Add Cosmos3 sound generation

[MaciejBalaNV]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (Optional) Release notes update. If your change is user-facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[linyueqian]

Thanks for this, Maciej. I read the full diff and ran the PR end to end on a Cosmos3-Nano deployment (offline Omni(), HTTP /v1/videos/sync, and a cfg_parallel_size=2 configuration). The feature works: stereo 48 kHz audio comes out of both the offline path and the HTTP endpoint, the AAC mux is correct on the server side, and the tuple-output path through CFGParallelMixin.combine_cfg_noise operates cleanly under CFG-parallel.

A few things worth addressing before merge. Specific suggestions inline; here is the summary:

🔴 Blocking

1. Two unit tests in tests/diffusion/models/cosmos3/test_cosmos3_transformer.py fail on this branch (test_forward_returns_video_prediction, test_forward_returns_video_and_sound_predictions). Both look like test-side oversights rather than model bugs.
2. The PR base is mbala/cosmos3_model, which has since landed as vllm-project#3454. A rebase onto current main produces conflicts in 4 files. Once that is done, CI can fairly evaluate the change.

🟡 Important
3. _is_sound_request accepts six aliases for what is effectively one user-facing flag. Narrowing this surface reduces the chance of a "set the wrong key, get silent video-only output" UX failure.
4. The duplicate self.time_embedder initialization at transformer_cosmos3.py:1003 is dead dtype configuration (later overridden by post_load_weights().to(torch.float32)), but it is confusing to read.

💡 Suggestion
5. The joint (video, sound) pack-then-step in diffuse._step relies on the scheduler being linearly separable per element. True for flow matching, but the assumption is worth a one-line comment.

Test evidence

• Offline T2V+sound, two prompts, 4.00 s stereo 48 kHz: distinct, prompt-responsive audio.
• HTTP /v1/videos/sync with generate_sound=true sound_duration=4.0: 1.9 MB mp4 with H264 video (2.04 s) + AAC stereo (4.02 s), muxed server-side.
• cfg_parallel_size=2 on 2 GPUs: produces the same shape audio cleanly.

[bastefaniak]

Closing, because already merged to vllm-omni main