feat(vllm-omni): qualify MiniMax-H3 T2VA on B200 #13937

Feature request

Qualify MiniMax-H3 text-to-video-with-audio (T2VA) through Dynamo's public video API on four NVIDIA B200 GPUs.

Provide a reproducible aggregated deployment and validation path that proves Dynamo can generate synchronized video and audio with MiniMax-H3. The T2VA configuration should load only the shared FL2VA checkpoint partition rather than the separate reference-conditioned partition.

Motivation

MiniMax-H3 support needs more than a model launcher: it needs a repeatable end-to-end qualification that exercises the public API and verifies the generated media. Without a checked-in validation path, model or runtime changes can silently break the video stream, audio stream, codec configuration, duration, or synchronization.

Proposed scope

• Add an aggregated launcher using four B200 GPUs with the required DiT and text-encoder parallelism.
• Pass MiniMax-H3 generation controls through the public video request as model-specific top-level fields.
• Add a deterministic HTTP validation script that generates a ten-second T2VA sample.
• Inspect the returned media and assert its video, audio, duration, resolution, frame-rate, sample-rate, channel, and non-silence properties.
• Document the required Dynamo and vLLM-Omni revisions, runtime/container setup, codec overlay, model defaults, hardware requirements, and deferred workflows.

Acceptance criteria

• A checked-in launcher starts the MiniMax-H3 T2VA configuration on four B200 GPUs.
• A request through Dynamo's public video API returns a playable MP4.
• The MP4 contains 448×256 H.264 video at 24 FPS.
• The MP4 contains non-silent 32-kHz stereo AAC audio.
• The validation confirms the expected ten-second duration and synchronized video/audio output.
• The exact source revisions, image, model, hardware, and commands required to reproduce the result are documented.
• Reference-conditioned or otherwise deferred MiniMax-H3 workflows are clearly identified as follow-up work.

Related work

• Parent diffusion tracker: [Roadmap]: Track Dynamo 2.0 diffusion work #13935
• Disaggregated media persistence: [FEATURE]: (vllm-omni): final-stage worker should write media-output-fs-url instead of shipping frames to the router #13805