Preserve node output dtype when lifting scalars to attrs in XNNPACK quantizer

[shoemoney]

Fixes #22062

_convert_scalars_to_attrs in backends/xnnpack/quantizer/xnnpack_quantizer_utils.py lifts every scalar argument of aten.add.Tensor and aten.mul.Tensor to a buffer created with torch.tensor(float(arg)), which is always float32. XNNPACKQuantizer.transform_for_annotation runs this unconditionally, so an integer chain such as x[:, torch.arange(4) + 0] (the position-ids pattern emitted by Hugging Face models) has its int64 add promoted to float32, and running the prepare_pt2e output fails with IndexError: tensors used as indices must be long, int, byte or bool tensors. The fix creates the lifted constant with the node's own output dtype, torch.tensor(args[i], dtype=n.meta["val"].dtype), which the function already relies on for fake_mode. Float behavior is unchanged.

Added test_int64_scalar_add_used_as_index to backends/xnnpack/test/quantizer/test_xnnpack_quantizer.py: it exports the index pattern above, runs transform_for_annotation, asserts the lifted constant keeps dtype int64, and executes the prepared module. The test fails with the float32 dtype assertion before this change and passes after. Neighboring scalar tests (test_add_mul_scalar, test_add_mul_long, test_mul_float32_max) still pass. flake8 and ufmt clean on both files.

Written in conjunction with my pair programmer Claude.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22065

• 📄 Preview Python docs built from this PR

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

❌ 16 Cancelled Jobs

As of commit 1f22872 with merge base cff6f4d ():

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• Build Presets / apple (ios-simulator) / build (gh)
• Build Presets / apple (ios) / build (gh)
• Build Presets / apple (llm) / build (gh)
• Build Presets / apple (macos) / build (gh)
• Build Presets / apple (profiling) / build (gh)
• Build Presets / apple (pybind) / build (gh)
• Lint / Get changed files / get-changed-files (gh)
• Lint / link-check / lint-file-size (gh)
• Lint / link-check / lint-urls (gh)
• Lint / link-check / lint-xrefs (gh)
• periodic / gather-models (gh)
• pull / CI run decision / decide (gh)
• pull / Get changed files / get-changed-files (gh)
• Test CoreML Backend / Get changed files / get-changed-files (gh)
• Test Cortex-M Backend / Get changed files / get-changed-files (gh)
• Test NXP Backend / Get changed files / get-changed-files (gh)

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

[meta-cla]

Hi @shoemoney!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: shoemoney / name: Jeremy Schoemaker (1f22872)

[Copilot]

Pull request overview

This pull request fixes an XNNPACK quantizer graph-rewrite bug where scalar constants lifted into module buffers were always created as float32, unintentionally promoting integer subgraphs (notably index expressions) and causing prepare_pt2e runtime failures. The fix preserves the original node’s output dtype when materializing the lifted scalar tensor, keeping integer index chains intact.

Changes:

• Update _convert_scalars_to_attrs to create lifted scalar buffers with dtype=n.meta["val"].dtype instead of forcing float32.
• Add a regression test covering an int64 scalar-add index pattern (torch.arange(4) + 0) to ensure dtype preservation and successful execution after prepare_pt2e.

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	Preserve node output dtype when lifting scalar constants into _tensor_constant_* buffers.
backends/xnnpack/test/quantizer/test_xnnpack_quantizer.py	Add regression test ensuring lifted scalar constants used in indexing remain int64 and prepared module runs.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[shoemoney]

@pytorchbot label "release notes: xnnpack"

[shoemoney]

/easycla

[meta-cla]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks!

[JakeStevens]

thanks for the fix!