XNNPACK: Run recommended pre-partition transforms in AOT flow

[mansnils]

Wire get_transform_passes() into the XNNPACK example AOT compiler so official exports run the same recommended pre-partition graph transforms used by tests.

Update XNNPACK docs and README snippets to show the optional transform_passes stage and clarify that it is recommended in general for XNNPACK before partitioning.

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/20446

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI jobs will have longer queue times due to CI migration

❌ 4 New Failures, 4 Unrelated Failures

As of commit f0f41be with merge base 82a605d ():

NEW FAILURES - The following jobs have failed:

• pull / unittest / linux / linux-job (gh)
  RuntimeError: Command docker exec -t a6665e4956a67465630db2c9e61984445441c965091c39e68b850c18810b9c3b /exec failed with exit code 1
• pull / unittest-editable / linux / linux-job (gh)
  RuntimeError: Command docker exec -t 9cc39d1e9cd2408d7240665cdfaa8512eaf744c79b5d828759411ca61abaf9e6 /exec failed with exit code 1
• trunk / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh)
  RuntimeError: Command docker exec -t a40a4b547bdb609c43a02fc695ec4291b619efc3507f13a4fca20e72c923d9f7 /exec failed with exit code 137
• trunk / unittest-release / linux / linux-job (gh)
  RuntimeError: Command docker exec -t a1397e66a33d5a80a737b0fbc9e192a6f8eaa7f4810044f99f98117f8acdcd3f /exec failed with exit code 1

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

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

• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh) (trunk failure)
• pull / unittest / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1
• pull / unittest-editable / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1
• trunk / unittest-release / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1

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

[Copilot]

Pull request overview

Wires XNNPACK’s recommended pre-partition graph transform passes into the XNNPACK example AOT export flow, and updates user-facing docs/snippets so “official” exports follow the same recommended lowering pipeline as tests prior to partitioning/delegation.

Changes:

• Run get_transform_passes() via transform_passes=... in the XNNPACK example to_edge_transform_and_lower(...) call.
• Update XNNPACK docs/READMEs to show and recommend the optional pre-partition transform stage for better pattern exposure to the partitioner.
• Add an additional copyright header line in the XNNPACK example AOT compiler script.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
examples/xnnpack/README.md	Documents that the example compiler runs the recommended pre-partition transform stage.
examples/xnnpack/aot_compiler.py	Applies get_transform_passes() in the AOT lowering call before XNNPACK partitioning.
docs/source/backends/xnnpack/xnnpack-quantization.md	Updates lowering snippet to include transform_passes and explains why it’s recommended for XNNPACK.
docs/source/backends/xnnpack/xnnpack-overview.md	Updates lowering snippet to include transform_passes and explains why it’s recommended for XNNPACK.
backends/xnnpack/README.md	Updates the “manual” lowering snippet to run edge.transform(get_transform_passes()) before partitioning.

---

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

[mansnils]

@claude review this PR

[claude]

Claude finished @mansnils's task in 1m 37s —— View job

---

PR Review: XNNPACK pre-partition transforms in AOT flow

• Read PR diff
• Examine get_transform_passes() and AOT compiler
• Check docs/snippet consistency
• Post review feedback

I reviewed the diff against origin/main. This is a small, well-scoped change that wires the recommended XNNPACK pre-partition transforms into the example AOT compiler and documents the optional transform_passes stage. LGTM — no blocking issues.

Correctness ✅

• get_transform_passes() (in backends/xnnpack/utils/configs.py:122) returns List[PassType]. This matches both API signatures it's passed to:
  • to_edge_transform_and_lower(..., transform_passes=...) accepts a Sequence[PassType] (exir/program/_program.py:1253). ✅
  • EdgeProgramManager.transform(passes) accepts a Sequence[PassType] (exir/program/_program.py:1537). ✅ — so the backends/xnnpack/README.md snippet edge.transform(get_transform_passes()) is valid.
• The pass set is currently just RemoveNoopExpandCopyPass(), so this is low-risk and applies uniformly to fp32 and quantized exports in examples/xnnpack/aot_compiler.py.

Observations (non-blocking)

1. Doc text duplication. The identical 3-line paragraph ("The transform_passes stage is optional in the generic lowering API. For XNNPACK, get_transform_passes() is recommended in general...") is repeated verbatim in both xnnpack-overview.md and xnnpack-quantization.md. That's acceptable since each page should stand alone, just flagging it.

2. Wording: "recommended in general". The phrase "recommended in general" reads slightly redundant. Consider "generally recommended" for clarity, e.g. "For XNNPACK, get_transform_passes() is generally recommended because it runs pre-partition graph transforms that expose supported patterns to the partitioner." Minor/optional.

3. backends/xnnpack/README.md snippet uses the legacy flow. That snippet uses to_edge(...) + to_backend(XnnpackPartitioner()), whereas the to_edge_transform_and_lower docstring states it is "the only API that should be used to lower to the edge dialect" for backends with non-decomposed ops, and "the primary recommended way to lower to the CPU based XNNPack backend" (exir/program/_program.py:1277-1281). This is pre-existing and out of scope for this PR, but since you're touching that snippet it may be worth aligning it with the recommended API in a follow-up.

4. Copyright header. Adding the Arm copyright line to examples/xnnpack/aot_compiler.py for a 3-line change is consistent with Arm's contribution conventions — no concern.

CI note

The reported new failures appear to be infra/docker-related (docker exec ... failed with exit code 1/137) rather than related to these doc/wiring changes, and several are marked as broken-trunk failures present on the merge base. Worth a rebase onto viable/strict to clear the broken-trunk ones.
· branch misc