test: add Ascend NPU UT for torch.autograd.profiler.parse_nvprof

已合并

test: add Ascend NPU UT for torch.autograd.profiler.parse_nvprof_trace #42230

yuhongming-2026创建于 7月20日从yuhongming-2026/pytorch_3228： part_271 合入到 Ascend/pytorch： v2.7.1

test: add Ascend NPU UT for torch.autograd.profiler.parse_nvprof_trace #42230

已合并

yuhongming-2026创建于 7月20日

从yuhongming-2026/pytorch_3228 part_271合入到Ascend/pytorch v2.7.1

共 1 个文件变更+123-0

Atest/profiler/test_parse_nvprof_trace.py+123-0

@@ -0,0 +1,123 @@
	1	+# Copyright (c) 2026 Huawei Technologies Co., Ltd
	2	+# All rights reserved.
	3	+#
	4	+# Licensed under the BSD 3-Clause License (the "License");
	5	+# you may not use this file except in compliance with the License.
	6	+# You may obtain a copy of the License at
	7	+#
	8	+# https://opensource.org/licenses/BSD-3-Clause
	9	+#
	10	+# Unless required by applicable law or agreed to in writing, software
	11	+# distributed under the License is distributed on an "AS IS" BASIS,
	12	+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	13	+# See the License for the specific language governing permissions and
	14	+# limitations under the License.
	15	+# Owner(s): ["oncall: profiler"]
	16	+"""Ascend NPU adaptation checks for torch.autograd.profiler.parse_nvprof_trace.
	17	+
	18	+``torch.autograd.profiler.parse_nvprof_trace`` / ``load_nvprof`` parse NVIDIA
	19	+nvprof CUPTI SQLite. Ascend does not emit that format; NPU profiling uses
	20	+``torch.autograd.profiler.profile(use_device="npu")``.
	21	+
	22	+Community pytorch has limited CUPTI-oriented coverage and does not provide an
	23	+Ascend/NPU adaptation suite for this API. This file therefore:
	24	+ 1. checks the APIs remain available under the torch_npu stack;
	25	+ 2. verifies NPU profiler semantics with real NPU ops;
	26	+ 3. exercises ``EnforceUnique`` (used inside ``parse_nvprof_trace``);
	27	+ 4. confirms ``parse_nvprof_trace`` / ``load_nvprof`` still reject missing
	28	+ nvprof DB paths after NPU profiling (negative path; no CUPTI fixture).
	29	+"""
	30	+
	31	+import os
	32	+import tempfile
	33	+
	34	+import torch
	35	+from torch.autograd.profiler import EnforceUnique, load_nvprof, parse_nvprof_trace
	36	+from torch.testing._internal.common_utils import TestCase, run_tests
	37	+
	38	+
	群群群青世界7月30日四个测试 PR 没有验证 NPU profiler 语义，构造的输入仍是 CUDA nvprof/CUPTI 格式。 yuhongming-2026 29 天前评论： 29 天前评论：
	39	+class TestParseNvprofTraceAscendNPU(TestCase):
	40	+ """NPU checks for parse_nvprof_trace availability and NPU profiler semantics."""
	41	+
	42	+ def setUp(self):
	43	+ if not hasattr(torch, "npu") or not torch.npu.is_available():
	44	+ self.skipTest("Requires Ascend NPU (CANN). Skip on CPU/CUDA-only hosts.")
	45	+ torch.npu.set_device(0)
	46	+ self.device = torch.device("npu:0")
	47	+
	48	+ def _mm_on_npu(self):
	49	+ a = torch.randn(8, 8, device=self.device)
	50	+ b = torch.randn(8, 8, device=self.device)
	51	+ return torch.mm(a, b)
	52	+
	53	+ def _event_names(self, prof):
	54	+ return [e.name for e in prof.function_events]
	55	+
	56	+ def test_api_available_on_npu_stack(self):
	57	+ self.assertTrue(hasattr(torch.autograd.profiler, "parse_nvprof_trace"))
	58	+ self.assertTrue(callable(torch.autograd.profiler.parse_nvprof_trace))
	59	+ self.assertTrue(hasattr(torch.autograd.profiler, "load_nvprof"))
	60	+ self.assertTrue(callable(torch.autograd.profiler.load_nvprof))
	61	+ self.assertIs(parse_nvprof_trace, torch.autograd.profiler.parse_nvprof_trace)
	62	+ self.assertIs(load_nvprof, torch.autograd.profiler.load_nvprof)
	63	+
	64	+ def test_npu_profiler_records_mm(self):
	65	+ """Real NPU profiler path: use_device='npu' must capture mm events."""
	66	+ with torch.autograd.profiler.profile(use_device="npu") as prof:
	67	+ out = self._mm_on_npu()
	68	+ torch.npu.synchronize()
	69	+ self.assertEqual(out.device.type, "npu")
	70	+ self.assertTrue(
	71	+ any("mm" in name for name in self._event_names(prof)),
	72	+ f"expected mm in NPU profiler events, got {self._event_names(prof)}",
	73	+ )
	74	+
	75	+ def test_npu_profiler_records_add_and_mm(self):
	76	+ """NPU profiler should record multiple ops executed on device."""
	77	+ with torch.autograd.profiler.profile(use_device="npu") as prof:
	78	+ a = torch.randn(16, 16, device=self.device)
	79	+ b = torch.randn(16, 16, device=self.device)
	80	+ c = torch.mm(a, b)
	81	+ d = c + a
	82	+ torch.npu.synchronize()
	83	+ self.assertEqual(d.device.type, "npu")
	84	+ names = self._event_names(prof)
	85	+ self.assertTrue(any("mm" in n for n in names), f"missing mm in {names}")
	86	+ self.assertTrue(
	87	+ any(("add" in n) or ("+" in n) for n in names),
	88	+ f"missing add in {names}",
	89	+ )
	90	+
	91	+ def test_npu_profiler_event_timing(self):
	92	+ """Captured NPU profiler events should expose non-negative CPU time."""
	93	+ with torch.autograd.profiler.profile(use_device="npu") as prof:
	94	+ _ = self._mm_on_npu()
	95	+ torch.npu.synchronize()
	96	+ mm_events = [e for e in prof.function_events if "mm" in e.name]
	97	+ self.assertGreater(len(mm_events), 0)
	98	+ for evt in mm_events:
	99	+ self.assertGreaterEqual(evt.cpu_time, 0.0)
	100	+
	101	+ def test_parse_apis_reject_missing_nvprof_db_after_npu_profile(self):
	102	+ """After NPU profiling, parse APIs remain callable and reject missing DB."""
	103	+ with torch.autograd.profiler.profile(use_device="npu") as prof:
	104	+ _ = self._mm_on_npu()
	105	+ torch.npu.synchronize()
	106	+ self.assertGreater(len(list(prof.function_events)), 0)
	107	+
	108	+ missing = os.path.join(tempfile.gettempdir(), "missing_nvprof_ascend.sqlite")
	109	+ with self.assertRaises(Exception):
	110	+ parse_nvprof_trace(missing)
	111	+ with self.assertRaises(Exception):
	112	+ load_nvprof(missing)
	113	+
	114	+ def test_enforce_unique_on_npu(self):
	115	+ _ = torch.randn(2, 2, device=self.device)
	116	+ unique = EnforceUnique()
	117	+ unique.see("a", 1)
	118	+ with self.assertRaises(RuntimeError):
	119	+ unique.see("a", 1)
	120	+
	121	+
	122	+if __name__ == "__main__":
	123	+ run_tests()