CUDA backend: fix output stride mismatch in delegate copy-back

[mergennachin]

The .pte serializes the output dim_order from PyTorch's SDPA composite
(which may return a non-contiguous transposed view, e.g., efficient
attention outputs [B,Lq,H,D] transposed to [B,H,Lq,D]). However, the
AOTI delegate always produces contiguous output in its own layout,
ignoring the .pte's expected dim_order. The runtime byte-copies the
GPU data to the CPU ETensor, but the ETensor interprets it with the
.pte's strides — causing silent data corruption when the layouts differ.

Fix: in copy_slimtensor_to_etensor, detect when the SlimTensor (GPU)
and ETensor (CPU) have different strides. When they match (common case),
use the fast byte-copy path. When they differ, copy GPU data to a temp
CPU buffer then rearrange element-by-element to match the ETensor's
expected layout.

Also enables the accuracy check in test_non_pow2_head_dim_with_bool_mask
and adds test_output_stride_rearrange.py exercising both fast and slow
copy paths with Triton ON and OFF.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17945

• 📄 Preview Python docs built from this PR

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

⏳ 12 Pending, 2 Unrelated Failures

As of commit bbc53d9 with merge base 4c96679 ():

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / unittest / macos / macos-job (gh) (trunk failure)
  AttributeError: '_OpNamespace' 'mkldnn' object has no attribute '_is_mkldnn_acl_supported'
• pull / unittest-editable / macos / macos-job (gh) (trunk failure)
  AttributeError: '_OpNamespace' 'mkldnn' object has no attribute '_is_mkldnn_acl_supported'

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

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[Copilot]

Pull request overview

This PR fixes a silent data corruption bug in the CUDA backend delegate copy-back path. When AOTI produces contiguous output but the .pte serializes a different dim_order (e.g., from SDPA's efficient attention returning a transposed view), the runtime would byte-copy GPU data into an ETensor that interprets it with wrong strides. The fix detects stride mismatches and, when strides differ, copies GPU data to a temporary CPU buffer and then rearranges it element-by-element to match the ETensor's expected layout.

Changes:

• Unified copy_slimtensor_to_etensor_async and copy_slimtensor_to_etensor into a new _copy_slimtensor_to_etensor_impl helper with a fast path (strides match → raw byte copy) and a slow path (strides differ → element-wise rearrange via _strided_copy).
• Re-enables the accuracy assertion in test_non_pow2_head_dim_with_bool_mask that was previously gated behind a TODO comment.
• Adds a new test file test_output_stride_rearrange.py exercising both fast and slow copy paths.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
backends/cuda/runtime/utils.h	Core fix: adds _strides_match, _strided_copy, and _copy_slimtensor_to_etensor_impl; refactors public copy functions to call it
backends/cuda/tests/test_triton_sdpa_nan.py	Enables the previously disabled accuracy check for the non-pow2 head dim + bool mask test
backends/cuda/tests/test_output_stride_rearrange.py	New test file exercising the fast and slow copy-back paths

---

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

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

---

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

[Gasoonjia]

Thx for fixing the issue!