[FayeSpica]

Contributor intent (required human answers)

1. What outcome do you want to achieve?

Fix the Qwen3-TTS startup failure during FlashInfer warmup in vLLM 0.32.

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

Correct identification of synthetic MRv2 warmup requests.

3. How will this work move that variable?

Identify the new warmup requests added in vLLM 0.32 and skip text validation for them, preventing the startup error.

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

Qwen3-TTS users need the server to start successfully. Release CI needs the TTS performance tests to run.

Purpose

Relates to #9233.

vLLM 0.32 adds warmup requests with a new ID prefix. Omni does not recognize them as warmup requests and tries to process them as normal TTS input. Because they contain no text, Qwen3-TTS fails to start.

This change adds a shared function to recognize both warmup ID prefixes and uses it across MRv2 runners and model states. Warmup requests then follow the existing warmup handling. No user configuration or API changes are needed, so user guides and examples remain unchanged.

Test Plan

Environment: A100 80GB, vLLM 0.32.0, PyTorch 2.13.0+cu130, Omni 798fef99a + this patch.

• MRv2 and model-state regression tests:

  python -m pytest -q -m cpu \
    tests/worker_v2/test_omni_model_state.py \
    tests/worker_v2/test_omni_gpu_model_runner.py \
    tests/model_executor/models/higgs_audio_v3/test_mrv2_state.py \
    tests/model_executor/models/fish_omni/test_fish_omni_model_state.py \
    tests/model_executor/models/moss_tts/test_local_model_state.py

• Qwen3-TTS two-stage startup with FlashInfer:

  vllm serve Qwen/Qwen3-TTS-12Hz-0.6B-Base --omni --trust-remote-code \
    --host 127.0.0.1 --port 19233 \
    --stage-overrides '{"0":{"attention_config":{"backend":"FLASHINFER"}},"1":{"devices":"1"}}'
  curl --fail http://127.0.0.1:19233/health

• Ruff and diff hygiene on changed files.

Test Result

• MRv2 and model-state regression tests: 161 passed. Before the fix, six cases covering mixed-warmup validity, preprocessing and data-plane registration failed.
• Qwen3-TTS two-stage startup with FlashInfer: the unpatched build reproduced the missing-text-conditioning error. With the patch, FlashInfer warmup completed, both stages initialized, and /health returned HTTP 200.
• Ruff and diff hygiene: lint, formatting and git diff --check passed.

Not tested: B200 / 1.7B Base or speech generation.

AI assistance: Codex assisted with implementation, tests, validation, and the PR description.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-22T15:50:43.230499Z We couldn’t complete this review. Comment @codex review to try again.	a5bc5ab	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/ar_runtime.md, docs/design/module/model_integration.md.

Module owners: @tzhouam @fake0fan @Gaohan123

Routing: @tzhouam via module of the changed files, CODEOWNERS; @fake0fan via module of the changed files; @Gaohan123 via module of the changed files

@FayeSpica, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[amy-why-3459]

Centralizing the warmup check is a good cleanup, and the parametrized tests cover the new IDs well. Two points on the helper's matching.

[amy-why-3459]

run_mixed_prefill_decode_warmup takes req_id_prefix as a parameter, and on vLLM main other callers already pass different prefixes: flashinfer_sparse_mla_warmup.py uses "_sparse_mla_v2_warmup" and step5_dsa_warmup.py uses "_step5_dsa_warmup". Those requests would miss this tuple and go through the normal TTS / model-state path again, so the same startup failure would come back on the next vLLM bump for any model whose worker enables those warmups. Could the check key off the shared shape instead, e.g. req_id.startswith("_") and "_warmup" in req_id, or at least add the two known prefixes plus a test?

Also, every call site you replaced used to coerce with str(...) first, but the helper now calls .startswith on the raw value. Doing str(req_id).startswith(...) here would keep the old behavior for non-str IDs.

[FayeSpica]

Since NewRequestData.req_id is typed as str and upstream warmup IDs are constructed as strings, I think we can keep the direct check here.

[amy-why-3459]

This one also drops the getattr(request_data, "req_id", "") fallback that was here before. If anything still puts objects without req_id into scheduled_new_reqs (test doubles, or the native data-plane shims), this now raises AttributeError instead of falling through to register_request. Was that intended?

[FayeSpica]

This method is specific to MRv2. Entries in scheduled_new_reqs are required to have req_id

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit a5bc5ab34205 produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[linyueqian]

Reviewed a5bc5ab. vLLM 0.32 builds its mixed prefill/decode warmup requests as _v2_mixed_warmup_decode_ and _v2_mixed_warmup_prefill_ (vllm/v1/worker/gpu/warmup.py, run_mixed_prefill_decode_warmup), so matching both prefixes in one helper covers every warmup site that used the old _warmup_ check, and no raw startswith("_warmup_") is left under vllm_omni/. The tests cover both prefixes and a normal request id.

Static review only; I did not run the tests locally. Main CUDA passes at a5bc5ab (build 17883).