[yuanheng-zhao]

PLEASE FILL IN THE PR DESCRIPTION HERE.

Contributor intent (required human answers)

All four questions must be answered by the human contributor before submission.
An agent may reuse explicit human answers with a public-safe source reference;
it must not invent, infer, or prefill missing answers. If an answer is missing
or materially ambiguous, ask the human and wait. Do not publish private
conversation content without authorization.

1. What outcome do you want to achieve?

Support deploying inclusionAI/Ming-3.0-tts-tiny with vLLM 0.32.

2. What is the highest-leverage variable for that outcome?

Successful model loading, service startup, speech generation, and basic features.

3. How will this work move that variable?

Add the integration needed to deploy and generate audio from unsupported.

4. Whose customer or user needs does this address, and what are those needs?

Users who need this model to generate audio for different TTS scenarios.

Purpose

Add Ming-3.0-TTS Tiny support (inclusionAI/Ming-3.0-tts-tiny), including checkpoint loading, AR/Flow decoding, streaming AudioVAE, voice cloning, instruction controls, and Podcast requests.

Co-authored with @runheyang

Test Plan

vLLM Version: 0.32.0

vLLM-Omni Commit: rebased on top of db23219

• Unit tests, contract/transport, shared Ming, adapter, and speech API regression tests
• Deployment and generation, check recipes/inclusionAI/Ming-3.0-TTS.md

Test Result

Example output files:

model_card_clone.wav
model_card_dialect.wav

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.

Note

To avoid agent spam, vLLM-Omni limits contributors without write access to 5 open PRs. If this blocks well-intentioned critical work, please contact a committer to request bypass list access.
(anything written below this line will be removed by GitHub Actions)

[amy-why-3459]

Re-checked 4e198958..a43e5029. The seeding and step-count points look addressed: _get_request_generator now seeds the per-request torch.Generator with generator.seed() instead of torch.seed(), so the process-wide RNG is left alone, and DEFAULT_CFM_STEPS now drives the graph sampler, the warmup/request solver noise, and the solver_noise shape check (derived from len(t) - 1).

The y0 layout point on cfm_cudagraph.py is still open at this head: the graph path still draws (B, patch_size, latent_dim) directly (L130-135), while the eager FlowLoss.sample and _draw_request_cfm_noises draw (B, latent_dim, patch_size) and transpose. With patch_size > 1, a fixed tts_local_seed can still get different first-step noise depending on whether the step runs through the graph or eager path.

[yuanheng-zhao]

PTAL @hsliuustc0106 @Gaohan123 The AMD/Intel CI error should be unrelated

[Sy0307]

Three inline comments on prefill handling, eager-mode configuration, and an unused shared AudioVAE branch. These are based on static code tracing; I have not run model/GPU validation.

[amy-why-3459]

Looked at the two new commits (59737ef chunked-prefill conditioning, 3db61eb CFM graph gating). The chunked-prefill path looks right to me: speaker/prompt-audio positions are global, each slice is clipped against [offset, offset + len), the reference work happens once per request, and the cache is dropped in on_requests_finished. One concern on the shared Ming module, inline.