Allow transposed convolution weights with a non-default dim order

[SuryanshSS1011]

Summary

Fixes #20804.

A transposed convolution whose weight has out_channels == 1 fails in the portable kernel with Check failed (tensor_is_default_or_channels_last_dim_order(weight)) (status 18). check_convolution_args runs tensor_is_default_or_channels_last_dim_order(weight) unconditionally, but a transposed conv weight is laid out as (in_channels, out_channels / groups, kH, kW), so when out_channels == 1 the weight comes through with a dim order such as [1, 0, 2, 3], which is neither contiguous ([0,1,2,3]) nor channels-last. For out_channels > 1 the dim order is [0,1,2,3] and the check passes, which matches the report.

The transposed path in op_convolution.cpp indexes the weight through strides derived from its dim order (dim_order_to_stride_nocheck then calculate_linear_index), so it reads the correct values regardless of whether the weight is in [0,1,2,3] or [1,0,2,3]. I checked this with the exact stride/index logic: every weight lookup matches between the two layouts, so the kernel already handles this weight and only the up-front check needs to change.

This skips the weight dim-order check when transposed is true, since the kernel does not rely on that layout, rather than reordering the weight upstream. Happy to switch to normalizing the weight instead if that's preferred.

Test plan

Added TransposedWeightNonDefaultDimOrder in kernels/test/op_convolution_test.cpp: a transposed conv with a (1, 1, 2, 2) weight built in dim order [1, 0, 2, 3], checked against a reference output computed with torch.nn.ConvTranspose2d. The test fails before this change (the weight is rejected) and passes after it.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21035

• 📄 Preview Python docs built from this PR

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

⚠️ 13 Awaiting Approval

As of commit 1433bf1 with merge base 55a71e6 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

• Build documentation (gh)
• Build Presets (gh)
• Cadence Build & Test (gh)
• Lint (gh)
• periodic (gh)
• pull (gh)
• Test ARM Backend (gh)
• Test CoreML Backend (gh)
• Test QNN Backend (gh)
• Test QNN Windows MSVC build (gh)
• Test Vulkan Backend (gh)
• Test WebGPU Backend (gh)
• Test XNNPACK Backend (gh)

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

[Copilot]

Pull request overview

This PR fixes a portability issue in ExecuTorch’s portable CPU convolution kernel argument validation for transposed convolutions, where valid conv_transpose2d weights can arrive with a non-default dim order (not contiguous / not channels-last) when out_channels == 1.

Changes:

• Skip the “default or channels-last dim order” validation for weight when transposed == true, since the transposed convolution implementation indexes weights via strides.
• Add a regression test intended to cover the non-default dim-order transposed-weight scenario (issue #20804).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
kernels/portable/cpu/util/kernel_ops_util.cpp	Adjusts convolution argument validation to allow non-default dim order weights for transposed convolutions.
kernels/test/op_convolution_test.cpp	Adds a regression test for transposed convolution weights with a non-default dim order.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[SuryanshSS1011]

@pytorchbot label "release notes: ops & kernels"

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.