[JakeStevens]

Summary

Enable the re-inplace pass by default for the Optimum XNNPACK export recipe. Callers can still A/B or opt out with recipe_kwargs={"run_reinplace_pass": False}. The pass currently rewrites eligible index_put calls.

Whisper Tiny A/B

Exported openai/whisper-tiny twice from the same Optimum revision using the local ExecuTorch source containing the output-liveness fix: re-inplace disabled vs enabled. With the 30-second trimmed audio sample used by tests/models/test_modeling_whisper.py, FP32 and four CPU threads, timed full transcription (encoder plus autoregressive decoder) after one warmup per artifact. Each variant had three measured runs; times exclude export, model loading and feature extraction.

Variant	Run times (s)	Median (s)
Off	18.5914, 18.5684, 18.8215	18.5914
On	4.8947, 4.7059, 4.7303	4.7303

This is 3.93x faster (74.6% lower median latency) on this host. The decoder changed from 16 aten::index_put.out instructions to 16 aten::index_put_ instructions. All runs produced the same 433-character transcript (SHA-256 0c9c4e433d2d0df6e3b034b8462e8ab8b018f967125c6ccd9065e53b2d55b5f6). As a control, off/on exports of SmolLM2-135M and Qwen2.5-0.5B with custom KV cache produced byte-identical PTE files; their cache updates use llama::update_cache, so this pass does not change those models.

Validation: the existing Whisper Tiny export/transcription integration test passed with the new default; Black, Ruff, and git diff --check passed.

Merge dependency

Draft until pytorch/executorch#23488 is merged and available to supported ExecuTorch versions. That PR fixes output liveness: without it, re-inplace can change a returned original tensor. The current executorch>=1.0.0 dependency does not ensure that fix is present, so before merging this PR we must raise the minimum version to a release containing the fix or add a reliable compatibility guard.

Authored with assistance from Codex (OpenAI AI assistant).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.