V2 quantizer: fix IO-boundary shared clusters left in float

[rascani]

Summary:
Shared-op clusters (e.g. cat, view, reshape) on the quantized IO boundary were silently left in float by the composable TOSA quantizer (_TOSAQuantizerV2), causing them to fall off the Ethos-U integer delegate onto CPU.

SharedQspecQuantizer propagates a qspec only from already-quantized neighbors. A cluster whose only quantized neighbors are a uint8 model input (intentionally skipped by _skip_shared_qspec_from_io to confine uint8 to the IO boundary) and/or an input-state placeholder with no output_qspec had no qspec to propagate, so it was rejected and remained in float.

The fix adds _is_quantized_io_boundary, which detects annotated placeholder/output nodes that signal the cluster is on the quantized data path even when their qspec is filtered. _get_shared_clique now returns a touches_quantized_io flag alongside the usual results. When _annotate_shared_cluster finds an empty adjacent_qspecs but a boundary-touching cluster, it initiates quantization from the global config input-activation qspec instead of rejecting. _TOSAQuantizerV2.set_global now also propagates to shared_qspec_quantizer.global_config so the fallback is wired automatically.

This restores the correctness fix from D107320847, which was abandoned because its other fix (parameter-operand weight misclassification) had already been resolved via the is_weight PARAMETER_TARGETS refactor.

This change was developed with assistance from Claude.

Differential Revision: D108662081

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20291

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 1 Cancelled Job

As of commit c0ac9b6 with merge base e257a71 ():

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

CANCELLED JOB - The following job was cancelled. Please retry:

• trunk / test-models-macos-coreml (vit) / macos-job (gh)
  ##[error]The operation was canceled.

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

[meta-codesync]

@rascani has exported this pull request. If you are a Meta employee, you can view the originating Diff in D108662081.

[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.

[AdrianLundell]

Thanks for the fix!

For some context, ideally we would just leave all nodes now handled by the SharedQspecQuantizer un-annotated and just let them be handled by dtype propagation, the reason it is done this way is mainly to ensure we know what nodes are quantized and not at partition-time. If we could do that in a more clever way maybe we could avoid the SharedQspecQuantizer altogether.

[digantdesai]

Thanks

[meta-codesync]

@rascani has imported this pull request. If you are a Meta employee, you can view this in D108662081.