Qwen3-Omni on vLLM: test video with no audio track, not just silent audio

A video that plays without sound is not necessarily a video without an audio track. For a Qwen3-Omni pipeline on vLLM, that distinction deserves its own input-admission check—not just a “silent video” label in the asset library.

The vLLM v0.28.0 release notes, under Model Support → Correctness, report a fix for “a Qwen3-Omni crash on video without an audio track,” linking pull request #48420. That is a reported crash fix, not a new video-quality claim or proof that a particular deployment now accepts these assets. The checks below are proposed validation, not results from a test run.

Keep two different fixtures

For this regression check, use the same visual content where practical, but distinguish the files by their actual track structure:

Fixture What to verify What it can test
Video with no audio track No audio track is present at the intended ingestion boundary The input condition named in the Qwen3-Omni crash fix
Video with a silent audio track An audio track is present, but its content is silent A comparison case that still contains audio structure

A successful run of the second file does not cover the first condition. Listening to the clip—or muting playback—is not a substitute for checking its tracks.

Make direct admission a measured decision

Before allowing no-audio-track assets through the intended Qwen3-Omni path:

  1. Inspect both fixtures and the preprocessing output. Confirm that the no-audio-track case still has no audio track when it reaches ingestion. If preprocessing adds one, that run has not tested the proposed direct path.
  2. Run each fixture through the path you intend to use on vLLM v0.28.0. Record the actual serving version, model identifier, preprocessing settings and asset format alongside completion or failure. Keep the two outcomes separate.
  3. Use the no-audio-track result to decide admission. A pass supports admitting the tested asset class under those tested constraints. A continuing crash, or only a silent-audio pass, is not enough to enable it on the strength of the release note.

If your pipeline already adds synthetic audio as a workaround, make its removal conditional on that direct-path check. The release note does not establish that this workaround was necessary, common or behaviorally equivalent; there is no reason to introduce it solely because it appears in a validation discussion.

Keep quality acceptance separate

Crash-free processing answers “can this tested input complete?” It does not answer “did Qwen3-Omni understand the video correctly?” Keep the workflow’s output-quality acceptance check separate, and do not extend a passing fixture to untested codecs, other Qwen variants or performance claims.

The useful change is narrow: preserve no audio track as a distinct input category, and test it directly before changing what the pipeline admits or how it alters incoming assets.

I would distinguish two conclusions in the test report: “this deployment accepts the fixture” and “the upgrade resolved a reproduced failure.” The vLLM v0.28.0 release notes identify a Qwen3-Omni crash on video without an audio track, but a successful run on the new version alone cannot establish the second conclusion.

For upgrade attribution, I would propose a small comparison in an isolated test environment: run the no-audio-track and silent-audio fixtures on both the previously deployed version and v0.28.0, holding the Qwen3-Omni checkpoint, request, preprocessing and other settings fixed where possible. Record any unavoidable configuration differences rather than treating the version as the only variable.

If the no-audio-track fixture fails on the old stack and completes on the new one, while the silent-audio control completes on both, that supports an upgrade-related resolution for this fixture. It still does not isolate pull request #48420 from other release changes. If both versions complete, the fixture can support the current admission decision, but it has not reproduced the reported regression. If the old stack is unavailable, label attribution untested; that need not prevent a separately justified admission decision on the new stack.

This comparison is a proposed evaluation, not an additional requirement from the release notes or a test result. It keeps a useful local acceptance check from becoming an unsupported claim about what the fix changed.