修改图优化用例及说明文档-pytorch-AtomGit

已合并

修改图优化用例及说明文档 #39031

修改图优化用例及说明文档 #39031

已合并

zzll创建于 6月23日

共 5 个文件变更+18-17

Mtest/_inductor/test_fold_redundant_ops_pass.py+4-4

	@@ -59,10 +59,10 @@ class TestFoldMultiShapeUnchangePass(TestUtils):
59	std_result = self.op_calc(arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1)	59	std_result = self.op_calc(arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1)
60	with torch.no_grad():	60	with torch.no_grad():
61	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")	61	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")
62	- inductor_result = compiled_op_calc(arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1)	62	+ inductor_result = compiled_op_calc(arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1)
63	- self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)	63	+ self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)
64	-	64	+
65	-	65	+
66	def test_ut_cases(self):	66	def test_ut_cases(self):
67	arg0_1 = torch.randn(289094, 64, dtype=torch.float32)	67	arg0_1 = torch.randn(289094, 64, dtype=torch.float32)
68	arg1_1 = torch.randint(0, 289094, (128,), dtype=torch.int64)	68	arg1_1 = torch.randint(0, 289094, (128,), dtype=torch.int64)

Mtest/_inductor/test_fold_slice_pass.py+4-4

	@@ -41,10 +41,10 @@ class TestFoldSlicePass(TestUtils):
41	std_result = self.op_calc(base, view, t1, t2, t3)	41	std_result = self.op_calc(base, view, t1, t2, t3)
42	with torch.no_grad():	42	with torch.no_grad():
43	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")	43	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")
44	- inductor_result = compiled_op_calc(base, view, t1, t2, t3)	44	+ inductor_result = compiled_op_calc(base, view, t1, t2, t3)
45	- self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)	45	+ self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)
46	-	46	+
47	-	47	+
48	def test_ut_cases(self):	48	def test_ut_cases(self):
49	base = torch.randn(8, 16, 32)	49	base = torch.randn(8, 16, 32)
50	view = torch.ones(8, 16, 32)	50	view = torch.ones(8, 16, 32)

Mtest/_inductor/test_fold_squeeze_pass.py+4-4

	@@ -34,10 +34,10 @@ class TestFoldSqueezePass(TestUtils):
34	std_result = self.op_calc(t1, t2)	34	std_result = self.op_calc(t1, t2)
35	with torch.no_grad():	35	with torch.no_grad():
36	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")	36	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")
37	- inductor_result = compiled_op_calc(t1, t2)	37	+ inductor_result = compiled_op_calc(t1, t2)
38	- self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)	38	+ self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)
39	-	39	+
40	-	40	+
41	def test_ut_cases(self):	41	def test_ut_cases(self):
42	t1 = torch.randn(2, 4)	42	t1 = torch.randn(2, 4)
43	t2 = torch.randn(2, 1, 1, 4)	43	t2 = torch.randn(2, 1, 1, 4)

Mtest/_inductor/test_fold_view_pass.py+4-4

	@@ -34,10 +34,10 @@ class TestFoldViewPass(TestUtils):
34	std_result = self.op_calc(t1, t2)	34	std_result = self.op_calc(t1, t2)
35	with torch.no_grad():	35	with torch.no_grad():
36	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")	36	compiled_op_calc = torch.compile(self.op_calc, backend="inductor")
37	- inductor_result = compiled_op_calc(t1, t2)	37	+ inductor_result = compiled_op_calc(t1, t2)
38	- self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)	38	+ self.assertEqual(std_result, inductor_result, atol=1e-3, rtol=1e-3)
39	-	39	+
40	-	40	+
41	def test_ut_cases(self):	41	def test_ut_cases(self):
42	t1 = torch.randn(1, 3, 1, 5)	42	t1 = torch.randn(1, 3, 1, 5)
43	t2 = torch.randn(128, 64)	43	t2 = torch.randn(128, 64)

Mtorch_npu/_inductor/docs/feature/graph_optimization/overview.md+2-1

	@@ -12,6 +12,7 @@
12	# pre	12	# pre
13	cat_slice_cat_fold_pass	13	cat_slice_cat_fold_pass
14	pad_slice_fold	14	pad_slice_fold
		15	+fusion_attention_v3_pass
15	```	16	```
16		17
17	```text	18	```text
	@@ -80,7 +81,7 @@ with torch.no_grad():
80	```python	81	```python
81	# 设置 inductor 日志级别以便观察 pass 生效情况	82	# 设置 inductor 日志级别以便观察 pass 生效情况
82	import os	83	import os
83	-os.environ["INDUCTOR_ASCEND_LOG_LEVEL"] = "INFO"	84	+os.environ["INDUCTOR_ASCEND_LOG_LEVEL"] = "DEBUG"
84	import torch	85	import torch
85	```	86	```
86		87