[fix][v2.8.0]aclgraph testcase add 910C support-pytorch-AtomGit

已合并

[fix][v2.8.0]aclgraph testcase add 910C support #31719

翰萨姆创建于 3月12日从yangch0324/pytorch_acltest： v2.8.0_acltest 合入到 Ascend/pytorch： v2.8.0

[fix][v2.8.0]aclgraph testcase add 910C support #31719

已合并

翰萨姆创建于 3月12日

从yangch0324/pytorch_acltest v2.8.0_acltest合入到Ascend/pytorch v2.8.0

共 3 个文件变更+16-16

Mtest/npu/test_aclgraph_launch_host_func.py+1-1

	@@ -31,7 +31,7 @@ class MyModel(nn.Module):
31		31
32	class TestAclgraphLaunchHostFunc(TestCase):	32	class TestAclgraphLaunchHostFunc(TestCase):
33		33
34	- @SupportedDevices(['Ascend910B'])	34	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
35	def test_launch_host_func(self):	35	def test_launch_host_func(self):
36	torch_npu.npu.set_compile_mode(jit_compile=False)	36	torch_npu.npu.set_compile_mode(jit_compile=False)
37	torch_npu.npu.set_device(0)	37	torch_npu.npu.set_device(0)

Mtest/npu/test_aclgraph_support_blocking.py+3-3

	@@ -14,7 +14,7 @@ os.environ["ASCEND_LAUNCH_BLOCKING"] = "1"
14		14
15		15
16	class TestIFAAclgraphUpdateSupportBlocking(TestCase):	16	class TestIFAAclgraphUpdateSupportBlocking(TestCase):
17	- @SupportedDevices(['Ascend910B', 'Ascend910C'])	17	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
18	def test_ifa_update_support_blocking(self):	18	def test_ifa_update_support_blocking(self):
19	torch.npu.set_device(0)	19	torch.npu.set_device(0)
20	length = [29]	20	length = [29]
	@@ -63,7 +63,7 @@ class TestIFAAclgraphUpdateSupportBlocking(TestCase):
63	self.assertEqual(output.cpu(), res_src[0].cpu())	63	self.assertEqual(output.cpu(), res_src[0].cpu())
64	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	64	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
65		65
66	- @SupportedDevices(['Ascend910B', 'Ascend910C'])	66	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
67	def test_ifa_update_with_auto_dispatch_capture_support_blocking(self):	67	def test_ifa_update_with_auto_dispatch_capture_support_blocking(self):
68	torch.npu.set_device(0)	68	torch.npu.set_device(0)
69	length = [29]	69	length = [29]
	@@ -97,7 +97,7 @@ class TestIFAAclgraphUpdateSupportBlocking(TestCase):
97	self.assertEqual(output.cpu(), res_src[0].cpu())	97	self.assertEqual(output.cpu(), res_src[0].cpu())
98	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	98	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
99		99
100	- @SupportedDevices(['Ascend910B', 'Ascend910C'])	100	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
101	def test_ifa_update_with_non_out_and_auto_dispatch_capture_support_blocking(self):	101	def test_ifa_update_with_non_out_and_auto_dispatch_capture_support_blocking(self):
102	torch.npu.set_device(0)	102	torch.npu.set_device(0)
103	length = [29]	103	length = [29]

Mtest/npu/test_aclgraph_update.py+12-12

	@@ -15,7 +15,7 @@ from torch_npu.testing.testcase import TestCase, run_tests
15		15
16	class TestIFAAclgraphUpdate(TestCase):	16	class TestIFAAclgraphUpdate(TestCase):
17		17
18	- @SupportedDevices(['Ascend910B'])	18	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
19	def test_ifa_update(self):	19	def test_ifa_update(self):
20	torch.npu.set_device(0)	20	torch.npu.set_device(0)
21	length = [29]	21	length = [29]
	@@ -64,7 +64,7 @@ class TestIFAAclgraphUpdate(TestCase):
64	self.assertEqual(output.cpu(), res_src[0].cpu())	64	self.assertEqual(output.cpu(), res_src[0].cpu())
65	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	65	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
66		66
67	- @SupportedDevices(['Ascend910B'])	67	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
68	def test_update_stream_globally_unique(self):	68	def test_update_stream_globally_unique(self):
69	torch.npu.set_device(0)	69	torch.npu.set_device(0)
70		70
	@@ -72,7 +72,7 @@ class TestIFAAclgraphUpdate(TestCase):
72	g2 = torch.npu.NPUGraph()	72	g2 = torch.npu.NPUGraph()
73	self.assertEqual(g1.graph_dispatch_mode.update_stream, g2.graph_dispatch_mode.update_stream)	73	self.assertEqual(g1.graph_dispatch_mode.update_stream, g2.graph_dispatch_mode.update_stream)
74		74
75	- @SupportedDevices(['Ascend910B'])	75	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
76	def test_ifa_update_with_auto_dispatch_capture(self):	76	def test_ifa_update_with_auto_dispatch_capture(self):
77	torch.npu.set_device(0)	77	torch.npu.set_device(0)
78	length = [29]	78	length = [29]
	@@ -106,7 +106,7 @@ class TestIFAAclgraphUpdate(TestCase):
106	self.assertEqual(output.cpu(), res_src[0].cpu())	106	self.assertEqual(output.cpu(), res_src[0].cpu())
107	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	107	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
108		108
109	- @SupportedDevices(['Ascend910B'])	109	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
110	def test_ifa_update_with_non_out_and_auto_dispatch_capture(self):	110	def test_ifa_update_with_non_out_and_auto_dispatch_capture(self):
111	torch.npu.set_device(0)	111	torch.npu.set_device(0)
112	length = [29]	112	length = [29]
	@@ -136,7 +136,7 @@ class TestIFAAclgraphUpdate(TestCase):
136	self.assertEqual(output.cpu(), res_src[0].cpu())	136	self.assertEqual(output.cpu(), res_src[0].cpu())
137	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	137	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
138		138
139	- @SupportedDevices(['Ascend910B'])	139	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
140	def test_ifa_v2_update_with_auto_dispatch_capture(self):	140	def test_ifa_v2_update_with_auto_dispatch_capture(self):
141	torch.npu.set_device(0)	141	torch.npu.set_device(0)
142	length = [1]	142	length = [1]
	@@ -169,7 +169,7 @@ class TestIFAAclgraphUpdate(TestCase):
169	g.replay()	169	g.replay()
170	self.assertEqual(output.cpu(), res_src[0].cpu())	170	self.assertEqual(output.cpu(), res_src[0].cpu())
171		171
172	- @SupportedDevices(['Ascend910B'])	172	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
173	def test_ifa_v2_update_with_non_out_and_auto_dispatch_capture(self):	173	def test_ifa_v2_update_with_non_out_and_auto_dispatch_capture(self):
174	torch.npu.set_device(0)	174	torch.npu.set_device(0)
175	length = [1]	175	length = [1]
	@@ -198,7 +198,7 @@ class TestIFAAclgraphUpdate(TestCase):
198	g.replay()	198	g.replay()
199	self.assertEqual(output.cpu(), res_src[0].cpu())	199	self.assertEqual(output.cpu(), res_src[0].cpu())
200		200
201	- @SupportedDevices(['Ascend910B'])	201	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
202	@unittest.skip("this cann version is not supported")	202	@unittest.skip("this cann version is not supported")
203	def test_npu_fused_infer_attention_score_v2(self):	203	def test_npu_fused_infer_attention_score_v2(self):
204	torch.npu.set_device(0)	204	torch.npu.set_device(0)
	@@ -246,7 +246,7 @@ class TestIFAAclgraphUpdate(TestCase):
246	self.assertEqual(output.cpu(), res_src[0].cpu())	246	self.assertEqual(output.cpu(), res_src[0].cpu())
247	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	247	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
248		248
249	- @SupportedDevices(['Ascend910B'])	249	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
250	def test_fia_out_4in1_with_graph(self, device="npu"):	250	def test_fia_out_4in1_with_graph(self, device="npu"):
251	q = torch.randn(1, 8, 164, 128, dtype=torch.float16).npu()	251	q = torch.randn(1, 8, 164, 128, dtype=torch.float16).npu()
252	k = torch.randn(1, 8, 1024, 128, dtype=torch.float16).npu()	252	k = torch.randn(1, 8, 1024, 128, dtype=torch.float16).npu()
	@@ -301,7 +301,7 @@ class TestIFAAclgraphUpdate(TestCase):
301	self.assertTrue(torch.allclose(output, output2, 1e-4, 1e-4))	301	self.assertTrue(torch.allclose(output, output2, 1e-4, 1e-4))
302	self.assertTrue(torch.allclose(softmax_lse, softmax_lse2, 1e-4, 1e-4))	302	self.assertTrue(torch.allclose(softmax_lse, softmax_lse2, 1e-4, 1e-4))
303		303
304	- @SupportedDevices(['Ascend910B'])	304	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
305	@unittest.skip("this cann version is not supported")	305	@unittest.skip("this cann version is not supported")
306	def test_npugraph_debug_dump(self):	306	def test_npugraph_debug_dump(self):
307	N, D_in, H, D_out = 640, 4096, 2048, 1024	307	N, D_in, H, D_out = 640, 4096, 2048, 1024
	@@ -330,7 +330,7 @@ class TestIFAAclgraphUpdate(TestCase):
330	self.assertTrue(os.path.getsize(file_path) > 0, "npugraph debug dump assert error")	330	self.assertTrue(os.path.getsize(file_path) > 0, "npugraph debug dump assert error")
331	os.remove(file_path)	331	os.remove(file_path)
332		332
333	- @SupportedDevices(['Ascend910B'])	333	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
334	def test_ifa_update_no_reset(self):	334	def test_ifa_update_no_reset(self):
335	torch.npu.set_device(0)	335	torch.npu.set_device(0)
336	length = [29]	336	length = [29]
	@@ -378,7 +378,7 @@ class TestIFAAclgraphUpdate(TestCase):
378	self.assertEqual(output.cpu(), res_src[0].cpu())	378	self.assertEqual(output.cpu(), res_src[0].cpu())
379	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())	379	self.assertEqual(softmax_lse.cpu(), res_src[1].cpu())
380		380
381	- @SupportedDevices(['Ascend910B'])	381	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
382	@unittest.skip("this cann version is not supported")	382	@unittest.skip("this cann version is not supported")
383	def test_npu_fused_infer_attention_score_v2_no_reset(self):	383	def test_npu_fused_infer_attention_score_v2_no_reset(self):
384	torch.npu.set_device(0)	384	torch.npu.set_device(0)
	@@ -548,7 +548,7 @@ class TestPAAclgraphUpdate(TestCase):
548	)	548	)
549	return params.output	549	return params.output
550		550
551	- @SupportedDevices(['Ascend910B'])	551	+ @SupportedDevices(['Ascend910B', 'Ascend910_93'])
552	def test_paged_attention_aclgraph_update(self):	552	def test_paged_attention_aclgraph_update(self):
553	params, golden_output = self.preprocess()	553	params, golden_output = self.preprocess()
554	output = None	554	output = None