[Inductor] fix AOTI update_constant_buffer bug-pytorch-AtomGit

已合并

[Inductor] fix AOTI update_constant_buffer bug #41499

zhucehw创建于 7月13日从zhucehw/pytorch： v2.10.0-dev 合入到 Ascend/pytorch： v2.10.0

[Inductor] fix AOTI update_constant_buffer bug #41499

已合并

zhucehw创建于 7月13日

从zhucehw/pytorch v2.10.0-dev合入到Ascend/pytorch v2.10.0

共 2 个文件变更+5-5

Mtest/_inductor/test_aoti_user_defined_op.py+4-4

	@@ -80,7 +80,7 @@ def activation_min_max(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
80	class Model(torch.nn.Module):	80	class Model(torch.nn.Module):
81	def __init__(self, dim=32):	81	def __init__(self, dim=32):
82	super().__init__()	82	super().__init__()
83	- self.fc1 = torch.nn.Linear(dim, dim, dtype=torch.float16)	83	+ self.fc1 = torch.nn.Linear(dim, dim, dtype=torch.float32)
84	self.sigmoid = torch.nn.Sigmoid()	84	self.sigmoid = torch.nn.Sigmoid()
85		85
86	def forward(self, x, y, z):	86	def forward(self, x, y, z):
	@@ -117,11 +117,11 @@ class Model(torch.nn.Module):
117	class TestAotiUserDefinedOp(TestUtils):	117	class TestAotiUserDefinedOp(TestUtils):
118	def generate_input_tensor(self, batch_size=8, dim=32, device="npu"):	118	def generate_input_tensor(self, batch_size=8, dim=32, device="npu"):
119	x_input = torch.arange(0, batch_size * dim, 1, device=device).reshape([batch_size, dim])	119	x_input = torch.arange(0, batch_size * dim, 1, device=device).reshape([batch_size, dim])
120	- x_input = 1.0 / x_input.to(torch.float16)	120	+ x_input = 1.0 / x_input.to(torch.float32)
121	y_input = torch.arange(batch_size * dim, 0, -1, device=device).reshape([batch_size, dim])	121	y_input = torch.arange(batch_size * dim, 0, -1, device=device).reshape([batch_size, dim])
122	- y_input = 1.0 / y_input.to(torch.float16)	122	+ y_input = 1.0 / y_input.to(torch.float32)
123	# z_input is a cpu input tensor	123	# z_input is a cpu input tensor
124	- z_input = torch.ones(batch_size * dim, device="cpu", dtype=torch.float16).reshape([batch_size, dim])	124	+ z_input = torch.ones(batch_size * dim, device="cpu", dtype=torch.float32).reshape([batch_size, dim])
125	return x_input, y_input, z_input	125	return x_input, y_input, z_input
126		126
127		127

Mtorch_npu/csrc/inductor/aoti_runtime/model_container.h+1-1

	@@ -442,7 +442,7 @@ public:
442	aoti_torch_get_storage_size(tensor, &constant_size);	442	aoti_torch_get_storage_size(tensor, &constant_size);
443	#if defined(USE_NPU)	443	#if defined(USE_NPU)
444	AOTI_RUNTIME_DEVICE_CHECK(aclrtMemcpy(internal_constants_ptr, constant_size, user_constant_ptr,	444	AOTI_RUNTIME_DEVICE_CHECK(aclrtMemcpy(internal_constants_ptr, constant_size, user_constant_ptr,
445	- constant_size, ACL_MEMCPY_HOST_TO_DEVICE));	445	+ constant_size, ACL_MEMCPY_DEFAULT));
446	#else	446	#else
447	memcpy(internal_constants_ptr, user_constant_ptr, constant_size);	447	memcpy(internal_constants_ptr, user_constant_ptr, constant_size);
448	#endif	448	#endif