[Bugfix][Audio] Restore soundfile-first automatic decoding

[AndreasKaratzas]

Speech jobs passed in AMD build 12635 before PR #51826 made TorchCodec the preferred automatic audio decoder. A local GPU git bisect identified its merge commit, 6a039f465e37, as the cause of the test_long_audio_request failures in build 12653 on MI300 and MI355. With TorchCodec 0.10 and FFmpeg 4.4.2, the Mary Vorbis fixture gains 192 padding samples at 16 kHz, changing repeated-audio boundaries and reducing recognized phrase repetitions from ten to seven.

• Restore automatic decoding order to soundfile → TorchCodec → PyAV, preserving the previous waveform length for soundfile-supported formats.
• Document the restored order and explicit TorchCodec selection for concurrent decoding; retain TorchCodec fallback for unsupported formats such as video containers.

Prepared with AI assistance. The recorded upstream duplicate audit found no matching open fix for this decoder-priority regression.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[mergify]

Documentation preview: https://vllm--55642.org.readthedocs.build/en/55642/

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Team

Run ID: ef360081-c72c-4af3-a842-90f31931e645

📥 Commits

Reviewing files that changed from the base of the PR and between f43ef15 and 7b5f6e5.

📒 Files selected for processing (3)

• docs/features/multimodal_inputs.md
• tests/multimodal/media/test_audio.py
• vllm/multimodal/media/audio.py

Included review availability: Your plan provides up to 10 included reviews per hour; 9 remain after this review.

---

📝 Summary

Summary by CodeRabbit

• Bug Fixes

  • Improved automatic audio decoding to preserve waveform lengths for supported formats, including Vorbis audio.
  • Added fallback handling for formats unsupported by the preferred decoder, including video containers.
  • Improved compatibility on platforms where optional audio decoding components are unavailable.

• Documentation

  • Updated audio decoding documentation to clarify the automatic backend preference and fallback behavior.

Walkthrough

The audio auto backend now prefers soundfile, then uses torchcodec and PyAV as fallbacks. A Vorbis length regression test and updated multimodal audio documentation cover the new order.

Changes

Audio backend order

Layer / File(s)	Summary
Automatic backend selection vllm/multimodal/media/audio.py	load_audio now tries soundfile, then torchcodec, then PyAV. Import and format-detection failures continue through the fallback chain.
Validation and documentation tests/multimodal/media/test_audio.py, docs/features/multimodal_inputs.md	The test checks that default Vorbis decoding matches soundfile output. The documentation describes the new order and platform behavior.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to 6eb61

Automatic audio decoding now preserves soundfile-supported waveform lengths while retaining fallbacks for unsupported formats. The implementation, regression coverage, and documentation are aligned, with no remaining merge-readiness risk identified.

Suggested reviewers: jaredforreal, jperezdealgaba

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 4 functions across 2 files. (1 skipped: 1 …
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly and concisely identifies the main change: restoring soundfile-first automatic audio decoding.
Description check	✅ Passed	The description directly explains the decoder regression, the affected audio behavior, the fix, documentation updates, and validation results.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndreasKaratzas]

/ci run

[github-actions]

✅ Triggered Buildkite CI #87481 for commit 6eb61b478f14.