[Bug]: --allowed-local-media-path is ignored for pure diffusion speech

Your current environment

Description

For pure diffusion speech models such as k2-fsa/OmniVoice, the --allowed-local-media-path option is accepted but is not propagated to the MediaConnector used to load ref_audio.

Requests using local file:// reference audio fail with:

RuntimeError: Cannot load local files without `--allowed-local-media-path`.

Environment

• vLLM: 0.26.0
• vLLM-Omni: main
• Model: k2-fsa/OmniVoice
• GPU: H200

Expected Behavior

Local audio under the directory specified by --allowed-local-media-path should be accepted.

Root Cause

The pure diffusion speech path creates MediaConnector() without passing the configured local media path.

Related PR

https://github.com/vllm-project/vllm-omni/pull/5788

Your code version

Your Code Version

v0.26.0

bfaaa3c2 (main, issue reproduced)
390c1a54 (fix, PR #5788)

🐛 Describe the bug

When serving a pure diffusion speech model such as k2-fsa/OmniVoice, --allowed-local-media-path is accepted by the CLI but is not propagated to the MediaConnector used by the diffusion speech handler.

As a result, a ref_audio file located inside the explicitly allowed directory is rejected with HTTP 500.

Reproduction

Start the server:

CUDA_VISIBLE_DEVICES=2 \
vllm serve k2-fsa/OmniVoice \
  --omni \
  --port 8108 \
  --trust-remote-code \
  --allowed-local-media-path /home/lcc/ref_audio

Place a valid reference audio file at:

/home/lcc/ref_audio/women_1.wav

Send a request:

curl http://127.0.0.1:8108/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "k2-fsa/OmniVoice",
    "input": "Hello, this is a voice cloning test.",
    "response_format": "wav",
    "ref_audio": "file:///home/lcc/ref_audio/women_1.wav"
  }' \
  --output output.wav

Expected behavior

The reference audio should be loaded because it is a child of the directory specified by --allowed-local-media-path.

Actual behavior

The request returns HTTP 500:

RuntimeError: Cannot load local files without `--allowed-local-media-path`.

Relevant traceback:

Diffusion speech generation failed: Cannot load local files without
`--allowed-local-media-path`.

Traceback (most recent call last):
  File "vllm_omni/entrypoints/openai/serving_speech.py", line 3641,
    in _create_diffusion_speech
    wav, sr = await self._resolve_ref_audio(request.ref_audio)
  File "vllm_omni/entrypoints/openai/serving_speech.py", line 2028,
    in _resolve_ref_audio
    wav_np, sr = await connector.fetch_audio_async(ref_audio_str)
  File "vllm/multimodal/media/connector.py", line 469,
    in fetch_audio_async
    return await self.load_from_url_async(...)
  File "vllm/multimodal/media/connector.py", line 441,
    in load_from_url_async
    return await future
  File "vllm/multimodal/media/connector.py", line 322,
    in _load_file_url
    raise RuntimeError(
RuntimeError: Cannot load local files without `--allowed-local-media-path`.

Root cause

The pure diffusion initialization path calls OmniOpenAIServingSpeech.for_diffusion() without passing allowed_local_media_path or allowed_media_domains.

Consequently, _resolve_ref_audio() creates an unconfigured connector:

if self._diffusion_mode:
    connector = MediaConnector()

The non-diffusion path correctly passes these settings from model_config.

A proposed fix propagates both media access settings through the pure diffusion speech initialization path.

Related PR: https://github.com/vllm-project/vllm-omni/pull/5788

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.