[FakeTensor] Supplement the relevant logic for converting conv1d to conv2d in meta_conv #160408

thenumberouscode · 2025-08-12T10:21:09Z

Fixes #159462

summary

the issue is caused by the wrong stride of conv1d's result generated by meta_conv:

Lines 2453 to 2471 in 4d5b3f2

    
           shape_out = calc_conv_nd_return_shape( 
        
               input_tensor, 
        
               weight, 
        
               stride, 
        
               padding, 
        
               dilation, 
        
               is_transposed, 
        
               groups, 
        
               output_padding if is_transposed else None, 
        
           ) 
        
           input_channels_dim = 1 
        
           output_channels_dim = 1 
        
           if input_tensor.size(input_channels_dim) == 0: 
        
               shape_out[output_channels_dim] = 0 
        
           out = input_tensor.new_empty(shape_out) 
        
           out = out.to(memory_format=pick_memory_format())  # type: ignore[call-overload] 
        
           return out

and the wrong stride will be used to codegen size assert in inductor:

pytorch/torch/_inductor/ir.py

Lines 6152 to 6163 in 4d5b3f2

    
           def codegen_size_asserts(self, wrapper: PythonWrapperCodegen) -> None: 
        
               if config.size_asserts and not V.graph.cpp_wrapper: 
        
                   # comparing strides for 0 size tensor is tricky. Ignore them for now. 
        
                   if sympy_product(self.get_size()) == 0: 
        
                       return 
        
                   size = V.graph.wrapper_code.codegen_shape_tuple(self.get_size()) 
        
                   stride = V.graph.wrapper_code.codegen_shape_tuple(self.get_stride()) 
        
                   op_name = self.get_op_name() 
        
                   wrapper.writeline( 
        
                       f"assert_size_stride({self.get_name()}, {size}, {stride}, {op_name!r})" 
        
                   )

reason

So why the computed stride is wrong in the meta_conv function? because the corresponding backend will convert conv1d to conv2d and change the input tensor' size and memory_format(channel last). but the meta_conv do not do this transformation, so a mismatch happend.

pytorch/aten/src/ATen/native/Convolution.cpp

Lines 1502 to 1510 in 4d5b3f2

    
           // Expand 1d -> 2d. 
        
           // This is only done for backends that don't natively support 1d spatial input. 
        
           if (k == 3 && !input.is_mkldnn() && !input.is_xpu()) { 
        
             // avoid accidentally going through NHWC for permuted 3d input. 
        
             input = input.contiguous(); 
        
             params.view1d_as_2d(); 
        
             input = view4d(input); 
        
             weight = view4d(weight); 
        
           }

just add corresponding logic in meta_conv.

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @eellison

…o convert2d in conv of meta kernel

pytorch-bot · 2025-08-12T10:21:13Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/160408

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 28a839d with merge base fea7e9d ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

thenumberouscode · 2025-08-12T10:22:54Z

@eellison Please review my PR and leave any comments if you are convenient.

eellison

thanks for pr - I would have expected us to need to update the fake impl. Which has access to the device information.

eellison · 2025-08-12T15:31:04Z

torch/_meta_registrations.py

@@ -2450,6 +2450,19 @@ def pick_memory_format():
        elif input_tensor.is_contiguous(memory_format=torch.preserve_format):
            return torch.preserve_format

+    # Expand 1d -> 2d.


Are you sure we don't need to update

pytorch/torch/_subclasses/fake_impls.py

Lines 1024 to 1026 in b7db866

elif k == 3 and not kwargs["input"].is_mkldnn and not kwargs["input"].is_xpu:

mem_fmt = None

else:

?

Hmm, not entirely sure yet. I tested the bug fix locally and it worked fine, but I'll double-check and confirm after further investigation.

[FakeTensor] Supplement the relevant logic for converting convert1d t…

28a839d

…o convert2d in conv of meta kernel

pytorch-bot bot added the module: inductor label Aug 12, 2025

thenumberouscode changed the title ~~[FakeTensor] Supplement the relevant logic for converting convert1d t…~~ [FakeTensor] Supplement the relevant logic for converting conv1d to conv2d in meta_conv Aug 12, 2025

pytorchbot added the open source label Aug 12, 2025

eellison reviewed Aug 12, 2025

View reviewed changes

jerryzh168 added the triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module label Aug 13, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[FakeTensor] Supplement the relevant logic for converting conv1d to conv2d in meta_conv #160408

[FakeTensor] Supplement the relevant logic for converting conv1d to conv2d in meta_conv #160408

thenumberouscode commented Aug 12, 2025 •

edited by pytorch-bot bot

Loading

Uh oh!

pytorch-bot bot commented Aug 12, 2025 •

edited

Loading

Uh oh!

thenumberouscode commented Aug 12, 2025

Uh oh!

eellison left a comment

Uh oh!

eellison Aug 12, 2025

Uh oh!

thenumberouscode Aug 13, 2025

Uh oh!

Uh oh!

	shape_out = calc_conv_nd_return_shape(
	input_tensor,
	weight,
	stride,
	padding,
	dilation,
	is_transposed,
	groups,
	output_padding if is_transposed else None,
	)

	input_channels_dim = 1
	output_channels_dim = 1
	if input_tensor.size(input_channels_dim) == 0:
	shape_out[output_channels_dim] = 0

	out = input_tensor.new_empty(shape_out)
	out = out.to(memory_format=pick_memory_format()) # type: ignore[call-overload]
	return out

	def codegen_size_asserts(self, wrapper: PythonWrapperCodegen) -> None:
	if config.size_asserts and not V.graph.cpp_wrapper:
	# comparing strides for 0 size tensor is tricky. Ignore them for now.
	if sympy_product(self.get_size()) == 0:
	return
	size = V.graph.wrapper_code.codegen_shape_tuple(self.get_size())
	stride = V.graph.wrapper_code.codegen_shape_tuple(self.get_stride())
	op_name = self.get_op_name()
	wrapper.writeline(
	f"assert_size_stride({self.get_name()}, {size}, {stride}, {op_name!r})"
	)

	// Expand 1d -> 2d.
	// This is only done for backends that don't natively support 1d spatial input.
	if (k == 3 && !input.is_mkldnn() && !input.is_xpu()) {
	// avoid accidentally going through NHWC for permuted 3d input.
	input = input.contiguous();
	params.view1d_as_2d();
	input = view4d(input);
	weight = view4d(weight);
	}

	elif k == 3 and not kwargs["input"].is_mkldnn and not kwargs["input"].is_xpu:
	mem_fmt = None
	else:

[FakeTensor] Supplement the relevant logic for converting conv1d to conv2d in meta_conv #160408

Are you sure you want to change the base?

[FakeTensor] Supplement the relevant logic for converting conv1d to conv2d in meta_conv #160408

Conversation

thenumberouscode commented Aug 12, 2025 • edited by pytorch-bot bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Fixes #159462

summary

reason

Uh oh!

pytorch-bot bot commented Aug 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/160408

✅ No Failures

Uh oh!

thenumberouscode commented Aug 12, 2025

Uh oh!

eellison left a comment

Choose a reason for hiding this comment

Uh oh!

eellison Aug 12, 2025

Choose a reason for hiding this comment

Uh oh!

thenumberouscode Aug 13, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

thenumberouscode commented Aug 12, 2025 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Aug 12, 2025 •

edited

Loading