[Bugfix][Refactor] Migrate Voxtral TTS config and parser registry

[yuanheng-zhao]

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

Purpose

Part of #3066

This PR

• refactored voxtral TTS config and parser registration to a unified place under vllm_omni/transformers_utils/, as consistent to upstream vllm
• removed the following unrelated logging/warning even if we serve models other than voxtral TTS

(APIServer pid=928653) WARNING 04-23 05:55:32 [config.py:347] Config format `mistral` is already registered, and will be overwritten by the new parser class `<class 'vllm_omni.model_executor.models.voxtral_tts.configuration_voxtral_tts.VoxtralTTSConfigParser'>`.
(APIServer pid=928653) INFO 04-23 05:55:32 [config.py:358] Registered config parser `<class 'vllm_omni.model_executor.models.voxtral_tts.configuration_voxtral_tts.VoxtralTTSConfigParser'>` with config format `mistral`

NOTE:
The following error happened on main branch and voxtral TTS repo resolution is utilizing a fallback path rather than it:

(APIServer pid=1263900) ERROR 04-23 14:31:03 [repo_utils.py:47] Error retrieving safetensors: 'mistralai/Voxtral-4B-TTS-2603' is not a safetensors repo. Couldn't find 'model.safetensors.index.json' or 'model.safetensors' files., retrying 1 of 2
(APIServer pid=1263900) ERROR 04-23 14:31:05 [repo_utils.py:45] Error retrieving safetensors: 'mistralai/Voxtral-4B-TTS-2603' is not a safetensors repo. Couldn't find 'model.safetensors.index.json' or 'model.safetensors' files.

This PR keep the path intact and won't include the above error fallback in scope.

cc @princepride , @linyueqian , @y123456y78

Test Plan

1. Running any model other than Voxtral-TTS to check if any unrelated voxtral-tts logging/warnings appear in init logs or not.
2. Existing e2e tests for Voxtral-TTS

pytest -s tests/e2e/offline_inference/test_voxtral_tts.py tests/e2e/online_serving/test_voxtral_tts.py

3. Voxtral-TTS offline examples

Test Result

1. Run with Ming

vllm serve Jonathan1909/Ming-flash-omni-2.0 \
    --omni \
    --stage-configs-path vllm_omni/model_executor/stage_configs/ming_flash_omni_tts.yaml \
    --port 8091 \
    --log-stats

Voxtral-TTS loggings/warnings no longer appear in init logs.

2. pytest

# pytest -s tests/e2e/offline_inference/test_voxtral_tts.py tests/e2e/online_serving/test_voxtral_tts.py

8 passed, 16 warnings in 537.44s (0:08:57)

3. offline example results: please check my subsequent comments in this PR

---

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)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d2d5acc830

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[yuanheng-zhao]

Voxtral-TTS offline examples

Script: examples/offline_inference/voxtral_tts/README.md

Streaming with neutral_female voice preset:

python3 examples/offline_inference/voxtral_tts/end2end.py \
    --num-prompts 32 --concurrency 8 --streaming --write-audio --voice neutral_female \
    --model mistralai/Voxtral-4B-TTS-2603 \
    --text "That eerie silence after the first storm was just the calm before another round of chaos, wasn't it?"

Out logs (partial)

Request 0: saved 174720 samples (7.28s) to output_audio/tts_output_0.wav
  Request 0 chunk 0: no_wait | arrived=0.0ms | chunk_dur=400.0ms
  Request 0 chunk 1: no_wait | arrived=42.2ms | chunk_dur=400.0ms
  Request 0 chunk 2: no_wait | arrived=99.4ms | chunk_dur=400.0ms
  Request 0 chunk 3: no_wait | arrived=135.7ms | chunk_dur=400.0ms
  Request 0 chunk 4: no_wait | arrived=209.9ms | chunk_dur=400.0ms
  Request 0 chunk 5: no_wait | arrived=514.1ms | chunk_dur=2000.0ms
  Request 0 chunk 6: no_wait | arrived=798.3ms | chunk_dur=2000.0ms
  Request 0 chunk 7: no_wait | arrived=1039.9ms | chunk_dur=1280.0ms
Request 0: TTFA=0.5367s | Generation=1.5769s | Audio=7.28s | RTF=4.6167 | WaitRate=0.00% (0/8)

All requests: Generation=1.5870s | TotalAudio=7.28s | Concurrency=1 | AvgTTFA=0.5367s | RTF(total)=4.5872 | RTF(per-request)=4.6167 | WaitRate=0.00% (0/8)

tts_output_0.wav

32 prompts, 8 concurrent requests per wave, streaming with neutral_female voice:

python3 examples/offline_inference/voxtral_tts/end2end.py \
    --num-prompts 32 --concurrency 8 --streaming --write-audio --voice neutral_female \
    --model mistralai/Voxtral-4B-TTS-2603 \
    --text "That eerie silence after the first storm was just the calm before another round of chaos, wasn't it?"

Out logs (partial)

All requests: Generation=9.6359s | TotalAudio=239.68s | Concurrency=8 | AvgTTFA=0.5348s | RTF(total)=24.8738 | RTF(per-request)=3.4928 | WaitRate=0.00% (0/262)

[yuanheng-zhao]

PTAL @princepride , @Gaohan123 , @linyueqian
cc @y123456y78

[princepride]

LGTM