fix: 修复 inductor 编译子进程重复初始化 NPU 的问题-pytorch-AtomGit

已合并

fix: 修复 inductor 编译子进程重复初始化 NPU 的问题 #44922

fix: 修复 inductor 编译子进程重复初始化 NPU 的问题 #44922

已合并

dingdairong创建于 13 天前

共 4 个文件变更+10-55

	@@ -13,23 +13,6 @@ from torch.testing._internal.common_utils import (
13	@unittest.skipIf(not torch.npu.is_available(), "requires an NPU device")	13	@unittest.skipIf(not torch.npu.is_available(), "requires an NPU device")
14	class TestAscendcBasic(TestCase):	14	class TestAscendcBasic(TestCase):
15		15
16	- @classmethod
17	- def setUpClass(cls):
18	- super().setUpClass()
19	- cls._ascendc_ok = False
20	- try:
21	- x = torch.randn(4, 4, device="npu")
22	- torch.compile(lambda t: t + 1, backend="inductor",
23	- options={"npu_backend": "ascendc"})(x)
24	- cls._ascendc_ok = True
25	- except Exception:
26	- pass
27	-
28	- def setUp(self):
29	- super().setUp()
30	- if not self._ascendc_ok:
31	- self.skipTest("ascendc backend not available")
32	-
33	@parametrize("dtype", [torch.float32, torch.float16])	16	@parametrize("dtype", [torch.float32, torch.float16])
34	def test_mul_sub(self, dtype):	17	def test_mul_sub(self, dtype):
35	"""Pointwise pattern: compiled output matches eager."""	18	"""Pointwise pattern: compiled output matches eager."""

	@@ -62,20 +62,6 @@ def _temporary_env(name, value):
62	os.environ[name] = old_value	62	os.environ[name] = old_value
63		63
64		64
65	-def _is_ascendc_backend_available():
66	- try:
67	- x = torch.randn(4, 4, device="npu")
68	- torch.compile(
69	- lambda t: t + 1,
70	- backend="inductor",
71	- options={"npu_backend": "ascendc"},
72	- )(x)
73	- torch.npu.synchronize()
74	- except Exception:
75	- return False
76	- return True
77	-
78	-
79	def _run_dynamic_actual_seq_case():	65	def _run_dynamic_actual_seq_case():
80	import torch_npu._inductor.ascendc.config as ascendc_config	66	import torch_npu._inductor.ascendc.config as ascendc_config
81		67
	@@ -166,15 +152,8 @@ def _assert_dynamic_actual_seq_key_reuse(test_case, logs):
166	@unittest.skipIf(not torch.npu.is_available(), "requires an NPU device")	152	@unittest.skipIf(not torch.npu.is_available(), "requires an NPU device")
167	class TestAscendcIFAv2DynamicActualSeq(TestCase):	153	class TestAscendcIFAv2DynamicActualSeq(TestCase):
168		154
169	- @classmethod
170	- def setUpClass(cls):
171	- super().setUpClass()
172	- cls._ascendc_ok = _is_ascendc_backend_available()
173	-
174	def setUp(self):	155	def setUp(self):
175	super().setUp()	156	super().setUp()
176	- if not self._ascendc_ok:
177	- self.skipTest("ascendc backend not available")
178	torch._dynamo.reset()	157	torch._dynamo.reset()
179		158
180	def tearDown(self):	159	def tearDown(self):

	@@ -8,23 +8,6 @@ from torch.testing._internal.common_utils import run_tests, TestCase
8	@unittest.skipIf(not torch.npu.is_available(), "requires an NPU device")	8	@unittest.skipIf(not torch.npu.is_available(), "requires an NPU device")
9	class TestAscendcReduceOverhead(TestCase):	9	class TestAscendcReduceOverhead(TestCase):
10		10
11	- @classmethod
12	- def setUpClass(cls):
13	- super().setUpClass()
14	- cls._ascendc_ok = False
15	- try:
16	- x = torch.randn(4, 4, device="npu")
17	- torch.compile(lambda t: t + 1, backend="inductor",
18	- options={"npu_backend": "ascendc"})(x)
19	- cls._ascendc_ok = True
20	- except Exception:
21	- pass
22	-
23	- def setUp(self):
24	- super().setUp()
25	- if not self._ascendc_ok:
26	- self.skipTest("ascendc backend not available")
27	-
28	def test_capture_replay(self):	11	def test_capture_replay(self):
29	"""Verify reduce-overhead triggers graph capture and results are correct."""	12	"""Verify reduce-overhead triggers graph capture and results are correct."""
30	def fn(x, y):	13	def fn(x, y):

	@@ -1,4 +1,14 @@
1	import os	1	import os
		2	+
		3	+ORG_AUTOLOAD = os.getenv("TORCH_DEVICE_BACKEND_AUTOLOAD", "1")
		4	+os.environ["TORCH_DEVICE_BACKEND_AUTOLOAD"] = "0"
		5	+from torch._inductor.async_compile import AsyncCompile
		6	+
		7	+
		8	+if os.environ.get("TORCH_WARM_POOL", "1") == "1":
		9	+ AsyncCompile.warm_pool()
		10	+os.environ["TORCH_DEVICE_BACKEND_AUTOLOAD"] = ORG_AUTOLOAD
		11	+
2	from torch_npu.utils._dynamo import (	12	from torch_npu.utils._dynamo import (
3	_dynamo_register_interface_for_device,	13	_dynamo_register_interface_for_device,
4	patch_SkipFunctionVariable,	14	patch_SkipFunctionVariable,