bugfix: remove the decomposition of expm1 from torchbench and add it to mlir.-pytorch-AtomGit

已合并

bugfix: remove the decomposition of expm1 from torchbench and add it to mlir. #36487

yvjc创建于 5月22日从已删除 :decomp7 合入到 Ascend/pytorch： v2.7.1

bugfix: remove the decomposition of expm1 from torchbench and add it to mlir. #36487

已合并

yvjc创建于 5月22日

从已删除 :decomp7合入到Ascend/pytorch v2.7.1

共 4 个文件变更+48-24

Mbenchmarks/torchbench/npu_support.py+0-6

	@@ -345,12 +345,6 @@ def _patch_model_10():
345		345
346	DeepRecommenderTrainBenchmark.__init__ = new_init	346	DeepRecommenderTrainBenchmark.__init__ = new_init
347		347
348	- def expm1(x):
349	- tensor = torch.exp(x) - torch.ones_like(x)
350	- return tensor
351	-
352	- inductor_decomp.register_decomposition(torch.ops.aten.expm1)(expm1)
353	-
354		348
355	@register_patch("resnet50", "resnet152", "resnext50_32x4d", "densenet121")	349	@register_patch("resnet50", "resnet152", "resnext50_32x4d", "densenet121")
356	def _patch_model_18():	350	def _patch_model_18():

Rtest/_inductor/test_inductor_fallback_list.py→test/_inductor/test_inductor_fallback_list_mlir.py+11-18

	@@ -1,27 +1,20 @@
1	-import torch
2	-from torch.testing._internal.common_utils import run_tests, parametrize, instantiate_parametrized_tests
3	-from testutils import TestUtils
4	-from torch._inductor.utils import run_and_get_code
5	-import torch_npu
6	import os	1	import os
		2	+
		3	+import torch
		4	+from torch.testing._internal.common_utils import run_tests
		5	+from torch._inductor.utils import run_and_get_code
		6	+
		7	+from testutils import TestUtils
		8	+
		9	+
7	os.environ["NPU_INDUCTOR_FALLBACK_LIST"] = "aten.div,aten.add.Tensor"	10	os.environ["NPU_INDUCTOR_FALLBACK_LIST"] = "aten.div,aten.add.Tensor"
8		11
		12	+
9	class TestFallback(TestUtils):	13	class TestFallback(TestUtils):
10		14
11	def add_op(self, x, y):	15	def add_op(self, x, y):
12	- return x / y	16	+ return x / y
13	-
14	- def test_add_fallback_detection(self):
15	-
16	- compiled_add = torch.compile(self.add_op, backend="inductor")
17	-
18	- x = torch.randn(4, 4, dtype=torch.float32).to("npu")
19	- y = torch.randn(4, 4, dtype=torch.float32).to("npu")
20		17
21	- _ , codes = run_and_get_code(compiled_add, x, y)
22	-
23	- self.assertTrue('unk_fused_div' not in codes[0])
24	-
25	def test_add_fallback_detection_mlir(self):	18	def test_add_fallback_detection_mlir(self):
26		19
27	compiled_add = torch.compile(self.add_op, backend="inductor", options={"npu_backend": "mlir"})	20	compiled_add = torch.compile(self.add_op, backend="inductor", options={"npu_backend": "mlir"})
	@@ -35,4 +28,4 @@ class TestFallback(TestUtils):
35		28
36		29
37	if __name__ == "__main__":	30	if __name__ == "__main__":
38	- run_tests()	31	+ run_tests()

Atest/_inductor/test_inductor_fallback_list_triton.py+31-0

@@ -0,0 +1,31 @@
	1	+import os
	2	+
	3	+import torch
	4	+from torch.testing._internal.common_utils import run_tests
	5	+from torch._inductor.utils import run_and_get_code
	6	+
	7	+from testutils import TestUtils
	8	+
	9	+
	10	+os.environ["NPU_INDUCTOR_FALLBACK_LIST"] = "aten.div,aten.add.Tensor"
	11	+
	12	+
	13	+class TestFallback(TestUtils):
	14	+
	15	+ def add_op(self, x, y):
	16	+ return x / y
	17	+
	18	+ def test_add_fallback_detection(self):
	19	+
	20	+ compiled_add = torch.compile(self.add_op, backend="inductor")
	21	+
	22	+ x = torch.randn(4, 4, dtype=torch.float32).to("npu")
	23	+ y = torch.randn(4, 4, dtype=torch.float32).to("npu")
	24	+
	25	+ _ , codes = run_and_get_code(compiled_add, x, y)
	26	+
	27	+ self.assertTrue('unk_fused_div' not in codes[0])
	28	+
	29	+
	30	+if __name__ == "__main__":
	31	+ run_tests()

Mtorch_npu/_inductor/ascend_npu_ir/ascend_npu_ir/npu/npu_decomp.py+6-0

	@@ -192,6 +192,12 @@ def gelu_backward(grad: Tensor, self: Tensor, approximate: str = "none"):
192		192
193	return grad * (left_derivative + right_derivative)	193	return grad * (left_derivative + right_derivative)
194		194
		195	+
		196	+def expm1(x):
		197	+ tensor = torch.exp(x) - torch.ones_like(x)
		198	+ return tensor
		199	+
		200	+inductor_decomp.register_decomposition(torch.ops.aten.expm1)(expm1)
195	inductor_decomp.register_decomposition(torch.ops.aten.convolution_backward)(npu_convolution_backward)	201	inductor_decomp.register_decomposition(torch.ops.aten.convolution_backward)(npu_convolution_backward)
196	inductor_decomp.register_decomposition(torch.ops.aten._softmax_backward_data.default)(npu__softmax_backward_data)	202	inductor_decomp.register_decomposition(torch.ops.aten._softmax_backward_data.default)(npu__softmax_backward_data)
197	inductor_decomp.register_decomposition(torch.ops.aten.gelu.default)(gelu)	203	inductor_decomp.register_decomposition(torch.ops.aten.gelu.default)(gelu)