fix(vllm-omni): harden generated video output

[GuanLuo]

Summary

• Harden vLLM-Omni video output handling for every video model, independent of MiniMax-H3.
• Reject unsupported output formats instead of returning MP4 under a mismatched format contract.
• Reject ambiguous channel-first tensors and mismatched audio batches instead of silently guessing or dropping audio.
• Resolve audio sample rate from emitted output metadata, with a documented 24-kHz fallback.
• Make the opt-in H.264/AAC codec overlay build against the current Dynamo vLLM runtime layout.

Validation

• pre-commit run --files passed for all files in this PR.
• Python compile validation passed.
• The final stack passed a four-B200 MiniMax-H3 T2VA Kubernetes qualification using image digest sha256:5b820b31f5a4b9b14eed29952623787756a38b283e2e73e35b82f96c6af86f5a.
• The qualification produced a 10.125-second 448x256 H.264 video with synchronized, non-silent 32-kHz stereo AAC in 24.206 seconds.
• Local pytest was unavailable in the workspace venv; the image build and end-to-end qualification exercised the runtime path.

Stack

Review from bottom to top:

1. feat(vllm-omni): preserve generated video audio #13707 — preserve generated video audio
2. feat(vllm-omni): pass model-specific video parameters #13708 — pass model-specific video parameters
3. feat(vllm-omni): qualify MiniMax-H3 T2VA on B200 #13589 — qualify MiniMax-H3 T2VA on B200
4. fix(vllm-omni): harden generated video output #14691 — universal video-output hardening
5. fix(vllm-omni): strengthen MiniMax-H3 qualification #14692 — H3-specific qualification corrections

[dynamo-review-agent]

vLLM-Omni's existing channel-first convention is CFHW, so valid CFHW outputs with 3 or 4 frames now satisfy both dimension checks and are rejected. This turns normal short generated videos into failed responses; the modified test explicitly drops those frame counts.

🤖 AI Fix

Preserve the established CFHW interpretation for channel-first tensors, or obtain an explicit layout indicator before rejecting a tensor as ambiguous.

[dynamo-review-agent]

A rank-2 (batch, samples) tensor is the standard representation for a batch of mono waveforms. For two generated videos with audio shaped (2, samples), this condition no longer splits the batch and the formatter returns a failed video response despite the audio batch matching the video batch.

🤖 AI Fix

Accept rank-2 tensors whose leading dimension equals expected_count as batched mono audio outputs and split them along axis zero.

[dynamo-review-agent]

The change removes the model-config sample-rate fallback and always labels/encodes metadata-less audio as 24 kHz. Models whose vocoder configuration specifies another rate, such as the previously covered 48-kHz configuration, will have their raw waveform muxed at the wrong rate, causing incorrect duration, pitch, and video/audio synchronization.

🤖 AI Fix

Retain the model-config-derived sample-rate fallback after checking emitted metadata, using 24 kHz only when neither source provides a valid rate.

[dynamo-review-agent]

The (3, 4, 8, 10) parameter is equivalent to the retained (4, 3, 8, 10) case: both only prove that a 4-D tensor whose first two dimensions are both channel-sized hits the same is_cfhw and is_fchw rejection path. Either case would catch the previous silent CFHW preference, so the second parameter adds no distinct contract coverage.

🤖 AI Fix

Collapse the shape parametrization to a single representative ambiguous shape, for example [(4, 3, 8, 10)].

[GuanLuo]

Superseded: this universal hardening commit has been folded into #13707 so the base video/audio feature is internally complete.