ph": "f", "id": 14649, "pid": 1728719, "tid": 1728719, "ts": 6950381120999.538, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120999.658, "dur": 0.090 , "args": { "External id": 4082,"cbid": 468, "correlation": 14650 } }, { "ph": "f", "id": 14650, "pid": 1728719, "tid": 1728719, "ts": 6950381120999.658, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381120999.918, "dur": 0.020 , "args": { "External id": 4082,"cbid": 1, "correlation": 14652 } }, { "ph": "f", "id": 14652, "pid": 1728719, "tid": 1728719, "ts": 6950381120999.918, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121000.029, "dur": 0.390 , "args": { "External id": 4082,"cbid": 273, "correlation": 14653 } }, { "ph": "f", "id": 14653, "pid": 1728719, "tid": 1728719, "ts": 6950381121000.029, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381229410.461, "dur": 233.565 , "args": { "External id": 4082,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14656, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14656, "pid": 0, "tid": 7, "ts": 6950381229410.461, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121000.820, "dur": 1.833 , "args": { "External id": 4082,"cbid": 211, "correlation": 14656 } }, { "ph": "s", "id": 14656, "pid": 1728719, "tid": 1728719, "ts": 6950381121000.820, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381229644.347, "dur": 8.831 , "args": { "External id": 4094,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14665, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14665, "pid": 0, "tid": 7, "ts": 6950381229644.347, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121008.444, "dur": 1.593 , "args": { "External id": 4094,"cbid": 307, "correlation": 14665 } }, { "ph": "s", "id": 14665, "pid": 1728719, "tid": 1728719, "ts": 6950381121008.444, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381229653.434, "dur": 28.448 , "args": { "External id": 4099,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14688, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14688, "pid": 0, "tid": 7, "ts": 6950381229653.434, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121027.600, "dur": 1.563 , "args": { "External id": 4099,"cbid": 307, "correlation": 14688 } }, { "ph": "s", "id": 14688, "pid": 1728719, "tid": 1728719, "ts": 6950381121027.600, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121037.619, "dur": 0.090 , "args": { "External id": 4100,"cbid": 200, "correlation": 14703 } }, { "ph": "f", "id": 14703, "pid": 1728719, "tid": 1728719, "ts": 6950381121037.619, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121037.799, "dur": 0.130 , "args": { "External id": 4100,"cbid": 468, "correlation": 14704 } }, { "ph": "f", "id": 14704, "pid": 1728719, "tid": 1728719, "ts": 6950381121037.799, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121038.139, "dur": 0.031 , "args": { "External id": 4100,"cbid": 1, "correlation": 14706 } }, { "ph": "f", "id": 14706, "pid": 1728719, "tid": 1728719, "ts": 6950381121038.139, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121038.240, "dur": 0.090 , "args": { "External id": 4100,"cbid": 468, "correlation": 14707 } }, { "ph": "f", "id": 14707, "pid": 1728719, "tid": 1728719, "ts": 6950381121038.240, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121038.480, "dur": 0.020 , "args": { "External id": 4100,"cbid": 1, "correlation": 14709 } }, { "ph": "f", "id": 14709, "pid": 1728719, "tid": 1728719, "ts": 6950381121038.480, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121038.600, "dur": 0.091 , "args": { "External id": 4100,"cbid": 468, "correlation": 14710 } }, { "ph": "f", "id": 14710, "pid": 1728719, "tid": 1728719, "ts": 6950381121038.600, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121038.851, "dur": 0.030 , "args": { "External id": 4100,"cbid": 1, "correlation": 14712 } }, { "ph": "f", "id": 14712, "pid": 1728719, "tid": 1728719, "ts": 6950381121038.851, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121038.961, "dur": 0.361 , "args": { "External id": 4100,"cbid": 273, "correlation": 14713 } }, { "ph": "f", "id": 14713, "pid": 1728719, "tid": 1728719, "ts": 6950381121038.961, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381229682.138, "dur": 235.934 , "args": { "External id": 4100,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14716, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14716, "pid": 0, "tid": 7, "ts": 6950381229682.138, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121039.702, "dur": 1.684 , "args": { "External id": 4100,"cbid": 211, "correlation": 14716 } }, { "ph": "s", "id": 14716, "pid": 1728719, "tid": 1728719, "ts": 6950381121039.702, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381229918.360, "dur": 59.295 , "args": { "External id": 4112,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14730, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14730, "pid": 0, "tid": 7, "ts": 6950381229918.360, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121048.940, "dur": 1.562 , "args": { "External id": 4112,"cbid": 307, "correlation": 14730 } }, { "ph": "s", "id": 14730, "pid": 1728719, "tid": 1728719, "ts": 6950381121048.940, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381229977.943, "dur": 64.607 , "args": { "External id": 4113,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14740, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14740, "pid": 0, "tid": 7, "ts": 6950381229977.943, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121058.788, "dur": 1.663 , "args": { "External id": 4113,"cbid": 307, "correlation": 14740 } }, { "ph": "s", "id": 14740, "pid": 1728719, "tid": 1728719, "ts": 6950381121058.788, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381230042.806, "dur": 39.840 , "args": { "External id": 4114,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14749, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14749, "pid": 0, "tid": 7, "ts": 6950381230042.806, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121068.406, "dur": 1.623 , "args": { "External id": 4114,"cbid": 307, "correlation": 14749 } }, { "ph": "s", "id": 14749, "pid": 1728719, "tid": 1728719, "ts": 6950381121068.406, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121083.624, "dur": 0.110 , "args": { "External id": 4122,"cbid": 200, "correlation": 14784 } }, { "ph": "f", "id": 14784, "pid": 1728719, "tid": 1728719, "ts": 6950381121083.624, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121083.804, "dur": 0.151 , "args": { "External id": 4122,"cbid": 200, "correlation": 14785 } }, { "ph": "f", "id": 14785, "pid": 1728719, "tid": 1728719, "ts": 6950381121083.804, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121084.055, "dur": 0.451 , "args": { "External id": 4122,"cbid": 273, "correlation": 14786 } }, { "ph": "f", "id": 14786, "pid": 1728719, "tid": 1728719, "ts": 6950381121084.055, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381230082.902, "dur": 857.751 , "args": { "External id": 4122,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14788, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14788, "pid": 0, "tid": 7, "ts": 6950381230082.902, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121084.786, "dur": 1.864 , "args": { "External id": 4122,"cbid": 211, "correlation": 14788 } }, { "ph": "s", "id": 14788, "pid": 1728719, "tid": 1728719, "ts": 6950381121084.786, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381230940.941, "dur": 9.247 , "args": { "External id": 4131,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14808, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14808, "pid": 0, "tid": 7, "ts": 6950381230940.941, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121101.077, "dur": 1.703 , "args": { "External id": 4131,"cbid": 652, "correlation": 14808 } }, { "ph": "s", "id": 14808, "pid": 1728719, "tid": 1728719, "ts": 6950381121101.077, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121111.977, "dur": 0.100 , "args": { "External id": 4132,"cbid": 200, "correlation": 14823 } }, { "ph": "f", "id": 14823, "pid": 1728719, "tid": 1728719, "ts": 6950381121111.977, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121112.167, "dur": 0.141 , "args": { "External id": 4132,"cbid": 468, "correlation": 14824 } }, { "ph": "f", "id": 14824, "pid": 1728719, "tid": 1728719, "ts": 6950381121112.167, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121112.588, "dur": 0.020 , "args": { "External id": 4132,"cbid": 1, "correlation": 14826 } }, { "ph": "f", "id": 14826, "pid": 1728719, "tid": 1728719, "ts": 6950381121112.588, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121112.668, "dur": 0.100 , "args": { "External id": 4132,"cbid": 468, "correlation": 14827 } }, { "ph": "f", "id": 14827, "pid": 1728719, "tid": 1728719, "ts": 6950381121112.668, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121112.919, "dur": 0.030 , "args": { "External id": 4132,"cbid": 1, "correlation": 14829 } }, { "ph": "f", "id": 14829, "pid": 1728719, "tid": 1728719, "ts": 6950381121112.919, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121113.039, "dur": 0.100 , "args": { "External id": 4132,"cbid": 468, "correlation": 14830 } }, { "ph": "f", "id": 14830, "pid": 1728719, "tid": 1728719, "ts": 6950381121113.039, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121113.309, "dur": 0.020 , "args": { "External id": 4132,"cbid": 1, "correlation": 14832 } }, { "ph": "f", "id": 14832, "pid": 1728719, "tid": 1728719, "ts": 6950381121113.309, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121113.420, "dur": 0.410 , "args": { "External id": 4132,"cbid": 273, "correlation": 14833 } }, { "ph": "f", "id": 14833, "pid": 1728719, "tid": 1728719, "ts": 6950381121113.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381230950.444, "dur": 212.446 , "args": { "External id": 4132,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14836, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14836, "pid": 0, "tid": 7, "ts": 6950381230950.444, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121114.211, "dur": 1.753 , "args": { "External id": 4132,"cbid": 211, "correlation": 14836 } }, { "ph": "s", "id": 14836, "pid": 1728719, "tid": 1728719, "ts": 6950381121114.211, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381231163.178, "dur": 46.528 , "args": { "External id": 4144,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14848, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14848, "pid": 0, "tid": 7, "ts": 6950381231163.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121123.849, "dur": 1.683 , "args": { "External id": 4144,"cbid": 307, "correlation": 14848 } }, { "ph": "s", "id": 14848, "pid": 1728719, "tid": 1728719, "ts": 6950381121123.849, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381231210.026, "dur": 13.567 , "args": { "External id": 4145,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14854, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14854, "pid": 0, "tid": 7, "ts": 6950381231210.026, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121131.644, "dur": 1.713 , "args": { "External id": 4145,"cbid": 652, "correlation": 14854 } }, { "ph": "s", "id": 14854, "pid": 1728719, "tid": 1728719, "ts": 6950381121131.644, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381231223.914, "dur": 13.503 , "args": { "External id": 4146,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14860, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14860, "pid": 0, "tid": 7, "ts": 6950381231223.914, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121140.159, "dur": 1.513 , "args": { "External id": 4146,"cbid": 652, "correlation": 14860 } }, { "ph": "s", "id": 14860, "pid": 1728719, "tid": 1728719, "ts": 6950381121140.159, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121151.631, "dur": 0.090 , "args": { "External id": 4147,"cbid": 200, "correlation": 14879 } }, { "ph": "f", "id": 14879, "pid": 1728719, "tid": 1728719, "ts": 6950381121151.631, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121151.821, "dur": 0.110 , "args": { "External id": 4147,"cbid": 468, "correlation": 14880 } }, { "ph": "f", "id": 14880, "pid": 1728719, "tid": 1728719, "ts": 6950381121151.821, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121152.162, "dur": 0.020 , "args": { "External id": 4147,"cbid": 1, "correlation": 14882 } }, { "ph": "f", "id": 14882, "pid": 1728719, "tid": 1728719, "ts": 6950381121152.162, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121152.262, "dur": 0.110 , "args": { "External id": 4147,"cbid": 468, "correlation": 14883 } }, { "ph": "f", "id": 14883, "pid": 1728719, "tid": 1728719, "ts": 6950381121152.262, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121152.512, "dur": 0.040 , "args": { "External id": 4147,"cbid": 1, "correlation": 14885 } }, { "ph": "f", "id": 14885, "pid": 1728719, "tid": 1728719, "ts": 6950381121152.512, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121152.643, "dur": 0.100 , "args": { "External id": 4147,"cbid": 468, "correlation": 14886 } }, { "ph": "f", "id": 14886, "pid": 1728719, "tid": 1728719, "ts": 6950381121152.643, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121152.903, "dur": 0.010 , "args": { "External id": 4147,"cbid": 1, "correlation": 14888 } }, { "ph": "f", "id": 14888, "pid": 1728719, "tid": 1728719, "ts": 6950381121152.903, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121153.013, "dur": 0.371 , "args": { "External id": 4147,"cbid": 273, "correlation": 14889 } }, { "ph": "f", "id": 14889, "pid": 1728719, "tid": 1728719, "ts": 6950381121153.013, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381231237.737, "dur": 679.609 , "args": { "External id": 4147,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14892, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14892, "pid": 0, "tid": 7, "ts": 6950381231237.737, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121153.745, "dur": 1.683 , "args": { "External id": 4147,"cbid": 211, "correlation": 14892 } }, { "ph": "s", "id": 14892, "pid": 1728719, "tid": 1728719, "ts": 6950381121153.745, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121165.356, "dur": 0.090 , "args": { "External id": 4159,"cbid": 200, "correlation": 14916 } }, { "ph": "f", "id": 14916, "pid": 1728719, "tid": 1728719, "ts": 6950381121165.356, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121165.537, "dur": 0.110 , "args": { "External id": 4159,"cbid": 468, "correlation": 14917 } }, { "ph": "f", "id": 14917, "pid": 1728719, "tid": 1728719, "ts": 6950381121165.537, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121165.847, "dur": 0.020 , "args": { "External id": 4159,"cbid": 1, "correlation": 14919 } }, { "ph": "f", "id": 14919, "pid": 1728719, "tid": 1728719, "ts": 6950381121165.847, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121165.937, "dur": 0.091 , "args": { "External id": 4159,"cbid": 468, "correlation": 14920 } }, { "ph": "f", "id": 14920, "pid": 1728719, "tid": 1728719, "ts": 6950381121165.937, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121166.178, "dur": 0.020 , "args": { "External id": 4159,"cbid": 1, "correlation": 14922 } }, { "ph": "f", "id": 14922, "pid": 1728719, "tid": 1728719, "ts": 6950381121166.178, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121166.298, "dur": 0.100 , "args": { "External id": 4159,"cbid": 468, "correlation": 14923 } }, { "ph": "f", "id": 14923, "pid": 1728719, "tid": 1728719, "ts": 6950381121166.298, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121166.559, "dur": 0.020 , "args": { "External id": 4159,"cbid": 1, "correlation": 14925 } }, { "ph": "f", "id": 14925, "pid": 1728719, "tid": 1728719, "ts": 6950381121166.559, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121166.659, "dur": 0.340 , "args": { "External id": 4159,"cbid": 273, "correlation": 14926 } }, { "ph": "f", "id": 14926, "pid": 1728719, "tid": 1728719, "ts": 6950381121166.659, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381231917.666, "dur": 695.384 , "args": { "External id": 4159,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14929, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14929, "pid": 0, "tid": 7, "ts": 6950381231917.666, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121167.350, "dur": 1.663 , "args": { "External id": 4159,"cbid": 211, "correlation": 14929 } }, { "ph": "s", "id": 14929, "pid": 1728719, "tid": 1728719, "ts": 6950381121167.350, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381232613.370, "dur": 153.854 , "args": { "External id": 4171,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14942, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14942, "pid": 0, "tid": 7, "ts": 6950381232613.370, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121176.497, "dur": 1.633 , "args": { "External id": 4171,"cbid": 307, "correlation": 14942 } }, { "ph": "s", "id": 14942, "pid": 1728719, "tid": 1728719, "ts": 6950381121176.497, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381232767.481, "dur": 57.247 , "args": { "External id": 4172,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14948, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14948, "pid": 0, "tid": 7, "ts": 6950381232767.481, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121185.163, "dur": 1.623 , "args": { "External id": 4172,"cbid": 652, "correlation": 14948 } }, { "ph": "s", "id": 14948, "pid": 1728719, "tid": 1728719, "ts": 6950381121185.163, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121196.464, "dur": 0.090 , "args": { "External id": 4173,"cbid": 200, "correlation": 14967 } }, { "ph": "f", "id": 14967, "pid": 1728719, "tid": 1728719, "ts": 6950381121196.464, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121196.645, "dur": 0.120 , "args": { "External id": 4173,"cbid": 468, "correlation": 14968 } }, { "ph": "f", "id": 14968, "pid": 1728719, "tid": 1728719, "ts": 6950381121196.645, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121196.955, "dur": 0.030 , "args": { "External id": 4173,"cbid": 1, "correlation": 14970 } }, { "ph": "f", "id": 14970, "pid": 1728719, "tid": 1728719, "ts": 6950381121196.955, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121197.055, "dur": 0.100 , "args": { "External id": 4173,"cbid": 468, "correlation": 14971 } }, { "ph": "f", "id": 14971, "pid": 1728719, "tid": 1728719, "ts": 6950381121197.055, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121197.296, "dur": 0.020 , "args": { "External id": 4173,"cbid": 1, "correlation": 14973 } }, { "ph": "f", "id": 14973, "pid": 1728719, "tid": 1728719, "ts": 6950381121197.296, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121197.416, "dur": 0.090 , "args": { "External id": 4173,"cbid": 468, "correlation": 14974 } }, { "ph": "f", "id": 14974, "pid": 1728719, "tid": 1728719, "ts": 6950381121197.416, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121197.666, "dur": 0.020 , "args": { "External id": 4173,"cbid": 1, "correlation": 14976 } }, { "ph": "f", "id": 14976, "pid": 1728719, "tid": 1728719, "ts": 6950381121197.666, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121197.767, "dur": 0.370 , "args": { "External id": 4173,"cbid": 273, "correlation": 14977 } }, { "ph": "f", "id": 14977, "pid": 1728719, "tid": 1728719, "ts": 6950381121197.767, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381232825.016, "dur": 693.657 , "args": { "External id": 4173,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14980, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14980, "pid": 0, "tid": 7, "ts": 6950381232825.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121198.498, "dur": 1.693 , "args": { "External id": 4173,"cbid": 211, "correlation": 14980 } }, { "ph": "s", "id": 14980, "pid": 1728719, "tid": 1728719, "ts": 6950381121198.498, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381233518.961, "dur": 39.807 , "args": { "External id": 4185,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 14991, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 14991, "pid": 0, "tid": 7, "ts": 6950381233518.961, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121208.176, "dur": 1.653 , "args": { "External id": 4185,"cbid": 307, "correlation": 14991 } }, { "ph": "s", "id": 14991, "pid": 1728719, "tid": 1728719, "ts": 6950381121208.176, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381233559.024, "dur": 21.536 , "args": { "External id": 4197,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15003, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15003, "pid": 0, "tid": 7, "ts": 6950381233559.024, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121284.137, "dur": 2.515 , "args": { "External id": 4197,"cbid": 307, "correlation": 15003 } }, { "ph": "s", "id": 15003, "pid": 1728719, "tid": 1728719, "ts": 6950381121284.137, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381233580.880, "dur": 13.248 , "args": { "External id": 4198,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15017, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15017, "pid": 0, "tid": 7, "ts": 6950381233580.880, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121298.073, "dur": 1.734 , "args": { "External id": 4198,"cbid": 652, "correlation": 15017 } }, { "ph": "s", "id": 15017, "pid": 1728719, "tid": 1728719, "ts": 6950381121298.073, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381233594.448, "dur": 12.832 , "args": { "External id": 4199,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15031, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15031, "pid": 0, "tid": 7, "ts": 6950381233594.448, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121309.735, "dur": 1.733 , "args": { "External id": 4199,"cbid": 652, "correlation": 15031 } }, { "ph": "s", "id": 15031, "pid": 1728719, "tid": 1728719, "ts": 6950381121309.735, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381233607.599, "dur": 10.912 , "args": { "External id": 4200,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15045, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15045, "pid": 0, "tid": 7, "ts": 6950381233607.599, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121320.415, "dur": 1.653 , "args": { "External id": 4200,"cbid": 652, "correlation": 15045 } }, { "ph": "s", "id": 15045, "pid": 1728719, "tid": 1728719, "ts": 6950381121320.415, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121332.247, "dur": 0.140 , "args": { "External id": 4201,"cbid": 200, "correlation": 15060 } }, { "ph": "f", "id": 15060, "pid": 1728719, "tid": 1728719, "ts": 6950381121332.247, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121332.527, "dur": 0.191 , "args": { "External id": 4201,"cbid": 468, "correlation": 15061 } }, { "ph": "f", "id": 15061, "pid": 1728719, "tid": 1728719, "ts": 6950381121332.527, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121333.038, "dur": 0.030 , "args": { "External id": 4201,"cbid": 1, "correlation": 15063 } }, { "ph": "f", "id": 15063, "pid": 1728719, "tid": 1728719, "ts": 6950381121333.038, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121333.149, "dur": 0.100 , "args": { "External id": 4201,"cbid": 468, "correlation": 15064 } }, { "ph": "f", "id": 15064, "pid": 1728719, "tid": 1728719, "ts": 6950381121333.149, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121333.509, "dur": 0.020 , "args": { "External id": 4201,"cbid": 1, "correlation": 15066 } }, { "ph": "f", "id": 15066, "pid": 1728719, "tid": 1728719, "ts": 6950381121333.509, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121333.619, "dur": 0.101 , "args": { "External id": 4201,"cbid": 468, "correlation": 15067 } }, { "ph": "f", "id": 15067, "pid": 1728719, "tid": 1728719, "ts": 6950381121333.619, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121333.890, "dur": 0.020 , "args": { "External id": 4201,"cbid": 1, "correlation": 15069 } }, { "ph": "f", "id": 15069, "pid": 1728719, "tid": 1728719, "ts": 6950381121333.890, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121334.040, "dur": 0.621 , "args": { "External id": 4201,"cbid": 273, "correlation": 15070 } }, { "ph": "f", "id": 15070, "pid": 1728719, "tid": 1728719, "ts": 6950381121334.040, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381233618.799, "dur": 220.510 , "args": { "External id": 4201,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15073, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15073, "pid": 0, "tid": 7, "ts": 6950381233618.799, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121335.112, "dur": 1.834 , "args": { "External id": 4201,"cbid": 211, "correlation": 15073 } }, { "ph": "s", "id": 15073, "pid": 1728719, "tid": 1728719, "ts": 6950381121335.112, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381233839.597, "dur": 8.896 , "args": { "External id": 4213,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15086, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15086, "pid": 0, "tid": 7, "ts": 6950381233839.597, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121345.792, "dur": 1.723 , "args": { "External id": 4213,"cbid": 307, "correlation": 15086 } }, { "ph": "s", "id": 15086, "pid": 1728719, "tid": 1728719, "ts": 6950381121345.792, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381233848.781, "dur": 37.727 , "args": { "External id": 4218,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15109, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15109, "pid": 0, "tid": 7, "ts": 6950381233848.781, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121364.788, "dur": 1.913 , "args": { "External id": 4218,"cbid": 307, "correlation": 15109 } }, { "ph": "s", "id": 15109, "pid": 1728719, "tid": 1728719, "ts": 6950381121364.788, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121375.247, "dur": 0.090 , "args": { "External id": 4219,"cbid": 200, "correlation": 15124 } }, { "ph": "f", "id": 15124, "pid": 1728719, "tid": 1728719, "ts": 6950381121375.247, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121375.437, "dur": 0.151 , "args": { "External id": 4219,"cbid": 468, "correlation": 15125 } }, { "ph": "f", "id": 15125, "pid": 1728719, "tid": 1728719, "ts": 6950381121375.437, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121375.838, "dur": 0.040 , "args": { "External id": 4219,"cbid": 1, "correlation": 15127 } }, { "ph": "f", "id": 15127, "pid": 1728719, "tid": 1728719, "ts": 6950381121375.838, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121375.948, "dur": 0.100 , "args": { "External id": 4219,"cbid": 468, "correlation": 15128 } }, { "ph": "f", "id": 15128, "pid": 1728719, "tid": 1728719, "ts": 6950381121375.948, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121376.199, "dur": 0.020 , "args": { "External id": 4219,"cbid": 1, "correlation": 15130 } }, { "ph": "f", "id": 15130, "pid": 1728719, "tid": 1728719, "ts": 6950381121376.199, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121376.309, "dur": 0.100 , "args": { "External id": 4219,"cbid": 468, "correlation": 15131 } }, { "ph": "f", "id": 15131, "pid": 1728719, "tid": 1728719, "ts": 6950381121376.309, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121376.559, "dur": 0.020 , "args": { "External id": 4219,"cbid": 1, "correlation": 15133 } }, { "ph": "f", "id": 15133, "pid": 1728719, "tid": 1728719, "ts": 6950381121376.559, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121376.670, "dur": 0.380 , "args": { "External id": 4219,"cbid": 273, "correlation": 15134 } }, { "ph": "f", "id": 15134, "pid": 1728719, "tid": 1728719, "ts": 6950381121376.670, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381233886.764, "dur": 234.174 , "args": { "External id": 4219,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15137, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15137, "pid": 0, "tid": 7, "ts": 6950381233886.764, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121377.451, "dur": 1.874 , "args": { "External id": 4219,"cbid": 211, "correlation": 15137 } }, { "ph": "s", "id": 15137, "pid": 1728719, "tid": 1728719, "ts": 6950381121377.451, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381234121.258, "dur": 8.896 , "args": { "External id": 4231,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15146, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15146, "pid": 0, "tid": 7, "ts": 6950381234121.258, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121385.366, "dur": 1.743 , "args": { "External id": 4231,"cbid": 307, "correlation": 15146 } }, { "ph": "s", "id": 15146, "pid": 1728719, "tid": 1728719, "ts": 6950381121385.366, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381234130.442, "dur": 25.247 , "args": { "External id": 4236,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15169, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15169, "pid": 0, "tid": 7, "ts": 6950381234130.442, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121401.656, "dur": 1.563 , "args": { "External id": 4236,"cbid": 307, "correlation": 15169 } }, { "ph": "s", "id": 15169, "pid": 1728719, "tid": 1728719, "ts": 6950381121401.656, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121411.735, "dur": 0.090 , "args": { "External id": 4237,"cbid": 200, "correlation": 15184 } }, { "ph": "f", "id": 15184, "pid": 1728719, "tid": 1728719, "ts": 6950381121411.735, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121411.905, "dur": 0.140 , "args": { "External id": 4237,"cbid": 468, "correlation": 15185 } }, { "ph": "f", "id": 15185, "pid": 1728719, "tid": 1728719, "ts": 6950381121411.905, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121412.246, "dur": 0.020 , "args": { "External id": 4237,"cbid": 1, "correlation": 15187 } }, { "ph": "f", "id": 15187, "pid": 1728719, "tid": 1728719, "ts": 6950381121412.246, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121412.336, "dur": 0.100 , "args": { "External id": 4237,"cbid": 468, "correlation": 15188 } }, { "ph": "f", "id": 15188, "pid": 1728719, "tid": 1728719, "ts": 6950381121412.336, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121412.586, "dur": 0.021 , "args": { "External id": 4237,"cbid": 1, "correlation": 15190 } }, { "ph": "f", "id": 15190, "pid": 1728719, "tid": 1728719, "ts": 6950381121412.586, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121412.697, "dur": 0.100 , "args": { "External id": 4237,"cbid": 468, "correlation": 15191 } }, { "ph": "f", "id": 15191, "pid": 1728719, "tid": 1728719, "ts": 6950381121412.697, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121412.947, "dur": 0.020 , "args": { "External id": 4237,"cbid": 1, "correlation": 15193 } }, { "ph": "f", "id": 15193, "pid": 1728719, "tid": 1728719, "ts": 6950381121412.947, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121413.047, "dur": 0.361 , "args": { "External id": 4237,"cbid": 273, "correlation": 15194 } }, { "ph": "f", "id": 15194, "pid": 1728719, "tid": 1728719, "ts": 6950381121413.047, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381234155.978, "dur": 236.253 , "args": { "External id": 4237,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15197, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15197, "pid": 0, "tid": 7, "ts": 6950381234155.978, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121413.769, "dur": 1.683 , "args": { "External id": 4237,"cbid": 211, "correlation": 15197 } }, { "ph": "s", "id": 15197, "pid": 1728719, "tid": 1728719, "ts": 6950381121413.769, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381234392.519, "dur": 58.559 , "args": { "External id": 4249,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15211, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15211, "pid": 0, "tid": 7, "ts": 6950381234392.519, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121422.956, "dur": 1.613 , "args": { "External id": 4249,"cbid": 307, "correlation": 15211 } }, { "ph": "s", "id": 15211, "pid": 1728719, "tid": 1728719, "ts": 6950381121422.956, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381234451.334, "dur": 66.016 , "args": { "External id": 4250,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15221, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15221, "pid": 0, "tid": 7, "ts": 6950381234451.334, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121433.065, "dur": 1.693 , "args": { "External id": 4250,"cbid": 307, "correlation": 15221 } }, { "ph": "s", "id": 15221, "pid": 1728719, "tid": 1728719, "ts": 6950381121433.065, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381234517.606, "dur": 39.647 , "args": { "External id": 4251,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15230, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15230, "pid": 0, "tid": 7, "ts": 6950381234517.606, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121442.652, "dur": 1.694 , "args": { "External id": 4251,"cbid": 307, "correlation": 15230 } }, { "ph": "s", "id": 15230, "pid": 1728719, "tid": 1728719, "ts": 6950381121442.652, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121457.570, "dur": 0.130 , "args": { "External id": 4259,"cbid": 200, "correlation": 15265 } }, { "ph": "f", "id": 15265, "pid": 1728719, "tid": 1728719, "ts": 6950381121457.570, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121457.781, "dur": 0.100 , "args": { "External id": 4259,"cbid": 200, "correlation": 15266 } }, { "ph": "f", "id": 15266, "pid": 1728719, "tid": 1728719, "ts": 6950381121457.781, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121457.971, "dur": 0.471 , "args": { "External id": 4259,"cbid": 273, "correlation": 15267 } }, { "ph": "f", "id": 15267, "pid": 1728719, "tid": 1728719, "ts": 6950381121457.971, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381234557.509, "dur": 861.143 , "args": { "External id": 4259,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15269, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15269, "pid": 0, "tid": 7, "ts": 6950381234557.509, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121458.742, "dur": 1.784 , "args": { "External id": 4259,"cbid": 211, "correlation": 15269 } }, { "ph": "s", "id": 15269, "pid": 1728719, "tid": 1728719, "ts": 6950381121458.742, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381235418.908, "dur": 9.504 , "args": { "External id": 4268,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15289, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15289, "pid": 0, "tid": 7, "ts": 6950381235418.908, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121474.532, "dur": 1.573 , "args": { "External id": 4268,"cbid": 652, "correlation": 15289 } }, { "ph": "s", "id": 15289, "pid": 1728719, "tid": 1728719, "ts": 6950381121474.532, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121485.502, "dur": 0.100 , "args": { "External id": 4269,"cbid": 200, "correlation": 15304 } }, { "ph": "f", "id": 15304, "pid": 1728719, "tid": 1728719, "ts": 6950381121485.502, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121485.703, "dur": 0.160 , "args": { "External id": 4269,"cbid": 468, "correlation": 15305 } }, { "ph": "f", "id": 15305, "pid": 1728719, "tid": 1728719, "ts": 6950381121485.703, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121486.133, "dur": 0.030 , "args": { "External id": 4269,"cbid": 1, "correlation": 15307 } }, { "ph": "f", "id": 15307, "pid": 1728719, "tid": 1728719, "ts": 6950381121486.133, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121486.234, "dur": 0.100 , "args": { "External id": 4269,"cbid": 468, "correlation": 15308 } }, { "ph": "f", "id": 15308, "pid": 1728719, "tid": 1728719, "ts": 6950381121486.234, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121486.494, "dur": 0.020 , "args": { "External id": 4269,"cbid": 1, "correlation": 15310 } }, { "ph": "f", "id": 15310, "pid": 1728719, "tid": 1728719, "ts": 6950381121486.494, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121486.604, "dur": 0.100 , "args": { "External id": 4269,"cbid": 468, "correlation": 15311 } }, { "ph": "f", "id": 15311, "pid": 1728719, "tid": 1728719, "ts": 6950381121486.604, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121486.865, "dur": 0.020 , "args": { "External id": 4269,"cbid": 1, "correlation": 15313 } }, { "ph": "f", "id": 15313, "pid": 1728719, "tid": 1728719, "ts": 6950381121486.865, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121486.975, "dur": 0.411 , "args": { "External id": 4269,"cbid": 273, "correlation": 15314 } }, { "ph": "f", "id": 15314, "pid": 1728719, "tid": 1728719, "ts": 6950381121486.975, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381235428.668, "dur": 212.510 , "args": { "External id": 4269,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15317, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15317, "pid": 0, "tid": 7, "ts": 6950381235428.668, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121487.766, "dur": 1.734 , "args": { "External id": 4269,"cbid": 211, "correlation": 15317 } }, { "ph": "s", "id": 15317, "pid": 1728719, "tid": 1728719, "ts": 6950381121487.766, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381235641.466, "dur": 44.639 , "args": { "External id": 4281,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15329, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15329, "pid": 0, "tid": 7, "ts": 6950381235641.466, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121497.444, "dur": 1.764 , "args": { "External id": 4281,"cbid": 307, "correlation": 15329 } }, { "ph": "s", "id": 15329, "pid": 1728719, "tid": 1728719, "ts": 6950381121497.444, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381235686.425, "dur": 13.600 , "args": { "External id": 4282,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15335, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15335, "pid": 0, "tid": 7, "ts": 6950381235686.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121505.469, "dur": 1.503 , "args": { "External id": 4282,"cbid": 652, "correlation": 15335 } }, { "ph": "s", "id": 15335, "pid": 1728719, "tid": 1728719, "ts": 6950381121505.469, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381235700.345, "dur": 13.632 , "args": { "External id": 4283,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15341, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15341, "pid": 0, "tid": 7, "ts": 6950381235700.345, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121514.015, "dur": 1.543 , "args": { "External id": 4283,"cbid": 652, "correlation": 15341 } }, { "ph": "s", "id": 15341, "pid": 1728719, "tid": 1728719, "ts": 6950381121514.015, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121525.477, "dur": 0.100 , "args": { "External id": 4284,"cbid": 200, "correlation": 15360 } }, { "ph": "f", "id": 15360, "pid": 1728719, "tid": 1728719, "ts": 6950381121525.477, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121525.687, "dur": 0.120 , "args": { "External id": 4284,"cbid": 468, "correlation": 15361 } }, { "ph": "f", "id": 15361, "pid": 1728719, "tid": 1728719, "ts": 6950381121525.687, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121526.058, "dur": 0.030 , "args": { "External id": 4284,"cbid": 1, "correlation": 15363 } }, { "ph": "f", "id": 15363, "pid": 1728719, "tid": 1728719, "ts": 6950381121526.058, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121526.168, "dur": 0.100 , "args": { "External id": 4284,"cbid": 468, "correlation": 15364 } }, { "ph": "f", "id": 15364, "pid": 1728719, "tid": 1728719, "ts": 6950381121526.168, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121526.438, "dur": 0.020 , "args": { "External id": 4284,"cbid": 1, "correlation": 15366 } }, { "ph": "f", "id": 15366, "pid": 1728719, "tid": 1728719, "ts": 6950381121526.438, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121526.569, "dur": 0.100 , "args": { "External id": 4284,"cbid": 468, "correlation": 15367 } }, { "ph": "f", "id": 15367, "pid": 1728719, "tid": 1728719, "ts": 6950381121526.569, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121526.839, "dur": 0.020 , "args": { "External id": 4284,"cbid": 1, "correlation": 15369 } }, { "ph": "f", "id": 15369, "pid": 1728719, "tid": 1728719, "ts": 6950381121526.839, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121526.939, "dur": 1.102 , "args": { "External id": 4284,"cbid": 273, "correlation": 15370 } }, { "ph": "f", "id": 15370, "pid": 1728719, "tid": 1728719, "ts": 6950381121526.939, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381235714.297, "dur": 679.032 , "args": { "External id": 4284,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15373, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15373, "pid": 0, "tid": 7, "ts": 6950381235714.297, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121528.412, "dur": 1.813 , "args": { "External id": 4284,"cbid": 211, "correlation": 15373 } }, { "ph": "s", "id": 15373, "pid": 1728719, "tid": 1728719, "ts": 6950381121528.412, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121543.170, "dur": 0.090 , "args": { "External id": 4296,"cbid": 200, "correlation": 15397 } }, { "ph": "f", "id": 15397, "pid": 1728719, "tid": 1728719, "ts": 6950381121543.170, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121543.340, "dur": 0.120 , "args": { "External id": 4296,"cbid": 468, "correlation": 15398 } }, { "ph": "f", "id": 15398, "pid": 1728719, "tid": 1728719, "ts": 6950381121543.340, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121543.660, "dur": 0.021 , "args": { "External id": 4296,"cbid": 1, "correlation": 15400 } }, { "ph": "f", "id": 15400, "pid": 1728719, "tid": 1728719, "ts": 6950381121543.660, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121543.751, "dur": 0.110 , "args": { "External id": 4296,"cbid": 468, "correlation": 15401 } }, { "ph": "f", "id": 15401, "pid": 1728719, "tid": 1728719, "ts": 6950381121543.751, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121544.001, "dur": 0.020 , "args": { "External id": 4296,"cbid": 1, "correlation": 15403 } }, { "ph": "f", "id": 15403, "pid": 1728719, "tid": 1728719, "ts": 6950381121544.001, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121544.121, "dur": 0.091 , "args": { "External id": 4296,"cbid": 468, "correlation": 15404 } }, { "ph": "f", "id": 15404, "pid": 1728719, "tid": 1728719, "ts": 6950381121544.121, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121544.372, "dur": 0.020 , "args": { "External id": 4296,"cbid": 1, "correlation": 15406 } }, { "ph": "f", "id": 15406, "pid": 1728719, "tid": 1728719, "ts": 6950381121544.372, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121544.482, "dur": 0.321 , "args": { "External id": 4296,"cbid": 273, "correlation": 15407 } }, { "ph": "f", "id": 15407, "pid": 1728719, "tid": 1728719, "ts": 6950381121544.482, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381236393.617, "dur": 690.233 , "args": { "External id": 4296,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15410, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15410, "pid": 0, "tid": 7, "ts": 6950381236393.617, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121545.163, "dur": 1.784 , "args": { "External id": 4296,"cbid": 211, "correlation": 15410 } }, { "ph": "s", "id": 15410, "pid": 1728719, "tid": 1728719, "ts": 6950381121545.163, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381237084.138, "dur": 153.534 , "args": { "External id": 4308,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15423, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15423, "pid": 0, "tid": 7, "ts": 6950381237084.138, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121554.831, "dur": 1.663 , "args": { "External id": 4308,"cbid": 307, "correlation": 15423 } }, { "ph": "s", "id": 15423, "pid": 1728719, "tid": 1728719, "ts": 6950381121554.831, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381237237.928, "dur": 59.616 , "args": { "External id": 4309,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15429, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15429, "pid": 0, "tid": 7, "ts": 6950381237237.928, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121563.347, "dur": 1.583 , "args": { "External id": 4309,"cbid": 652, "correlation": 15429 } }, { "ph": "s", "id": 15429, "pid": 1728719, "tid": 1728719, "ts": 6950381121563.347, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121574.808, "dur": 0.101 , "args": { "External id": 4310,"cbid": 200, "correlation": 15448 } }, { "ph": "f", "id": 15448, "pid": 1728719, "tid": 1728719, "ts": 6950381121574.808, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121575.009, "dur": 0.120 , "args": { "External id": 4310,"cbid": 468, "correlation": 15449 } }, { "ph": "f", "id": 15449, "pid": 1728719, "tid": 1728719, "ts": 6950381121575.009, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121575.400, "dur": 0.020 , "args": { "External id": 4310,"cbid": 1, "correlation": 15451 } }, { "ph": "f", "id": 15451, "pid": 1728719, "tid": 1728719, "ts": 6950381121575.400, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121575.580, "dur": 0.110 , "args": { "External id": 4310,"cbid": 468, "correlation": 15452 } }, { "ph": "f", "id": 15452, "pid": 1728719, "tid": 1728719, "ts": 6950381121575.580, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121575.830, "dur": 0.030 , "args": { "External id": 4310,"cbid": 1, "correlation": 15454 } }, { "ph": "f", "id": 15454, "pid": 1728719, "tid": 1728719, "ts": 6950381121575.830, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121575.941, "dur": 0.110 , "args": { "External id": 4310,"cbid": 468, "correlation": 15455 } }, { "ph": "f", "id": 15455, "pid": 1728719, "tid": 1728719, "ts": 6950381121575.941, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121576.201, "dur": 0.020 , "args": { "External id": 4310,"cbid": 1, "correlation": 15457 } }, { "ph": "f", "id": 15457, "pid": 1728719, "tid": 1728719, "ts": 6950381121576.201, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121576.301, "dur": 0.361 , "args": { "External id": 4310,"cbid": 273, "correlation": 15458 } }, { "ph": "f", "id": 15458, "pid": 1728719, "tid": 1728719, "ts": 6950381121576.301, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381237297.832, "dur": 690.904 , "args": { "External id": 4310,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15461, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15461, "pid": 0, "tid": 7, "ts": 6950381237297.832, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121577.023, "dur": 1.723 , "args": { "External id": 4310,"cbid": 211, "correlation": 15461 } }, { "ph": "s", "id": 15461, "pid": 1728719, "tid": 1728719, "ts": 6950381121577.023, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381237989.024, "dur": 40.992 , "args": { "External id": 4322,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15472, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15472, "pid": 0, "tid": 7, "ts": 6950381237989.024, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121587.292, "dur": 1.663 , "args": { "External id": 4322,"cbid": 307, "correlation": 15472 } }, { "ph": "s", "id": 15472, "pid": 1728719, "tid": 1728719, "ts": 6950381121587.292, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381238030.272, "dur": 21.439 , "args": { "External id": 4334,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15484, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15484, "pid": 0, "tid": 7, "ts": 6950381238030.272, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121665.477, "dur": 2.395 , "args": { "External id": 4334,"cbid": 307, "correlation": 15484 } }, { "ph": "s", "id": 15484, "pid": 1728719, "tid": 1728719, "ts": 6950381121665.477, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381238052.032, "dur": 12.703 , "args": { "External id": 4335,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15498, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15498, "pid": 0, "tid": 7, "ts": 6950381238052.032, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121679.313, "dur": 1.603 , "args": { "External id": 4335,"cbid": 652, "correlation": 15498 } }, { "ph": "s", "id": 15498, "pid": 1728719, "tid": 1728719, "ts": 6950381121679.313, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381238065.055, "dur": 12.448 , "args": { "External id": 4336,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15512, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15512, "pid": 0, "tid": 7, "ts": 6950381238065.055, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121690.975, "dur": 1.623 , "args": { "External id": 4336,"cbid": 652, "correlation": 15512 } }, { "ph": "s", "id": 15512, "pid": 1728719, "tid": 1728719, "ts": 6950381121690.975, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381238077.791, "dur": 11.424 , "args": { "External id": 4337,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15526, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15526, "pid": 0, "tid": 7, "ts": 6950381238077.791, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121701.554, "dur": 1.503 , "args": { "External id": 4337,"cbid": 652, "correlation": 15526 } }, { "ph": "s", "id": 15526, "pid": 1728719, "tid": 1728719, "ts": 6950381121701.554, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121713.336, "dur": 0.171 , "args": { "External id": 4338,"cbid": 200, "correlation": 15541 } }, { "ph": "f", "id": 15541, "pid": 1728719, "tid": 1728719, "ts": 6950381121713.336, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121713.647, "dur": 0.200 , "args": { "External id": 4338,"cbid": 468, "correlation": 15542 } }, { "ph": "f", "id": 15542, "pid": 1728719, "tid": 1728719, "ts": 6950381121713.647, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121714.218, "dur": 0.030 , "args": { "External id": 4338,"cbid": 1, "correlation": 15544 } }, { "ph": "f", "id": 15544, "pid": 1728719, "tid": 1728719, "ts": 6950381121714.218, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121714.338, "dur": 0.100 , "args": { "External id": 4338,"cbid": 468, "correlation": 15545 } }, { "ph": "f", "id": 15545, "pid": 1728719, "tid": 1728719, "ts": 6950381121714.338, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121714.589, "dur": 0.030 , "args": { "External id": 4338,"cbid": 1, "correlation": 15547 } }, { "ph": "f", "id": 15547, "pid": 1728719, "tid": 1728719, "ts": 6950381121714.589, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121714.709, "dur": 0.100 , "args": { "External id": 4338,"cbid": 468, "correlation": 15548 } }, { "ph": "f", "id": 15548, "pid": 1728719, "tid": 1728719, "ts": 6950381121714.709, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121714.979, "dur": 0.020 , "args": { "External id": 4338,"cbid": 1, "correlation": 15550 } }, { "ph": "f", "id": 15550, "pid": 1728719, "tid": 1728719, "ts": 6950381121714.979, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121715.130, "dur": 0.651 , "args": { "External id": 4338,"cbid": 273, "correlation": 15551 } }, { "ph": "f", "id": 15551, "pid": 1728719, "tid": 1728719, "ts": 6950381121715.130, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381238089.535, "dur": 220.126 , "args": { "External id": 4338,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15554, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15554, "pid": 0, "tid": 7, "ts": 6950381238089.535, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121716.272, "dur": 1.913 , "args": { "External id": 4338,"cbid": 211, "correlation": 15554 } }, { "ph": "s", "id": 15554, "pid": 1728719, "tid": 1728719, "ts": 6950381121716.272, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381238309.949, "dur": 8.992 , "args": { "External id": 4350,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15567, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15567, "pid": 0, "tid": 7, "ts": 6950381238309.949, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121727.092, "dur": 1.623 , "args": { "External id": 4350,"cbid": 307, "correlation": 15567 } }, { "ph": "s", "id": 15567, "pid": 1728719, "tid": 1728719, "ts": 6950381121727.092, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381238319.229, "dur": 37.183 , "args": { "External id": 4355,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15590, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15590, "pid": 0, "tid": 7, "ts": 6950381238319.229, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121745.937, "dur": 1.713 , "args": { "External id": 4355,"cbid": 307, "correlation": 15590 } }, { "ph": "s", "id": 15590, "pid": 1728719, "tid": 1728719, "ts": 6950381121745.937, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121756.046, "dur": 0.110 , "args": { "External id": 4356,"cbid": 200, "correlation": 15605 } }, { "ph": "f", "id": 15605, "pid": 1728719, "tid": 1728719, "ts": 6950381121756.046, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121756.246, "dur": 0.130 , "args": { "External id": 4356,"cbid": 468, "correlation": 15606 } }, { "ph": "f", "id": 15606, "pid": 1728719, "tid": 1728719, "ts": 6950381121756.246, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121756.607, "dur": 0.020 , "args": { "External id": 4356,"cbid": 1, "correlation": 15608 } }, { "ph": "f", "id": 15608, "pid": 1728719, "tid": 1728719, "ts": 6950381121756.607, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121756.717, "dur": 0.100 , "args": { "External id": 4356,"cbid": 468, "correlation": 15609 } }, { "ph": "f", "id": 15609, "pid": 1728719, "tid": 1728719, "ts": 6950381121756.717, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121756.978, "dur": 0.020 , "args": { "External id": 4356,"cbid": 1, "correlation": 15611 } }, { "ph": "f", "id": 15611, "pid": 1728719, "tid": 1728719, "ts": 6950381121756.978, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121757.088, "dur": 0.100 , "args": { "External id": 4356,"cbid": 468, "correlation": 15612 } }, { "ph": "f", "id": 15612, "pid": 1728719, "tid": 1728719, "ts": 6950381121757.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121757.348, "dur": 0.020 , "args": { "External id": 4356,"cbid": 1, "correlation": 15614 } }, { "ph": "f", "id": 15614, "pid": 1728719, "tid": 1728719, "ts": 6950381121757.348, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121757.458, "dur": 0.421 , "args": { "External id": 4356,"cbid": 273, "correlation": 15615 } }, { "ph": "f", "id": 15615, "pid": 1728719, "tid": 1728719, "ts": 6950381121757.458, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381238356.700, "dur": 234.494 , "args": { "External id": 4356,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15618, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15618, "pid": 0, "tid": 7, "ts": 6950381238356.700, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121758.260, "dur": 1.823 , "args": { "External id": 4356,"cbid": 211, "correlation": 15618 } }, { "ph": "s", "id": 15618, "pid": 1728719, "tid": 1728719, "ts": 6950381121758.260, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381238591.514, "dur": 8.864 , "args": { "External id": 4368,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15627, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15627, "pid": 0, "tid": 7, "ts": 6950381238591.514, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121766.475, "dur": 1.633 , "args": { "External id": 4368,"cbid": 307, "correlation": 15627 } }, { "ph": "s", "id": 15627, "pid": 1728719, "tid": 1728719, "ts": 6950381121766.475, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381238600.666, "dur": 25.631 , "args": { "External id": 4373,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15650, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15650, "pid": 0, "tid": 7, "ts": 6950381238600.666, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121782.725, "dur": 1.573 , "args": { "External id": 4373,"cbid": 307, "correlation": 15650 } }, { "ph": "s", "id": 15650, "pid": 1728719, "tid": 1728719, "ts": 6950381121782.725, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121792.814, "dur": 0.100 , "args": { "External id": 4374,"cbid": 200, "correlation": 15665 } }, { "ph": "f", "id": 15665, "pid": 1728719, "tid": 1728719, "ts": 6950381121792.814, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121793.005, "dur": 0.130 , "args": { "External id": 4374,"cbid": 468, "correlation": 15666 } }, { "ph": "f", "id": 15666, "pid": 1728719, "tid": 1728719, "ts": 6950381121793.005, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121793.355, "dur": 0.040 , "args": { "External id": 4374,"cbid": 1, "correlation": 15668 } }, { "ph": "f", "id": 15668, "pid": 1728719, "tid": 1728719, "ts": 6950381121793.355, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121793.465, "dur": 0.101 , "args": { "External id": 4374,"cbid": 468, "correlation": 15669 } }, { "ph": "f", "id": 15669, "pid": 1728719, "tid": 1728719, "ts": 6950381121793.465, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121793.726, "dur": 0.020 , "args": { "External id": 4374,"cbid": 1, "correlation": 15671 } }, { "ph": "f", "id": 15671, "pid": 1728719, "tid": 1728719, "ts": 6950381121793.726, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121793.836, "dur": 0.100 , "args": { "External id": 4374,"cbid": 468, "correlation": 15672 } }, { "ph": "f", "id": 15672, "pid": 1728719, "tid": 1728719, "ts": 6950381121793.836, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121794.107, "dur": 0.020 , "args": { "External id": 4374,"cbid": 1, "correlation": 15674 } }, { "ph": "f", "id": 15674, "pid": 1728719, "tid": 1728719, "ts": 6950381121794.107, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121794.227, "dur": 0.381 , "args": { "External id": 4374,"cbid": 273, "correlation": 15675 } }, { "ph": "f", "id": 15675, "pid": 1728719, "tid": 1728719, "ts": 6950381121794.227, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381238626.585, "dur": 236.574 , "args": { "External id": 4374,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15678, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15678, "pid": 0, "tid": 7, "ts": 6950381238626.585, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121794.968, "dur": 1.713 , "args": { "External id": 4374,"cbid": 211, "correlation": 15678 } }, { "ph": "s", "id": 15678, "pid": 1728719, "tid": 1728719, "ts": 6950381121794.968, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381238863.447, "dur": 59.935 , "args": { "External id": 4386,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15692, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15692, "pid": 0, "tid": 7, "ts": 6950381238863.447, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121804.296, "dur": 1.613 , "args": { "External id": 4386,"cbid": 307, "correlation": 15692 } }, { "ph": "s", "id": 15692, "pid": 1728719, "tid": 1728719, "ts": 6950381121804.296, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381238923.670, "dur": 63.999 , "args": { "External id": 4387,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15702, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15702, "pid": 0, "tid": 7, "ts": 6950381238923.670, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121814.184, "dur": 1.894 , "args": { "External id": 4387,"cbid": 307, "correlation": 15702 } }, { "ph": "s", "id": 15702, "pid": 1728719, "tid": 1728719, "ts": 6950381121814.184, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381238987.957, "dur": 39.776 , "args": { "External id": 4388,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15711, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15711, "pid": 0, "tid": 7, "ts": 6950381238987.957, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121824.072, "dur": 1.754 , "args": { "External id": 4388,"cbid": 307, "correlation": 15711 } }, { "ph": "s", "id": 15711, "pid": 1728719, "tid": 1728719, "ts": 6950381121824.072, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121839.561, "dur": 0.120 , "args": { "External id": 4396,"cbid": 200, "correlation": 15746 } }, { "ph": "f", "id": 15746, "pid": 1728719, "tid": 1728719, "ts": 6950381121839.561, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121839.752, "dur": 0.080 , "args": { "External id": 4396,"cbid": 200, "correlation": 15747 } }, { "ph": "f", "id": 15747, "pid": 1728719, "tid": 1728719, "ts": 6950381121839.752, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121839.932, "dur": 0.451 , "args": { "External id": 4396,"cbid": 273, "correlation": 15748 } }, { "ph": "f", "id": 15748, "pid": 1728719, "tid": 1728719, "ts": 6950381121839.932, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381239028.021, "dur": 861.686 , "args": { "External id": 4396,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15750, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15750, "pid": 0, "tid": 7, "ts": 6950381239028.021, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121840.774, "dur": 1.953 , "args": { "External id": 4396,"cbid": 211, "correlation": 15750 } }, { "ph": "s", "id": 15750, "pid": 1728719, "tid": 1728719, "ts": 6950381121840.774, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381239889.995, "dur": 9.440 , "args": { "External id": 4405,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15770, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15770, "pid": 0, "tid": 7, "ts": 6950381239889.995, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121856.833, "dur": 1.623 , "args": { "External id": 4405,"cbid": 652, "correlation": 15770 } }, { "ph": "s", "id": 15770, "pid": 1728719, "tid": 1728719, "ts": 6950381121856.833, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121867.694, "dur": 0.110 , "args": { "External id": 4406,"cbid": 200, "correlation": 15785 } }, { "ph": "f", "id": 15785, "pid": 1728719, "tid": 1728719, "ts": 6950381121867.694, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121867.914, "dur": 0.140 , "args": { "External id": 4406,"cbid": 468, "correlation": 15786 } }, { "ph": "f", "id": 15786, "pid": 1728719, "tid": 1728719, "ts": 6950381121867.914, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121868.305, "dur": 0.020 , "args": { "External id": 4406,"cbid": 1, "correlation": 15788 } }, { "ph": "f", "id": 15788, "pid": 1728719, "tid": 1728719, "ts": 6950381121868.305, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121868.395, "dur": 0.100 , "args": { "External id": 4406,"cbid": 468, "correlation": 15789 } }, { "ph": "f", "id": 15789, "pid": 1728719, "tid": 1728719, "ts": 6950381121868.395, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121868.645, "dur": 0.020 , "args": { "External id": 4406,"cbid": 1, "correlation": 15791 } }, { "ph": "f", "id": 15791, "pid": 1728719, "tid": 1728719, "ts": 6950381121868.645, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121868.756, "dur": 0.100 , "args": { "External id": 4406,"cbid": 468, "correlation": 15792 } }, { "ph": "f", "id": 15792, "pid": 1728719, "tid": 1728719, "ts": 6950381121868.756, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121869.016, "dur": 0.020 , "args": { "External id": 4406,"cbid": 1, "correlation": 15794 } }, { "ph": "f", "id": 15794, "pid": 1728719, "tid": 1728719, "ts": 6950381121869.016, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121869.136, "dur": 0.411 , "args": { "External id": 4406,"cbid": 273, "correlation": 15795 } }, { "ph": "f", "id": 15795, "pid": 1728719, "tid": 1728719, "ts": 6950381121869.136, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381239899.691, "dur": 211.806 , "args": { "External id": 4406,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15798, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15798, "pid": 0, "tid": 7, "ts": 6950381239899.691, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121869.898, "dur": 1.773 , "args": { "External id": 4406,"cbid": 211, "correlation": 15798 } }, { "ph": "s", "id": 15798, "pid": 1728719, "tid": 1728719, "ts": 6950381121869.898, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381240111.785, "dur": 46.880 , "args": { "External id": 4418,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15810, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15810, "pid": 0, "tid": 7, "ts": 6950381240111.785, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121879.335, "dur": 1.774 , "args": { "External id": 4418,"cbid": 307, "correlation": 15810 } }, { "ph": "s", "id": 15810, "pid": 1728719, "tid": 1728719, "ts": 6950381121879.335, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381240158.984, "dur": 13.184 , "args": { "External id": 4419,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15816, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15816, "pid": 0, "tid": 7, "ts": 6950381240158.984, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121887.551, "dur": 1.512 , "args": { "External id": 4419,"cbid": 652, "correlation": 15816 } }, { "ph": "s", "id": 15816, "pid": 1728719, "tid": 1728719, "ts": 6950381121887.551, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381240172.456, "dur": 12.992 , "args": { "External id": 4420,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15822, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15822, "pid": 0, "tid": 7, "ts": 6950381240172.456, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121895.806, "dur": 1.453 , "args": { "External id": 4420,"cbid": 652, "correlation": 15822 } }, { "ph": "s", "id": 15822, "pid": 1728719, "tid": 1728719, "ts": 6950381121895.806, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121907.147, "dur": 0.130 , "args": { "External id": 4421,"cbid": 200, "correlation": 15841 } }, { "ph": "f", "id": 15841, "pid": 1728719, "tid": 1728719, "ts": 6950381121907.147, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121907.398, "dur": 0.120 , "args": { "External id": 4421,"cbid": 468, "correlation": 15842 } }, { "ph": "f", "id": 15842, "pid": 1728719, "tid": 1728719, "ts": 6950381121907.398, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121907.878, "dur": 0.020 , "args": { "External id": 4421,"cbid": 1, "correlation": 15844 } }, { "ph": "f", "id": 15844, "pid": 1728719, "tid": 1728719, "ts": 6950381121907.878, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121907.969, "dur": 0.100 , "args": { "External id": 4421,"cbid": 468, "correlation": 15845 } }, { "ph": "f", "id": 15845, "pid": 1728719, "tid": 1728719, "ts": 6950381121907.969, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121908.219, "dur": 0.020 , "args": { "External id": 4421,"cbid": 1, "correlation": 15847 } }, { "ph": "f", "id": 15847, "pid": 1728719, "tid": 1728719, "ts": 6950381121908.219, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121908.329, "dur": 0.100 , "args": { "External id": 4421,"cbid": 468, "correlation": 15848 } }, { "ph": "f", "id": 15848, "pid": 1728719, "tid": 1728719, "ts": 6950381121908.329, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121908.580, "dur": 0.020 , "args": { "External id": 4421,"cbid": 1, "correlation": 15850 } }, { "ph": "f", "id": 15850, "pid": 1728719, "tid": 1728719, "ts": 6950381121908.580, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121908.690, "dur": 0.381 , "args": { "External id": 4421,"cbid": 273, "correlation": 15851 } }, { "ph": "f", "id": 15851, "pid": 1728719, "tid": 1728719, "ts": 6950381121908.690, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381240185.768, "dur": 677.369 , "args": { "External id": 4421,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15854, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15854, "pid": 0, "tid": 7, "ts": 6950381240185.768, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121909.431, "dur": 1.794 , "args": { "External id": 4421,"cbid": 211, "correlation": 15854 } }, { "ph": "s", "id": 15854, "pid": 1728719, "tid": 1728719, "ts": 6950381121909.431, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121921.173, "dur": 0.090 , "args": { "External id": 4433,"cbid": 200, "correlation": 15878 } }, { "ph": "f", "id": 15878, "pid": 1728719, "tid": 1728719, "ts": 6950381121921.173, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121921.364, "dur": 0.120 , "args": { "External id": 4433,"cbid": 468, "correlation": 15879 } }, { "ph": "f", "id": 15879, "pid": 1728719, "tid": 1728719, "ts": 6950381121921.364, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121921.674, "dur": 0.020 , "args": { "External id": 4433,"cbid": 1, "correlation": 15881 } }, { "ph": "f", "id": 15881, "pid": 1728719, "tid": 1728719, "ts": 6950381121921.674, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121921.764, "dur": 0.100 , "args": { "External id": 4433,"cbid": 468, "correlation": 15882 } }, { "ph": "f", "id": 15882, "pid": 1728719, "tid": 1728719, "ts": 6950381121921.764, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121922.045, "dur": 0.020 , "args": { "External id": 4433,"cbid": 1, "correlation": 15884 } }, { "ph": "f", "id": 15884, "pid": 1728719, "tid": 1728719, "ts": 6950381121922.045, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121922.155, "dur": 0.090 , "args": { "External id": 4433,"cbid": 468, "correlation": 15885 } }, { "ph": "f", "id": 15885, "pid": 1728719, "tid": 1728719, "ts": 6950381121922.155, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121922.405, "dur": 0.021 , "args": { "External id": 4433,"cbid": 1, "correlation": 15887 } }, { "ph": "f", "id": 15887, "pid": 1728719, "tid": 1728719, "ts": 6950381121922.405, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121922.506, "dur": 0.330 , "args": { "External id": 4433,"cbid": 273, "correlation": 15888 } }, { "ph": "f", "id": 15888, "pid": 1728719, "tid": 1728719, "ts": 6950381121922.506, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381240863.425, "dur": 690.488 , "args": { "External id": 4433,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15891, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15891, "pid": 0, "tid": 7, "ts": 6950381240863.425, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121923.187, "dur": 1.663 , "args": { "External id": 4433,"cbid": 211, "correlation": 15891 } }, { "ph": "s", "id": 15891, "pid": 1728719, "tid": 1728719, "ts": 6950381121923.187, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381241554.202, "dur": 152.830 , "args": { "External id": 4445,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15904, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15904, "pid": 0, "tid": 7, "ts": 6950381241554.202, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121932.264, "dur": 1.843 , "args": { "External id": 4445,"cbid": 307, "correlation": 15904 } }, { "ph": "s", "id": 15904, "pid": 1728719, "tid": 1728719, "ts": 6950381121932.264, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381241707.288, "dur": 59.391 , "args": { "External id": 4446,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15910, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15910, "pid": 0, "tid": 7, "ts": 6950381241707.288, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381121940.870, "dur": 1.823 , "args": { "External id": 4446,"cbid": 652, "correlation": 15910 } }, { "ph": "s", "id": 15910, "pid": 1728719, "tid": 1728719, "ts": 6950381121940.870, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121952.371, "dur": 0.090 , "args": { "External id": 4447,"cbid": 200, "correlation": 15929 } }, { "ph": "f", "id": 15929, "pid": 1728719, "tid": 1728719, "ts": 6950381121952.371, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121952.572, "dur": 0.130 , "args": { "External id": 4447,"cbid": 468, "correlation": 15930 } }, { "ph": "f", "id": 15930, "pid": 1728719, "tid": 1728719, "ts": 6950381121952.572, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121952.882, "dur": 0.020 , "args": { "External id": 4447,"cbid": 1, "correlation": 15932 } }, { "ph": "f", "id": 15932, "pid": 1728719, "tid": 1728719, "ts": 6950381121952.882, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121952.972, "dur": 0.101 , "args": { "External id": 4447,"cbid": 468, "correlation": 15933 } }, { "ph": "f", "id": 15933, "pid": 1728719, "tid": 1728719, "ts": 6950381121952.972, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121953.223, "dur": 0.020 , "args": { "External id": 4447,"cbid": 1, "correlation": 15935 } }, { "ph": "f", "id": 15935, "pid": 1728719, "tid": 1728719, "ts": 6950381121953.223, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121953.333, "dur": 0.100 , "args": { "External id": 4447,"cbid": 468, "correlation": 15936 } }, { "ph": "f", "id": 15936, "pid": 1728719, "tid": 1728719, "ts": 6950381121953.333, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381121953.584, "dur": 0.020 , "args": { "External id": 4447,"cbid": 1, "correlation": 15938 } }, { "ph": "f", "id": 15938, "pid": 1728719, "tid": 1728719, "ts": 6950381121953.584, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381121953.684, "dur": 0.350 , "args": { "External id": 4447,"cbid": 273, "correlation": 15939 } }, { "ph": "f", "id": 15939, "pid": 1728719, "tid": 1728719, "ts": 6950381121953.684, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381241766.967, "dur": 692.569 , "args": { "External id": 4447,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15942, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15942, "pid": 0, "tid": 7, "ts": 6950381241766.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121954.405, "dur": 1.723 , "args": { "External id": 4447,"cbid": 211, "correlation": 15942 } }, { "ph": "s", "id": 15942, "pid": 1728719, "tid": 1728719, "ts": 6950381121954.405, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381242459.824, "dur": 40.703 , "args": { "External id": 4459,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15953, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15953, "pid": 0, "tid": 7, "ts": 6950381242459.824, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381121964.123, "dur": 1.723 , "args": { "External id": 4459,"cbid": 307, "correlation": 15953 } }, { "ph": "s", "id": 15953, "pid": 1728719, "tid": 1728719, "ts": 6950381121964.123, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381242500.815, "dur": 21.568 , "args": { "External id": 4471,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15965, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15965, "pid": 0, "tid": 7, "ts": 6950381242500.815, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122041.868, "dur": 2.525 , "args": { "External id": 4471,"cbid": 307, "correlation": 15965 } }, { "ph": "s", "id": 15965, "pid": 1728719, "tid": 1728719, "ts": 6950381122041.868, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381242522.703, "dur": 13.184 , "args": { "External id": 4472,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15979, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15979, "pid": 0, "tid": 7, "ts": 6950381242522.703, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122055.744, "dur": 1.763 , "args": { "External id": 4472,"cbid": 652, "correlation": 15979 } }, { "ph": "s", "id": 15979, "pid": 1728719, "tid": 1728719, "ts": 6950381122055.744, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381242536.207, "dur": 12.384 , "args": { "External id": 4473,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 15993, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 15993, "pid": 0, "tid": 7, "ts": 6950381242536.207, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122067.656, "dur": 1.683 , "args": { "External id": 4473,"cbid": 652, "correlation": 15993 } }, { "ph": "s", "id": 15993, "pid": 1728719, "tid": 1728719, "ts": 6950381122067.656, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381242548.879, "dur": 10.848 , "args": { "External id": 4474,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16007, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16007, "pid": 0, "tid": 7, "ts": 6950381242548.879, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122078.195, "dur": 1.553 , "args": { "External id": 4474,"cbid": 652, "correlation": 16007 } }, { "ph": "s", "id": 16007, "pid": 1728719, "tid": 1728719, "ts": 6950381122078.195, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122089.697, "dur": 0.150 , "args": { "External id": 4475,"cbid": 200, "correlation": 16022 } }, { "ph": "f", "id": 16022, "pid": 1728719, "tid": 1728719, "ts": 6950381122089.697, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122089.967, "dur": 0.171 , "args": { "External id": 4475,"cbid": 468, "correlation": 16023 } }, { "ph": "f", "id": 16023, "pid": 1728719, "tid": 1728719, "ts": 6950381122089.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122090.448, "dur": 0.030 , "args": { "External id": 4475,"cbid": 1, "correlation": 16025 } }, { "ph": "f", "id": 16025, "pid": 1728719, "tid": 1728719, "ts": 6950381122090.448, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122090.548, "dur": 0.101 , "args": { "External id": 4475,"cbid": 468, "correlation": 16026 } }, { "ph": "f", "id": 16026, "pid": 1728719, "tid": 1728719, "ts": 6950381122090.548, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122090.799, "dur": 0.020 , "args": { "External id": 4475,"cbid": 1, "correlation": 16028 } }, { "ph": "f", "id": 16028, "pid": 1728719, "tid": 1728719, "ts": 6950381122090.799, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122090.929, "dur": 0.100 , "args": { "External id": 4475,"cbid": 468, "correlation": 16029 } }, { "ph": "f", "id": 16029, "pid": 1728719, "tid": 1728719, "ts": 6950381122090.929, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122091.200, "dur": 0.020 , "args": { "External id": 4475,"cbid": 1, "correlation": 16031 } }, { "ph": "f", "id": 16031, "pid": 1728719, "tid": 1728719, "ts": 6950381122091.200, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122091.390, "dur": 0.641 , "args": { "External id": 4475,"cbid": 273, "correlation": 16032 } }, { "ph": "f", "id": 16032, "pid": 1728719, "tid": 1728719, "ts": 6950381122091.390, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381242560.014, "dur": 220.350 , "args": { "External id": 4475,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16035, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16035, "pid": 0, "tid": 7, "ts": 6950381242560.014, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122092.532, "dur": 1.864 , "args": { "External id": 4475,"cbid": 211, "correlation": 16035 } }, { "ph": "s", "id": 16035, "pid": 1728719, "tid": 1728719, "ts": 6950381122092.532, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381242780.652, "dur": 8.928 , "args": { "External id": 4487,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16048, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16048, "pid": 0, "tid": 7, "ts": 6950381242780.652, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122102.891, "dur": 1.644 , "args": { "External id": 4487,"cbid": 307, "correlation": 16048 } }, { "ph": "s", "id": 16048, "pid": 1728719, "tid": 1728719, "ts": 6950381122102.891, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381242789.836, "dur": 37.440 , "args": { "External id": 4492,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16071, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16071, "pid": 0, "tid": 7, "ts": 6950381242789.836, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122121.546, "dur": 1.683 , "args": { "External id": 4492,"cbid": 307, "correlation": 16071 } }, { "ph": "s", "id": 16071, "pid": 1728719, "tid": 1728719, "ts": 6950381122121.546, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122132.056, "dur": 0.100 , "args": { "External id": 4493,"cbid": 200, "correlation": 16086 } }, { "ph": "f", "id": 16086, "pid": 1728719, "tid": 1728719, "ts": 6950381122132.056, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122132.246, "dur": 0.150 , "args": { "External id": 4493,"cbid": 468, "correlation": 16087 } }, { "ph": "f", "id": 16087, "pid": 1728719, "tid": 1728719, "ts": 6950381122132.246, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122132.627, "dur": 0.030 , "args": { "External id": 4493,"cbid": 1, "correlation": 16089 } }, { "ph": "f", "id": 16089, "pid": 1728719, "tid": 1728719, "ts": 6950381122132.627, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122132.717, "dur": 0.100 , "args": { "External id": 4493,"cbid": 468, "correlation": 16090 } }, { "ph": "f", "id": 16090, "pid": 1728719, "tid": 1728719, "ts": 6950381122132.717, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122132.967, "dur": 0.020 , "args": { "External id": 4493,"cbid": 1, "correlation": 16092 } }, { "ph": "f", "id": 16092, "pid": 1728719, "tid": 1728719, "ts": 6950381122132.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122133.088, "dur": 0.100 , "args": { "External id": 4493,"cbid": 468, "correlation": 16093 } }, { "ph": "f", "id": 16093, "pid": 1728719, "tid": 1728719, "ts": 6950381122133.088, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122133.348, "dur": 0.020 , "args": { "External id": 4493,"cbid": 1, "correlation": 16095 } }, { "ph": "f", "id": 16095, "pid": 1728719, "tid": 1728719, "ts": 6950381122133.348, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122133.468, "dur": 0.391 , "args": { "External id": 4493,"cbid": 273, "correlation": 16096 } }, { "ph": "f", "id": 16096, "pid": 1728719, "tid": 1728719, "ts": 6950381122133.468, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381242827.596, "dur": 233.629 , "args": { "External id": 4493,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16099, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16099, "pid": 0, "tid": 7, "ts": 6950381242827.596, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122134.240, "dur": 1.813 , "args": { "External id": 4493,"cbid": 211, "correlation": 16099 } }, { "ph": "s", "id": 16099, "pid": 1728719, "tid": 1728719, "ts": 6950381122134.240, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381243061.545, "dur": 8.864 , "args": { "External id": 4505,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16108, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16108, "pid": 0, "tid": 7, "ts": 6950381243061.545, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122141.994, "dur": 1.553 , "args": { "External id": 4505,"cbid": 307, "correlation": 16108 } }, { "ph": "s", "id": 16108, "pid": 1728719, "tid": 1728719, "ts": 6950381122141.994, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381243070.697, "dur": 25.888 , "args": { "External id": 4510,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16131, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16131, "pid": 0, "tid": 7, "ts": 6950381243070.697, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122157.613, "dur": 1.794 , "args": { "External id": 4510,"cbid": 307, "correlation": 16131 } }, { "ph": "s", "id": 16131, "pid": 1728719, "tid": 1728719, "ts": 6950381122157.613, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122167.662, "dur": 0.090 , "args": { "External id": 4511,"cbid": 200, "correlation": 16146 } }, { "ph": "f", "id": 16146, "pid": 1728719, "tid": 1728719, "ts": 6950381122167.662, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122167.852, "dur": 0.141 , "args": { "External id": 4511,"cbid": 468, "correlation": 16147 } }, { "ph": "f", "id": 16147, "pid": 1728719, "tid": 1728719, "ts": 6950381122167.852, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122168.193, "dur": 0.030 , "args": { "External id": 4511,"cbid": 1, "correlation": 16149 } }, { "ph": "f", "id": 16149, "pid": 1728719, "tid": 1728719, "ts": 6950381122168.193, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122168.303, "dur": 0.100 , "args": { "External id": 4511,"cbid": 468, "correlation": 16150 } }, { "ph": "f", "id": 16150, "pid": 1728719, "tid": 1728719, "ts": 6950381122168.303, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122168.554, "dur": 0.020 , "args": { "External id": 4511,"cbid": 1, "correlation": 16152 } }, { "ph": "f", "id": 16152, "pid": 1728719, "tid": 1728719, "ts": 6950381122168.554, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122168.664, "dur": 0.100 , "args": { "External id": 4511,"cbid": 468, "correlation": 16153 } }, { "ph": "f", "id": 16153, "pid": 1728719, "tid": 1728719, "ts": 6950381122168.664, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122168.924, "dur": 0.020 , "args": { "External id": 4511,"cbid": 1, "correlation": 16155 } }, { "ph": "f", "id": 16155, "pid": 1728719, "tid": 1728719, "ts": 6950381122168.924, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122169.045, "dur": 0.370 , "args": { "External id": 4511,"cbid": 273, "correlation": 16156 } }, { "ph": "f", "id": 16156, "pid": 1728719, "tid": 1728719, "ts": 6950381122169.045, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381243096.873, "dur": 237.821 , "args": { "External id": 4511,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16159, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16159, "pid": 0, "tid": 7, "ts": 6950381243096.873, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122169.806, "dur": 1.783 , "args": { "External id": 4511,"cbid": 211, "correlation": 16159 } }, { "ph": "s", "id": 16159, "pid": 1728719, "tid": 1728719, "ts": 6950381122169.806, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381243334.982, "dur": 58.335 , "args": { "External id": 4523,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16173, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16173, "pid": 0, "tid": 7, "ts": 6950381243334.982, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122178.973, "dur": 1.573 , "args": { "External id": 4523,"cbid": 307, "correlation": 16173 } }, { "ph": "s", "id": 16173, "pid": 1728719, "tid": 1728719, "ts": 6950381122178.973, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381243393.605, "dur": 65.504 , "args": { "External id": 4524,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16183, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16183, "pid": 0, "tid": 7, "ts": 6950381243393.605, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122188.751, "dur": 1.663 , "args": { "External id": 4524,"cbid": 307, "correlation": 16183 } }, { "ph": "s", "id": 16183, "pid": 1728719, "tid": 1728719, "ts": 6950381122188.751, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381243459.365, "dur": 38.879 , "args": { "External id": 4525,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16192, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16192, "pid": 0, "tid": 7, "ts": 6950381243459.365, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122198.129, "dur": 1.653 , "args": { "External id": 4525,"cbid": 307, "correlation": 16192 } }, { "ph": "s", "id": 16192, "pid": 1728719, "tid": 1728719, "ts": 6950381122198.129, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122213.447, "dur": 0.120 , "args": { "External id": 4533,"cbid": 200, "correlation": 16227 } }, { "ph": "f", "id": 16227, "pid": 1728719, "tid": 1728719, "ts": 6950381122213.447, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122213.638, "dur": 0.150 , "args": { "External id": 4533,"cbid": 200, "correlation": 16228 } }, { "ph": "f", "id": 16228, "pid": 1728719, "tid": 1728719, "ts": 6950381122213.638, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122213.898, "dur": 0.451 , "args": { "External id": 4533,"cbid": 273, "correlation": 16229 } }, { "ph": "f", "id": 16229, "pid": 1728719, "tid": 1728719, "ts": 6950381122213.898, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381243498.532, "dur": 861.527 , "args": { "External id": 4533,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16231, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16231, "pid": 0, "tid": 7, "ts": 6950381243498.532, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122214.880, "dur": 1.813 , "args": { "External id": 4533,"cbid": 211, "correlation": 16231 } }, { "ph": "s", "id": 16231, "pid": 1728719, "tid": 1728719, "ts": 6950381122214.880, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381244360.315, "dur": 9.504 , "args": { "External id": 4542,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16251, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16251, "pid": 0, "tid": 7, "ts": 6950381244360.315, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122230.820, "dur": 1.593 , "args": { "External id": 4542,"cbid": 652, "correlation": 16251 } }, { "ph": "s", "id": 16251, "pid": 1728719, "tid": 1728719, "ts": 6950381122230.820, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122241.540, "dur": 0.110 , "args": { "External id": 4543,"cbid": 200, "correlation": 16266 } }, { "ph": "f", "id": 16266, "pid": 1728719, "tid": 1728719, "ts": 6950381122241.540, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122241.750, "dur": 0.170 , "args": { "External id": 4543,"cbid": 468, "correlation": 16267 } }, { "ph": "f", "id": 16267, "pid": 1728719, "tid": 1728719, "ts": 6950381122241.750, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122242.171, "dur": 0.030 , "args": { "External id": 4543,"cbid": 1, "correlation": 16269 } }, { "ph": "f", "id": 16269, "pid": 1728719, "tid": 1728719, "ts": 6950381122242.171, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122242.271, "dur": 0.100 , "args": { "External id": 4543,"cbid": 468, "correlation": 16270 } }, { "ph": "f", "id": 16270, "pid": 1728719, "tid": 1728719, "ts": 6950381122242.271, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122242.531, "dur": 0.020 , "args": { "External id": 4543,"cbid": 1, "correlation": 16272 } }, { "ph": "f", "id": 16272, "pid": 1728719, "tid": 1728719, "ts": 6950381122242.531, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122242.652, "dur": 0.100 , "args": { "External id": 4543,"cbid": 468, "correlation": 16273 } }, { "ph": "f", "id": 16273, "pid": 1728719, "tid": 1728719, "ts": 6950381122242.652, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122242.912, "dur": 0.020 , "args": { "External id": 4543,"cbid": 1, "correlation": 16275 } }, { "ph": "f", "id": 16275, "pid": 1728719, "tid": 1728719, "ts": 6950381122242.912, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122243.022, "dur": 0.361 , "args": { "External id": 4543,"cbid": 273, "correlation": 16276 } }, { "ph": "f", "id": 16276, "pid": 1728719, "tid": 1728719, "ts": 6950381122243.022, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381244370.107, "dur": 213.502 , "args": { "External id": 4543,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16279, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16279, "pid": 0, "tid": 7, "ts": 6950381244370.107, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122243.744, "dur": 1.763 , "args": { "External id": 4543,"cbid": 211, "correlation": 16279 } }, { "ph": "s", "id": 16279, "pid": 1728719, "tid": 1728719, "ts": 6950381122243.744, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381244583.897, "dur": 45.599 , "args": { "External id": 4555,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16291, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16291, "pid": 0, "tid": 7, "ts": 6950381244583.897, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122253.442, "dur": 1.693 , "args": { "External id": 4555,"cbid": 307, "correlation": 16291 } }, { "ph": "s", "id": 16291, "pid": 1728719, "tid": 1728719, "ts": 6950381122253.442, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381244629.816, "dur": 13.440 , "args": { "External id": 4556,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16297, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16297, "pid": 0, "tid": 7, "ts": 6950381244629.816, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122261.246, "dur": 1.523 , "args": { "External id": 4556,"cbid": 652, "correlation": 16297 } }, { "ph": "s", "id": 16297, "pid": 1728719, "tid": 1728719, "ts": 6950381122261.246, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381244643.576, "dur": 13.184 , "args": { "External id": 4557,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16303, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16303, "pid": 0, "tid": 7, "ts": 6950381244643.576, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122269.431, "dur": 1.523 , "args": { "External id": 4557,"cbid": 652, "correlation": 16303 } }, { "ph": "s", "id": 16303, "pid": 1728719, "tid": 1728719, "ts": 6950381122269.431, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122281.043, "dur": 0.100 , "args": { "External id": 4558,"cbid": 200, "correlation": 16322 } }, { "ph": "f", "id": 16322, "pid": 1728719, "tid": 1728719, "ts": 6950381122281.043, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122281.253, "dur": 0.121 , "args": { "External id": 4558,"cbid": 468, "correlation": 16323 } }, { "ph": "f", "id": 16323, "pid": 1728719, "tid": 1728719, "ts": 6950381122281.253, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122281.574, "dur": 0.030 , "args": { "External id": 4558,"cbid": 1, "correlation": 16325 } }, { "ph": "f", "id": 16325, "pid": 1728719, "tid": 1728719, "ts": 6950381122281.574, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122281.674, "dur": 0.100 , "args": { "External id": 4558,"cbid": 468, "correlation": 16326 } }, { "ph": "f", "id": 16326, "pid": 1728719, "tid": 1728719, "ts": 6950381122281.674, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122281.925, "dur": 0.020 , "args": { "External id": 4558,"cbid": 1, "correlation": 16328 } }, { "ph": "f", "id": 16328, "pid": 1728719, "tid": 1728719, "ts": 6950381122281.925, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122282.025, "dur": 0.100 , "args": { "External id": 4558,"cbid": 468, "correlation": 16329 } }, { "ph": "f", "id": 16329, "pid": 1728719, "tid": 1728719, "ts": 6950381122282.025, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122282.295, "dur": 0.020 , "args": { "External id": 4558,"cbid": 1, "correlation": 16331 } }, { "ph": "f", "id": 16331, "pid": 1728719, "tid": 1728719, "ts": 6950381122282.295, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122282.416, "dur": 0.370 , "args": { "External id": 4558,"cbid": 273, "correlation": 16332 } }, { "ph": "f", "id": 16332, "pid": 1728719, "tid": 1728719, "ts": 6950381122282.416, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381244657.080, "dur": 684.697 , "args": { "External id": 4558,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16335, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16335, "pid": 0, "tid": 7, "ts": 6950381244657.080, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122283.167, "dur": 1.723 , "args": { "External id": 4558,"cbid": 211, "correlation": 16335 } }, { "ph": "s", "id": 16335, "pid": 1728719, "tid": 1728719, "ts": 6950381122283.167, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122294.949, "dur": 0.090 , "args": { "External id": 4570,"cbid": 200, "correlation": 16359 } }, { "ph": "f", "id": 16359, "pid": 1728719, "tid": 1728719, "ts": 6950381122294.949, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122295.139, "dur": 0.121 , "args": { "External id": 4570,"cbid": 468, "correlation": 16360 } }, { "ph": "f", "id": 16360, "pid": 1728719, "tid": 1728719, "ts": 6950381122295.139, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122295.440, "dur": 0.020 , "args": { "External id": 4570,"cbid": 1, "correlation": 16362 } }, { "ph": "f", "id": 16362, "pid": 1728719, "tid": 1728719, "ts": 6950381122295.440, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122295.530, "dur": 0.090 , "args": { "External id": 4570,"cbid": 468, "correlation": 16363 } }, { "ph": "f", "id": 16363, "pid": 1728719, "tid": 1728719, "ts": 6950381122295.530, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122295.801, "dur": 0.020 , "args": { "External id": 4570,"cbid": 1, "correlation": 16365 } }, { "ph": "f", "id": 16365, "pid": 1728719, "tid": 1728719, "ts": 6950381122295.801, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122295.911, "dur": 0.090 , "args": { "External id": 4570,"cbid": 468, "correlation": 16366 } }, { "ph": "f", "id": 16366, "pid": 1728719, "tid": 1728719, "ts": 6950381122295.911, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122296.161, "dur": 0.020 , "args": { "External id": 4570,"cbid": 1, "correlation": 16368 } }, { "ph": "f", "id": 16368, "pid": 1728719, "tid": 1728719, "ts": 6950381122296.161, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122296.271, "dur": 0.331 , "args": { "External id": 4570,"cbid": 273, "correlation": 16369 } }, { "ph": "f", "id": 16369, "pid": 1728719, "tid": 1728719, "ts": 6950381122296.271, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381245342.064, "dur": 696.953 , "args": { "External id": 4570,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16372, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16372, "pid": 0, "tid": 7, "ts": 6950381245342.064, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122296.943, "dur": 1.853 , "args": { "External id": 4570,"cbid": 211, "correlation": 16372 } }, { "ph": "s", "id": 16372, "pid": 1728719, "tid": 1728719, "ts": 6950381122296.943, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381246039.337, "dur": 154.014 , "args": { "External id": 4582,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16385, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16385, "pid": 0, "tid": 7, "ts": 6950381246039.337, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122306.140, "dur": 1.603 , "args": { "External id": 4582,"cbid": 307, "correlation": 16385 } }, { "ph": "s", "id": 16385, "pid": 1728719, "tid": 1728719, "ts": 6950381122306.140, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381246193.639, "dur": 61.951 , "args": { "External id": 4583,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16391, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16391, "pid": 0, "tid": 7, "ts": 6950381246193.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122314.165, "dur": 1.613 , "args": { "External id": 4583,"cbid": 652, "correlation": 16391 } }, { "ph": "s", "id": 16391, "pid": 1728719, "tid": 1728719, "ts": 6950381122314.165, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122325.205, "dur": 0.090 , "args": { "External id": 4584,"cbid": 200, "correlation": 16410 } }, { "ph": "f", "id": 16410, "pid": 1728719, "tid": 1728719, "ts": 6950381122325.205, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122325.416, "dur": 0.120 , "args": { "External id": 4584,"cbid": 468, "correlation": 16411 } }, { "ph": "f", "id": 16411, "pid": 1728719, "tid": 1728719, "ts": 6950381122325.416, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122325.726, "dur": 0.020 , "args": { "External id": 4584,"cbid": 1, "correlation": 16413 } }, { "ph": "f", "id": 16413, "pid": 1728719, "tid": 1728719, "ts": 6950381122325.726, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122325.836, "dur": 0.241 , "args": { "External id": 4584,"cbid": 468, "correlation": 16414 } }, { "ph": "f", "id": 16414, "pid": 1728719, "tid": 1728719, "ts": 6950381122325.836, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122326.217, "dur": 0.020 , "args": { "External id": 4584,"cbid": 1, "correlation": 16416 } }, { "ph": "f", "id": 16416, "pid": 1728719, "tid": 1728719, "ts": 6950381122326.217, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122326.357, "dur": 0.101 , "args": { "External id": 4584,"cbid": 468, "correlation": 16417 } }, { "ph": "f", "id": 16417, "pid": 1728719, "tid": 1728719, "ts": 6950381122326.357, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122326.608, "dur": 0.030 , "args": { "External id": 4584,"cbid": 1, "correlation": 16419 } }, { "ph": "f", "id": 16419, "pid": 1728719, "tid": 1728719, "ts": 6950381122326.608, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122326.738, "dur": 0.371 , "args": { "External id": 4584,"cbid": 273, "correlation": 16420 } }, { "ph": "f", "id": 16420, "pid": 1728719, "tid": 1728719, "ts": 6950381122326.738, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381246255.879, "dur": 694.264 , "args": { "External id": 4584,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16423, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16423, "pid": 0, "tid": 7, "ts": 6950381246255.879, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122327.469, "dur": 1.714 , "args": { "External id": 4584,"cbid": 211, "correlation": 16423 } }, { "ph": "s", "id": 16423, "pid": 1728719, "tid": 1728719, "ts": 6950381122327.469, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381246950.431, "dur": 40.479 , "args": { "External id": 4596,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16434, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16434, "pid": 0, "tid": 7, "ts": 6950381246950.431, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122337.198, "dur": 1.633 , "args": { "External id": 4596,"cbid": 307, "correlation": 16434 } }, { "ph": "s", "id": 16434, "pid": 1728719, "tid": 1728719, "ts": 6950381122337.198, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_unsqueeze_0", "pid": 0, "tid": 7, "ts": 6950381246991.198, "dur": 22.144 , "args": { "External id": 4608,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16446, "registers per thread": 33, "shared memory": 192, "blocks per SM": 21.787233352661133, "warps per SM": 348.5957336425781, "grid": [4096, 1, 1], "block": [512, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 3, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 3, "blockLimitSharedMem": 80, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 1280}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16446, "pid": 0, "tid": 7, "ts": 6950381246991.198, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122415.032, "dur": 2.415 , "args": { "External id": 4608,"cbid": 307, "correlation": 16446 } }, { "ph": "s", "id": 16446, "pid": 1728719, "tid": 1728719, "ts": 6950381122415.032, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381247013.662, "dur": 13.088 , "args": { "External id": 4609,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16460, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16460, "pid": 0, "tid": 7, "ts": 6950381247013.662, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122431.934, "dur": 1.703 , "args": { "External id": 4609,"cbid": 652, "correlation": 16460 } }, { "ph": "s", "id": 16460, "pid": 1728719, "tid": 1728719, "ts": 6950381122431.934, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381247027.070, "dur": 12.736 , "args": { "External id": 4610,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16474, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16474, "pid": 0, "tid": 7, "ts": 6950381247027.070, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122443.175, "dur": 1.553 , "args": { "External id": 4610,"cbid": 652, "correlation": 16474 } }, { "ph": "s", "id": 16474, "pid": 1728719, "tid": 1728719, "ts": 6950381122443.175, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381247040.094, "dur": 10.752 , "args": { "External id": 4611,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16488, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16488, "pid": 0, "tid": 7, "ts": 6950381247040.094, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122453.654, "dur": 1.613 , "args": { "External id": 4611,"cbid": 652, "correlation": 16488 } }, { "ph": "s", "id": 16488, "pid": 1728719, "tid": 1728719, "ts": 6950381122453.654, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122465.396, "dur": 0.151 , "args": { "External id": 4612,"cbid": 200, "correlation": 16503 } }, { "ph": "f", "id": 16503, "pid": 1728719, "tid": 1728719, "ts": 6950381122465.396, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122465.687, "dur": 0.250 , "args": { "External id": 4612,"cbid": 468, "correlation": 16504 } }, { "ph": "f", "id": 16504, "pid": 1728719, "tid": 1728719, "ts": 6950381122465.687, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122466.278, "dur": 0.030 , "args": { "External id": 4612,"cbid": 1, "correlation": 16506 } }, { "ph": "f", "id": 16506, "pid": 1728719, "tid": 1728719, "ts": 6950381122466.278, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122466.388, "dur": 0.100 , "args": { "External id": 4612,"cbid": 468, "correlation": 16507 } }, { "ph": "f", "id": 16507, "pid": 1728719, "tid": 1728719, "ts": 6950381122466.388, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122466.639, "dur": 0.020 , "args": { "External id": 4612,"cbid": 1, "correlation": 16509 } }, { "ph": "f", "id": 16509, "pid": 1728719, "tid": 1728719, "ts": 6950381122466.639, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122466.759, "dur": 0.100 , "args": { "External id": 4612,"cbid": 468, "correlation": 16510 } }, { "ph": "f", "id": 16510, "pid": 1728719, "tid": 1728719, "ts": 6950381122466.759, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122467.019, "dur": 0.020 , "args": { "External id": 4612,"cbid": 1, "correlation": 16512 } }, { "ph": "f", "id": 16512, "pid": 1728719, "tid": 1728719, "ts": 6950381122467.019, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122467.190, "dur": 0.631 , "args": { "External id": 4612,"cbid": 273, "correlation": 16513 } }, { "ph": "f", "id": 16513, "pid": 1728719, "tid": 1728719, "ts": 6950381122467.190, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381247051.134, "dur": 222.237 , "args": { "External id": 4612,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16516, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16516, "pid": 0, "tid": 7, "ts": 6950381247051.134, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122468.262, "dur": 1.993 , "args": { "External id": 4612,"cbid": 211, "correlation": 16516 } }, { "ph": "s", "id": 16516, "pid": 1728719, "tid": 1728719, "ts": 6950381122468.262, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381247273.691, "dur": 8.896 , "args": { "External id": 4624,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16529, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16529, "pid": 0, "tid": 7, "ts": 6950381247273.691, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122478.801, "dur": 1.733 , "args": { "External id": 4624,"cbid": 307, "correlation": 16529 } }, { "ph": "s", "id": 16529, "pid": 1728719, "tid": 1728719, "ts": 6950381122478.801, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381247282.875, "dur": 38.144 , "args": { "External id": 4629,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16552, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16552, "pid": 0, "tid": 7, "ts": 6950381247282.875, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122497.396, "dur": 1.743 , "args": { "External id": 4629,"cbid": 307, "correlation": 16552 } }, { "ph": "s", "id": 16552, "pid": 1728719, "tid": 1728719, "ts": 6950381122497.396, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122507.515, "dur": 0.090 , "args": { "External id": 4630,"cbid": 200, "correlation": 16567 } }, { "ph": "f", "id": 16567, "pid": 1728719, "tid": 1728719, "ts": 6950381122507.515, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122507.715, "dur": 0.150 , "args": { "External id": 4630,"cbid": 468, "correlation": 16568 } }, { "ph": "f", "id": 16568, "pid": 1728719, "tid": 1728719, "ts": 6950381122507.715, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122508.096, "dur": 0.060 , "args": { "External id": 4630,"cbid": 1, "correlation": 16570 } }, { "ph": "f", "id": 16570, "pid": 1728719, "tid": 1728719, "ts": 6950381122508.096, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122508.226, "dur": 0.100 , "args": { "External id": 4630,"cbid": 468, "correlation": 16571 } }, { "ph": "f", "id": 16571, "pid": 1728719, "tid": 1728719, "ts": 6950381122508.226, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122508.476, "dur": 0.020 , "args": { "External id": 4630,"cbid": 1, "correlation": 16573 } }, { "ph": "f", "id": 16573, "pid": 1728719, "tid": 1728719, "ts": 6950381122508.476, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122508.597, "dur": 0.090 , "args": { "External id": 4630,"cbid": 468, "correlation": 16574 } }, { "ph": "f", "id": 16574, "pid": 1728719, "tid": 1728719, "ts": 6950381122508.597, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122508.847, "dur": 0.020 , "args": { "External id": 4630,"cbid": 1, "correlation": 16576 } }, { "ph": "f", "id": 16576, "pid": 1728719, "tid": 1728719, "ts": 6950381122508.847, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122508.967, "dur": 0.431 , "args": { "External id": 4630,"cbid": 273, "correlation": 16577 } }, { "ph": "f", "id": 16577, "pid": 1728719, "tid": 1728719, "ts": 6950381122508.967, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381247321.307, "dur": 232.285 , "args": { "External id": 4630,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16580, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16580, "pid": 0, "tid": 7, "ts": 6950381247321.307, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122509.769, "dur": 1.813 , "args": { "External id": 4630,"cbid": 211, "correlation": 16580 } }, { "ph": "s", "id": 16580, "pid": 1728719, "tid": 1728719, "ts": 6950381122509.769, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mean_pow_view_1", "pid": 0, "tid": 7, "ts": 6950381247553.912, "dur": 8.928 , "args": { "External id": 4642,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16589, "registers per thread": 20, "shared memory": 32, "blocks per SM": 87.14893341064453, "warps per SM": 174.29786682128906, "grid": [16384, 1, 1], "block": [64, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 24, "limitingFactors": "WARPS|BLOCKS", "blockLimitRegs": 42, "blockLimitSharedMem": 88, "blockLimitWarps": 24, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 1536, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16589, "pid": 0, "tid": 7, "ts": 6950381247553.912, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122517.473, "dur": 1.633 , "args": { "External id": 4642,"cbid": 307, "correlation": 16589 } }, { "ph": "s", "id": 16589, "pid": 1728719, "tid": 1728719, "ts": 6950381122517.473, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_mean_mul_pow_rsqrt_select_sub_unsqueeze_view_2", "pid": 0, "tid": 7, "ts": 6950381247563.128, "dur": 26.336 , "args": { "External id": 4647,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16612, "registers per thread": 30, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 697.1914672851562, "grid": [16384, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 8, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 8192, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16612, "pid": 0, "tid": 7, "ts": 6950381247563.128, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122534.365, "dur": 1.653 , "args": { "External id": 4647,"cbid": 307, "correlation": 16612 } }, { "ph": "s", "id": 16612, "pid": 1728719, "tid": 1728719, "ts": 6950381122534.365, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122548.130, "dur": 0.090 , "args": { "External id": 4648,"cbid": 200, "correlation": 16627 } }, { "ph": "f", "id": 16627, "pid": 1728719, "tid": 1728719, "ts": 6950381122548.130, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122548.341, "dur": 0.140 , "args": { "External id": 4648,"cbid": 468, "correlation": 16628 } }, { "ph": "f", "id": 16628, "pid": 1728719, "tid": 1728719, "ts": 6950381122548.341, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122548.771, "dur": 0.030 , "args": { "External id": 4648,"cbid": 1, "correlation": 16630 } }, { "ph": "f", "id": 16630, "pid": 1728719, "tid": 1728719, "ts": 6950381122548.771, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122548.872, "dur": 0.100 , "args": { "External id": 4648,"cbid": 468, "correlation": 16631 } }, { "ph": "f", "id": 16631, "pid": 1728719, "tid": 1728719, "ts": 6950381122548.872, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122549.142, "dur": 0.020 , "args": { "External id": 4648,"cbid": 1, "correlation": 16633 } }, { "ph": "f", "id": 16633, "pid": 1728719, "tid": 1728719, "ts": 6950381122549.142, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122549.252, "dur": 0.101 , "args": { "External id": 4648,"cbid": 468, "correlation": 16634 } }, { "ph": "f", "id": 16634, "pid": 1728719, "tid": 1728719, "ts": 6950381122549.252, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122549.503, "dur": 0.020 , "args": { "External id": 4648,"cbid": 1, "correlation": 16636 } }, { "ph": "f", "id": 16636, "pid": 1728719, "tid": 1728719, "ts": 6950381122549.503, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122549.643, "dur": 0.411 , "args": { "External id": 4648,"cbid": 273, "correlation": 16637 } }, { "ph": "f", "id": 16637, "pid": 1728719, "tid": 1728719, "ts": 6950381122549.643, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381247589.752, "dur": 236.413 , "args": { "External id": 4648,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16640, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16640, "pid": 0, "tid": 7, "ts": 6950381247589.752, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122550.435, "dur": 1.733 , "args": { "External id": 4648,"cbid": 211, "correlation": 16640 } }, { "ph": "s", "id": 16640, "pid": 1728719, "tid": 1728719, "ts": 6950381122550.435, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_3", "pid": 0, "tid": 7, "ts": 6950381247826.485, "dur": 60.096 , "args": { "External id": 4660,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16654, "registers per thread": 16, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 1394.3829345703125, "grid": [32768, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16654, "pid": 0, "tid": 7, "ts": 6950381247826.485, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122559.942, "dur": 1.663 , "args": { "External id": 4660,"cbid": 307, "correlation": 16654 } }, { "ph": "s", "id": 16654, "pid": 1728719, "tid": 1728719, "ts": 6950381122559.942, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_4", "pid": 0, "tid": 7, "ts": 6950381247886.837, "dur": 63.295 , "args": { "External id": 4661,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16664, "registers per thread": 34, "shared memory": 0, "blocks per SM": 87.85106658935547, "warps per SM": 351.4042663574219, "grid": [16516, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16664, "pid": 0, "tid": 7, "ts": 6950381247886.837, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122570.221, "dur": 1.683 , "args": { "External id": 4661,"cbid": 307, "correlation": 16664 } }, { "ph": "s", "id": 16664, "pid": 1728719, "tid": 1728719, "ts": 6950381122570.221, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused__scaled_dot_product_flash_attention__to_copy_cat_stack_transpose_view_5", "pid": 0, "tid": 7, "ts": 6950381247950.420, "dur": 38.752 , "args": { "External id": 4662,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16673, "registers per thread": 16, "shared memory": 0, "blocks per SM": 175.70213317871094, "warps per SM": 1405.6170654296875, "grid": [33032, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16673, "pid": 0, "tid": 7, "ts": 6950381247950.420, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122579.509, "dur": 1.703 , "args": { "External id": 4662,"cbid": 307, "correlation": 16673 } }, { "ph": "s", "id": 16673, "pid": 1728719, "tid": 1728719, "ts": 6950381122579.509, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122595.348, "dur": 0.110 , "args": { "External id": 4670,"cbid": 200, "correlation": 16708 } }, { "ph": "f", "id": 16708, "pid": 1728719, "tid": 1728719, "ts": 6950381122595.348, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122595.528, "dur": 0.071 , "args": { "External id": 4670,"cbid": 200, "correlation": 16709 } }, { "ph": "f", "id": 16709, "pid": 1728719, "tid": 1728719, "ts": 6950381122595.528, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122595.679, "dur": 0.461 , "args": { "External id": 4670,"cbid": 273, "correlation": 16710 } }, { "ph": "f", "id": 16710, "pid": 1728719, "tid": 1728719, "ts": 6950381122595.679, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7, "ts": 6950381247989.428, "dur": 857.303 , "args": { "External id": 4670,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16712, "registers per thread": 255, "shared memory": 65536, "blocks per SM": 5.446808338165283, "warps per SM": 21.787233352661133, "grid": [32, 1, 32], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 2, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 32768, "allocatedSharedMemPerBlock": 66560}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16712, "pid": 0, "tid": 7, "ts": 6950381247989.428, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122596.430, "dur": 1.924 , "args": { "External id": 4670,"cbid": 211, "correlation": 16712 } }, { "ph": "s", "id": 16712, "pid": 1728719, "tid": 1728719, "ts": 6950381122596.430, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381248846.986, "dur": 9.408 , "args": { "External id": 4679,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16732, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16732, "pid": 0, "tid": 7, "ts": 6950381248846.986, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122613.051, "dur": 1.603 , "args": { "External id": 4679,"cbid": 652, "correlation": 16732 } }, { "ph": "s", "id": 16732, "pid": 1728719, "tid": 1728719, "ts": 6950381122613.051, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122624.132, "dur": 0.100 , "args": { "External id": 4680,"cbid": 200, "correlation": 16747 } }, { "ph": "f", "id": 16747, "pid": 1728719, "tid": 1728719, "ts": 6950381122624.132, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122624.342, "dur": 0.170 , "args": { "External id": 4680,"cbid": 468, "correlation": 16748 } }, { "ph": "f", "id": 16748, "pid": 1728719, "tid": 1728719, "ts": 6950381122624.342, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122624.763, "dur": 0.050 , "args": { "External id": 4680,"cbid": 1, "correlation": 16750 } }, { "ph": "f", "id": 16750, "pid": 1728719, "tid": 1728719, "ts": 6950381122624.763, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122624.883, "dur": 0.100 , "args": { "External id": 4680,"cbid": 468, "correlation": 16751 } }, { "ph": "f", "id": 16751, "pid": 1728719, "tid": 1728719, "ts": 6950381122624.883, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122625.124, "dur": 0.030 , "args": { "External id": 4680,"cbid": 1, "correlation": 16753 } }, { "ph": "f", "id": 16753, "pid": 1728719, "tid": 1728719, "ts": 6950381122625.124, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122625.244, "dur": 0.100 , "args": { "External id": 4680,"cbid": 468, "correlation": 16754 } }, { "ph": "f", "id": 16754, "pid": 1728719, "tid": 1728719, "ts": 6950381122625.244, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122625.504, "dur": 0.020 , "args": { "External id": 4680,"cbid": 1, "correlation": 16756 } }, { "ph": "f", "id": 16756, "pid": 1728719, "tid": 1728719, "ts": 6950381122625.504, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122625.614, "dur": 0.391 , "args": { "External id": 4680,"cbid": 273, "correlation": 16757 } }, { "ph": "f", "id": 16757, "pid": 1728719, "tid": 1728719, "ts": 6950381122625.614, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381248856.650, "dur": 211.198 , "args": { "External id": 4680,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16760, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16760, "pid": 0, "tid": 7, "ts": 6950381248856.650, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122626.386, "dur": 1.733 , "args": { "External id": 4680,"cbid": 211, "correlation": 16760 } }, { "ph": "s", "id": 16760, "pid": 1728719, "tid": 1728719, "ts": 6950381122626.386, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_red_fused_add_mul_native_layer_norm_slice_split_tanh_unsqueeze_view_6", "pid": 0, "tid": 7, "ts": 6950381249068.136, "dur": 47.712 , "args": { "External id": 4692,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16772, "registers per thread": 40, "shared memory": 96, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16772, "pid": 0, "tid": 7, "ts": 6950381249068.136, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122636.124, "dur": 1.683 , "args": { "External id": 4692,"cbid": 307, "correlation": 16772 } }, { "ph": "s", "id": 16772, "pid": 1728719, "tid": 1728719, "ts": 6950381122636.124, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381249116.168, "dur": 13.119 , "args": { "External id": 4693,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16778, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16778, "pid": 0, "tid": 7, "ts": 6950381249116.168, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122644.229, "dur": 1.513 , "args": { "External id": 4693,"cbid": 652, "correlation": 16778 } }, { "ph": "s", "id": 16778, "pid": 1728719, "tid": 1728719, "ts": 6950381122644.229, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381249129.607, "dur": 13.088 , "args": { "External id": 4694,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16784, "registers per thread": 36, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 6, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 10240, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16784, "pid": 0, "tid": 7, "ts": 6950381249129.607, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122652.665, "dur": 1.483 , "args": { "External id": 4694,"cbid": 652, "correlation": 16784 } }, { "ph": "s", "id": 16784, "pid": 1728719, "tid": 1728719, "ts": 6950381122652.665, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122664.106, "dur": 0.090 , "args": { "External id": 4695,"cbid": 200, "correlation": 16803 } }, { "ph": "f", "id": 16803, "pid": 1728719, "tid": 1728719, "ts": 6950381122664.106, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122664.306, "dur": 0.111 , "args": { "External id": 4695,"cbid": 468, "correlation": 16804 } }, { "ph": "f", "id": 16804, "pid": 1728719, "tid": 1728719, "ts": 6950381122664.306, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122664.617, "dur": 0.020 , "args": { "External id": 4695,"cbid": 1, "correlation": 16806 } }, { "ph": "f", "id": 16806, "pid": 1728719, "tid": 1728719, "ts": 6950381122664.617, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122664.707, "dur": 0.100 , "args": { "External id": 4695,"cbid": 468, "correlation": 16807 } }, { "ph": "f", "id": 16807, "pid": 1728719, "tid": 1728719, "ts": 6950381122664.707, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122664.958, "dur": 0.020 , "args": { "External id": 4695,"cbid": 1, "correlation": 16809 } }, { "ph": "f", "id": 16809, "pid": 1728719, "tid": 1728719, "ts": 6950381122664.958, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122665.068, "dur": 0.100 , "args": { "External id": 4695,"cbid": 468, "correlation": 16810 } }, { "ph": "f", "id": 16810, "pid": 1728719, "tid": 1728719, "ts": 6950381122665.068, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122665.338, "dur": 0.020 , "args": { "External id": 4695,"cbid": 1, "correlation": 16812 } }, { "ph": "f", "id": 16812, "pid": 1728719, "tid": 1728719, "ts": 6950381122665.338, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122665.449, "dur": 0.380 , "args": { "External id": 4695,"cbid": 273, "correlation": 16813 } }, { "ph": "f", "id": 16813, "pid": 1728719, "tid": 1728719, "ts": 6950381122665.449, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381249143.015, "dur": 681.241 , "args": { "External id": 4695,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16816, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16816, "pid": 0, "tid": 7, "ts": 6950381249143.015, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122666.190, "dur": 1.833 , "args": { "External id": 4695,"cbid": 211, "correlation": 16816 } }, { "ph": "s", "id": 16816, "pid": 1728719, "tid": 1728719, "ts": 6950381122666.190, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122677.992, "dur": 0.080 , "args": { "External id": 4707,"cbid": 200, "correlation": 16840 } }, { "ph": "f", "id": 16840, "pid": 1728719, "tid": 1728719, "ts": 6950381122677.992, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122678.162, "dur": 0.120 , "args": { "External id": 4707,"cbid": 468, "correlation": 16841 } }, { "ph": "f", "id": 16841, "pid": 1728719, "tid": 1728719, "ts": 6950381122678.162, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122678.463, "dur": 0.020 , "args": { "External id": 4707,"cbid": 1, "correlation": 16843 } }, { "ph": "f", "id": 16843, "pid": 1728719, "tid": 1728719, "ts": 6950381122678.463, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122678.553, "dur": 0.100 , "args": { "External id": 4707,"cbid": 468, "correlation": 16844 } }, { "ph": "f", "id": 16844, "pid": 1728719, "tid": 1728719, "ts": 6950381122678.553, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122678.793, "dur": 0.020 , "args": { "External id": 4707,"cbid": 1, "correlation": 16846 } }, { "ph": "f", "id": 16846, "pid": 1728719, "tid": 1728719, "ts": 6950381122678.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122678.904, "dur": 0.110 , "args": { "External id": 4707,"cbid": 468, "correlation": 16847 } }, { "ph": "f", "id": 16847, "pid": 1728719, "tid": 1728719, "ts": 6950381122678.904, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122679.184, "dur": 0.020 , "args": { "External id": 4707,"cbid": 1, "correlation": 16849 } }, { "ph": "f", "id": 16849, "pid": 1728719, "tid": 1728719, "ts": 6950381122679.184, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122679.284, "dur": 0.361 , "args": { "External id": 4707,"cbid": 273, "correlation": 16850 } }, { "ph": "f", "id": 16850, "pid": 1728719, "tid": 1728719, "ts": 6950381122679.284, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381249824.576, "dur": 694.968 , "args": { "External id": 4707,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16853, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 16.340425491333008, "warps per SM": 196.08511352539062, "grid": [32, 96, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16853, "pid": 0, "tid": 7, "ts": 6950381249824.576, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122679.996, "dur": 1.793 , "args": { "External id": 4707,"cbid": 211, "correlation": 16853 } }, { "ph": "s", "id": 16853, "pid": 1728719, "tid": 1728719, "ts": 6950381122679.996, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_mul_silu_view_7", "pid": 0, "tid": 7, "ts": 6950381250519.832, "dur": 157.151 , "args": { "External id": 4719,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16866, "registers per thread": 16, "shared memory": 0, "blocks per SM": 522.8936157226562, "warps per SM": 4183.14892578125, "grid": [98304, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 6, "limitingFactors": "WARPS", "blockLimitRegs": 16, "blockLimitSharedMem": 100, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 4096, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16866, "pid": 0, "tid": 7, "ts": 6950381250519.832, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122689.523, "dur": 1.623 , "args": { "External id": 4719,"cbid": 307, "correlation": 16866 } }, { "ph": "s", "id": 16866, "pid": 1728719, "tid": 1728719, "ts": 6950381122689.523, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "_quantize_rows", "pid": 0, "tid": 7, "ts": 6950381250677.239, "dur": 57.727 , "args": { "External id": 4720,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16872, "registers per thread": 64, "shared memory": 32, "blocks per SM": 21.787233352661133, "warps per SM": 174.29786682128906, "grid": [4096, 1, 1], "block": [256, 1, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 4, "limitingFactors": "REGS", "blockLimitRegs": 4, "blockLimitSharedMem": 88, "blockLimitWarps": 6, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 16384, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16872, "pid": 0, "tid": 7, "ts": 6950381250677.239, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 1728719, "tid": 1728719, "ts": 6950381122697.839, "dur": 1.653 , "args": { "External id": 4720,"cbid": 652, "correlation": 16872 } }, { "ph": "s", "id": 16872, "pid": 1728719, "tid": 1728719, "ts": 6950381122697.839, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122709.260, "dur": 0.090 , "args": { "External id": 4721,"cbid": 200, "correlation": 16891 } }, { "ph": "f", "id": 16891, "pid": 1728719, "tid": 1728719, "ts": 6950381122709.260, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122709.440, "dur": 0.121 , "args": { "External id": 4721,"cbid": 468, "correlation": 16892 } }, { "ph": "f", "id": 16892, "pid": 1728719, "tid": 1728719, "ts": 6950381122709.440, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122709.761, "dur": 0.020 , "args": { "External id": 4721,"cbid": 1, "correlation": 16894 } }, { "ph": "f", "id": 16894, "pid": 1728719, "tid": 1728719, "ts": 6950381122709.761, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122709.851, "dur": 0.100 , "args": { "External id": 4721,"cbid": 468, "correlation": 16895 } }, { "ph": "f", "id": 16895, "pid": 1728719, "tid": 1728719, "ts": 6950381122709.851, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122710.102, "dur": 0.020 , "args": { "External id": 4721,"cbid": 1, "correlation": 16897 } }, { "ph": "f", "id": 16897, "pid": 1728719, "tid": 1728719, "ts": 6950381122710.102, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPointByVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122710.212, "dur": 0.090 , "args": { "External id": 4721,"cbid": 468, "correlation": 16898 } }, { "ph": "f", "id": 16898, "pid": 1728719, "tid": 1728719, "ts": 6950381122710.212, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDriverGetVersion", "pid": 1728719, "tid": 1728719, "ts": 6950381122710.462, "dur": 0.020 , "args": { "External id": 4721,"cbid": 1, "correlation": 16900 } }, { "ph": "f", "id": 16900, "pid": 1728719, "tid": 1728719, "ts": 6950381122710.462, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122710.573, "dur": 0.360 , "args": { "External id": 4721,"cbid": 273, "correlation": 16901 } }, { "ph": "f", "id": 16901, "pid": 1728719, "tid": 1728719, "ts": 6950381122710.573, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "_ZN7cutlass13device_kernelIN2at4cuda6detail34enable_3x_kernel_for_sm10_or_laterINS_4gemm6kernel13GemmUniversalIN4cute5tupleIJiiiEEENS5_10collective13CollectiveMmaINS5_31MainloopSm120TmaWarpSpecializedILi2ELi2ENS9_IJNS8_1CILi1EEESF_SF_EEENS5_40KernelTmaWarpSpecializedCooperativeSm120ILi2EEEEENS9_IJNSE_ILi128EEESK_SK_EEENS_12float_e4m3_tENS9_IJlSF_lEEESM_SN_NS8_8TiledMMAINS8_8MMA_AtomIJNS8_16SM120_16x8x32_TNISM_SM_fEEEEENS8_6LayoutINS9_IJNSE_ILi4EEENSE_ILi2EEESF_EEENS9_IJSF_SU_NSE_ILi0EEEEEEEENS9_IJSK_NSE_ILi32EEES10_EEEEENS8_13SM90_TMA_LOADENS8_14ComposedLayoutINS8_7SwizzleILi3ELi4ELi3EEENS8_18smem_ptr_flag_bitsILi8EEENST_INS9_IJNSE_ILi8EEESK_EEENS9_IJSK_SF_EEEEEEENS8_9Copy_AtomIJNS8_17SM75_U32x4_LDSM_NEhEEENS8_8identityES13_S1D_S1G_S1H_EENS_8epilogue10collective18CollectiveEpilogueINS1J_22Sm90TmaWarpSpecializedILi2ELi2ELi4ELb0ELb1EEEJSL_NS9_IJNSE_ILi64EEES10_EEEvSN_NS_10bfloat16_tESN_NS1J_6fusion15Sm90TreeVisitorINS1R_11Sm90ComputeINS1J_6thread8IdentityES1Q_fLNS_15FloatRoundStyleE2EvEEJNS1S_INS1T_INS_4plusEffLS1W_2EvEEJNS1R_16Sm90RowBroadcastILi0ESL_ffNS9_IJSX_SF_SX_EEELi4ELb1EEENS1S_INS1T_INS_10multipliesEffLS1W_2EvEEJS22_NS1S_IS24_JNS1R_16Sm90ColBroadcastILi0ESL_ffNS9_IJSF_SX_SX_EEELi4ELb1EEENS1R_12Sm90AccFetchEEEEEEEEEEEEES13_NS14_INS15_ILi2ELi4ELi3EEENS17_ILi16EEENST_INS9_IJS19_S10_EEENS9_IJS10_SF_EEEEEEENS8_17SM75_U32x2_LDSM_NENS8_14SM90_TMA_STOREES2I_NS8_17SM90_U32x2_STSM_NENS1E_IJS2L_NS_6half_tEEEEvEEEvvEEEEEEvNT_6ParamsE", "pid": 0, "tid": 7, "ts": 6950381250735.254, "dur": 700.056 , "args": { "External id": 4721,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16904, "registers per thread": 168, "shared memory": 75776, "blocks per SM": 5.446808338165283, "warps per SM": 65.36170196533203, "grid": [32, 32, 1], "block": [384, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 1, "blockLimitSharedMem": 0, "blockLimitWarps": 4, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 64512, "allocatedSharedMemPerBlock": 76800}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16904, "pid": 0, "tid": 7, "ts": 6950381250735.254, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122711.294, "dur": 1.833 , "args": { "External id": 4721,"cbid": 211, "correlation": 16904 } }, { "ph": "s", "id": 16904, "pid": 1728719, "tid": 1728719, "ts": 6950381122711.294, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_slice_split_tanh_unsqueeze_view_8", "pid": 0, "tid": 7, "ts": 6950381251435.598, "dur": 40.128 , "args": { "External id": 4733,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16915, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16915, "pid": 0, "tid": 7, "ts": 6950381251435.598, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122721.132, "dur": 1.703 , "args": { "External id": 4733,"cbid": 307, "correlation": 16915 } }, { "ph": "s", "id": 16915, "pid": 1728719, "tid": 1728719, "ts": 6950381122721.132, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#6}::operator()() const::{lambda(c10::BFloat16)#1}, std::array, false>(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#6}::operator()() const::{lambda(c10::BFloat16)#1}, std::array)", "pid": 0, "tid": 7, "ts": 6950381251475.982, "dur": 2.784 , "args": { "External id": 4734,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16924, "registers per thread": 38, "shared memory": 0, "blocks per SM": 0.042553190141916275, "warps per SM": 0.1702127605676651, "grid": [8, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16924, "pid": 0, "tid": 7, "ts": 6950381251475.982, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122743.684, "dur": 2.425 , "args": { "External id": 4734,"cbid": 211, "correlation": 16924 } }, { "ph": "s", "id": 16924, "pid": 1728719, "tid": 1728719, "ts": 6950381122743.684, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122765.705, "dur": 0.231 , "args": { "External id": 4741,"cbid": 200, "correlation": 16939 } }, { "ph": "f", "id": 16939, "pid": 1728719, "tid": 1728719, "ts": 6950381122765.705, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7, "ts": 6950381251479.310, "dur": 0.320 , "args": { "External id": 4741,"device": 0, "context": 1, "stream": 7, "correlation": 16941, "bytes": 512, "memory bandwidth (GB/s)": 1.6, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16941, "pid": 0, "tid": 7, "ts": 6950381251479.310, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaMemsetAsync", "pid": 1728719, "tid": 1728719, "ts": 6950381122766.797, "dur": 5.150 , "args": { "External id": 4741,"cbid": 51, "correlation": 16941 } }, { "ph": "s", "id": 16941, "pid": 1728719, "tid": 1728719, "ts": 6950381122766.797, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x1_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381251480.782, "dur": 26.624 , "args": { "External id": 4741,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16942, "registers per thread": 72, "shared memory": 17408, "blocks per SM": 2.7234041690826416, "warps per SM": 10.893616676330566, "grid": [8, 16, 4], "block": [128, 1, 1], "est. achieved occupancy %": 23, "occupancy": {"activeBlocksPerMultiprocessor": 5, "limitingFactors": "SMEM", "blockLimitRegs": 7, "blockLimitSharedMem": 5, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 9216, "allocatedSharedMemPerBlock": 18432}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16942, "pid": 0, "tid": 7, "ts": 6950381251480.782, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122772.097, "dur": 2.405 , "args": { "External id": 4741,"cbid": 307, "correlation": 16942 } }, { "ph": "s", "id": 16942, "pid": 1728719, "tid": 1728719, "ts": 6950381122772.097, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::(anonymous namespace)::where_kernel_impl(at::TensorIterator&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(bool, unsigned short, unsigned short)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::(anonymous namespace)::where_kernel_impl(at::TensorIterator&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(bool, unsigned short, unsigned short)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, "ts": 6950381251507.726, "dur": 23.231 , "args": { "External id": 4751,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16965, "registers per thread": 23, "shared memory": 0, "blocks per SM": 174.29786682128906, "warps per SM": 697.1914672851562, "grid": [32768, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS", "blockLimitRegs": 21, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 3072, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16965, "pid": 0, "tid": 7, "ts": 6950381251507.726, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122800.009, "dur": 2.996 , "args": { "External id": 4751,"cbid": 211, "correlation": 16965 } }, { "ph": "s", "id": 16965, "pid": 1728719, "tid": 1728719, "ts": 6950381122800.009, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::vectorized_layer_norm_kernel(int, float, c10::BFloat16 const*, c10::BFloat16 const*, c10::BFloat16 const*, float*, float*, c10::BFloat16*)", "pid": 0, "tid": 7, "ts": 6950381251531.214, "dur": 15.263 , "args": { "External id": 4755,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 16995, "registers per thread": 40, "shared memory": 24, "blocks per SM": 21.787233352661133, "warps per SM": 87.14893341064453, "grid": [4096, 1, 1], "block": [32, 4, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 88, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1152}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 16995, "pid": 0, "tid": 7, "ts": 6950381251531.214, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122825.537, "dur": 2.554 , "args": { "External id": 4755,"cbid": 211, "correlation": 16995 } }, { "ph": "s", "id": 16995, "pid": 1728719, "tid": 1728719, "ts": 6950381122825.537, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, std::array, false>(int, at::native::CUDAFunctorOnSelf_add, std::array)", "pid": 0, "tid": 7, "ts": 6950381251546.733, "dur": 15.904 , "args": { "External id": 4761,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 17006, "registers per thread": 40, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 100, "occupancy": {"activeBlocksPerMultiprocessor": 12, "limitingFactors": "WARPS|REGS", "blockLimitRegs": 12, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 5120, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 17006, "pid": 0, "tid": 7, "ts": 6950381251546.733, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122839.082, "dur": 2.354 , "args": { "External id": 4761,"cbid": 211, "correlation": 17006 } }, { "ph": "s", "id": 17006, "pid": 1728719, "tid": 1728719, "ts": 6950381122839.082, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, std::array, false>(int, at::native::BinaryFunctor >, std::array)", "pid": 0, "tid": 7, "ts": 6950381251562.925, "dur": 15.232 , "args": { "External id": 4762,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 17017, "registers per thread": 46, "shared memory": 0, "blocks per SM": 87.14893341064453, "warps per SM": 348.5957336425781, "grid": [16384, 1, 1], "block": [128, 1, 1], "est. achieved occupancy %": 83, "occupancy": {"activeBlocksPerMultiprocessor": 10, "limitingFactors": "REGS", "blockLimitRegs": 10, "blockLimitSharedMem": 100, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 6144, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 17017, "pid": 0, "tid": 7, "ts": 6950381251562.925, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122846.886, "dur": 2.375 , "args": { "External id": 4762,"cbid": 211, "correlation": 17017 } }, { "ph": "s", "id": 17017, "pid": 1728719, "tid": 1728719, "ts": 6950381122846.886, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 1728719, "tid": 1728719, "ts": 6950381122866.793, "dur": 0.161 , "args": { "External id": 4770,"cbid": 200, "correlation": 17032 } }, { "ph": "f", "id": 17032, "pid": 1728719, "tid": 1728719, "ts": 6950381122866.793, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x128_64x3_tn_align8::Params)", "pid": 0, "tid": 7, "ts": 6950381251578.445, "dur": 11.232 , "args": { "External id": 4770,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 17035, "registers per thread": 158, "shared memory": 73728, "blocks per SM": 0.8510638475418091, "warps per SM": 3.4042553901672363, "grid": [32, 1, 5], "block": [128, 1, 1], "est. achieved occupancy %": 0, "occupancy": {"activeBlocksPerMultiprocessor": 0, "limitingFactors": "SMEM", "blockLimitRegs": 3, "blockLimitSharedMem": 0, "blockLimitWarps": 12, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 20480, "allocatedSharedMemPerBlock": 74752}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 17035, "pid": 0, "tid": 7, "ts": 6950381251578.445, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 1728719, "tid": 1728719, "ts": 6950381122867.905, "dur": 2.285 , "args": { "External id": 4770,"cbid": 307, "correlation": 17035 } }, { "ph": "s", "id": 17035, "pid": 1728719, "tid": 1728719, "ts": 6950381122867.905, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7, "ts": 6950381251589.965, "dur": 2.720 , "args": { "External id": 4770,"queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 17037, "registers per thread": 42, "shared memory": 0, "blocks per SM": 2.7234041690826416, "warps per SM": 43.574466705322266, "grid": [2, 256, 1], "block": [32, 16, 1], "est. achieved occupancy %": 67, "occupancy": {"activeBlocksPerMultiprocessor": 2, "limitingFactors": "REGS", "blockLimitRegs": 2, "blockLimitSharedMem": 100, "blockLimitWarps": 3, "blockLimitBlocks": 24, "blockLimitBarriers": 2147483647, "allocatedRegistersPerBlock": 24576, "allocatedSharedMemPerBlock": 1024}, "graph id": 0, "graph node id": 0, "channel": 0, "channel_type": 1 } }, { "ph": "f", "id": 17037, "pid": 0, "tid": 7, "ts": 6950381251589.965, "cat": "ac2g", "name": "ac2g" , "bp": "e" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernelExC", "pid": 1728719, "tid": 1728719, "ts": 6950381122871.091, "dur": 2.455 , "args": { "External id": 4770,"cbid": 430, "correlation": 17037 } }, { "ph": "s", "id": 17037, "pid": 1728719, "tid": 1728719, "ts": 6950381122871.091, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381122908.401, "dur": 128686.642 , "args": { "cbid": 165, "correlation": 17043 } }, { "ph": "s", "id": 17043, "pid": 1728719, "tid": 1728719, "ts": 6950381122908.401, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 1728719, "tid": 1728719, "ts": 6950381251613.868, "dur": 2.084 , "args": { "cbid": 165, "correlation": 17047 } }, { "ph": "s", "id": 17047, "pid": 1728719, "tid": 1728719, "ts": 6950381251613.868, "cat": "ac2g", "name": "ac2g" }, { "ph": "X", "cat": "Trace", "name": "PyTorch Profiler (0)", "pid": "Spans", "tid": "PyTorch Profiler", "ts": 6950381108123.930, "dur": 143498.072 , "args": { "Op count": 0 } }, { "ph": "M", "name": "process_sort_index", "ts": 6950381108123.930, "pid": "Spans", "tid": 0, "args": { "sort_index": 536870912 } }, { "ph": "i", "s": "g", "name": "Iteration Start: PyTorch Profiler", "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 6950381108123.930 }, { "ph": "i", "s": "g", "name": "Record Window End", "pid": "", "tid": "", "ts": 6950381252670.894 } ],"traceName": "/home/user/.local/share/rtx-pro-apps/qwen-image-2.1-fp8/optimization-20260920/final-optimized/denoising-trace.json" }