XNNPACK: Fix cat qparams for quantized ViT token paths

[mansnils]

Bug: Cat annotation always used the first input as the shared qparam source. In DeiT Tiny, the first input to token concatenation is the class-token path:

cls_token -> expand -> cat

The patch-token path is a later cat input:

conv -> flatten -> transpose -> cat

The conv output has annotated activation qparams, and flatten/transpose are qparam-preserving view/layout ops. They should carry the same qparams with SharedQuantizationSpec. Anchoring cat qparams on the class-token path can leave the patch-token static transpose with different input/output qparams. XNNPACK rejects this during runtime initialization because static transpose only reorders bytes.

Fix: Choose the first cat input that traces through qparam-preserving ops to annotated output qparams, falling back to the first input otherwise. Also propagate shared qparams through reshape and transpose so static transpose nodes keep identical input and output qparams.

Add a quantized XNNPACK regression covering conv, flatten, transpose, cat, and add.

cc @GregoryComer @digantdesai @cbilgin @freddan80 @per @zingo @oscarandersson8218 @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20280

• 📄 Preview Python docs built from this PR

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

❌ 3 New Failures

As of commit 194bb19 with merge base e88fd04 ():

NEW FAILURES - The following jobs have failed:

• Cadence Build & Test / hifi-build / hifi4 (gh)
  Input required and not supplied: aws-region
• Cadence Build & Test / vision-build / vision (gh)
  Input required and not supplied: aws-region
• pull / test-voxtral-realtime-xnnpack-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 0b8d2d8897492fd303d703f1a7213ee3e770da1fe5b88a5d99c3cdd51a1c8972 /exec failed with exit code 1

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 quantization annotation bug in the XNNPACK backend where cat shared quantization parameters were always anchored on the first input, which can be incorrect for quantized ViT token concatenation patterns (e.g., DeiT Tiny). It adjusts the cat qparam source selection to prefer an input that traces through qparam-preserving ops to an already-annotated activation producer, and extends qparam propagation through additional view/layout ops to keep static transpose nodes’ input/output qparams identical.

Changes:

• Update cat quantization annotation to select a better shared-qparam “source” input (prefer inputs tracing to annotated activation qparams).
• Treat aten.transpose.int and aten.reshape.default as qparam-preserving ops for annotation propagation.
• Add a quantized XNNPACK regression test covering conv -> flatten -> transpose -> cat -> add consistent with ViT token paths.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
backends/xnnpack/quantizer/xnnpack_quantizer_utils.py	Improves cat qparam-source selection and extends qparam propagation through transpose/reshape.
backends/xnnpack/test/ops/test_cat.py	Adds a quantized regression test reproducing the ViT-style token concat path and validating successful lowering/execution.

---

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