[Bug-fix] Fix Bugs in Qwen3/Qwen2.5 Omni Rebased Support

[tzhouam]

Purpose

This PR is to fix the 2 bug for Qwen 3 Omni and Qwen 2.5 Omni support.

1. The stage stuck at retrieveing data from queue.
2. The multimodal list error for Qwen 2.5 Omni.

Test Plan

Test both model on all modalities they support.

For Qwen 2.5:

export PYTHONPATH="""YOUR VLLM OMNI DIR"""
# export HF_HOME="""YOUR HF HOME DIR"""
cd vllm_omni/model_executor/models/qwen2_5_omni

Modify the "--query-type" in run_single_prompt.sh, the values include mixed_modalities, use_audio_in_video, multi_audios, text.

bash run_single_prompt.sh

For Qwen 3:

export PYTHONPATH="""YOUR VLLM OMNI DIR"""
# export HF_HOME="""YOUR HF HOME DIR"""
cd vllm_omni/model_executor/models/qwen3_omni

Modify the "--query-type" in run_single_prompt.sh, the values include text, use_audio, use_image, use_video.

bash run_single_prompt.sh

Test Result

The Qwen 3 passed
The Qwen 2.5 passed

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, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (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.

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".

[chatgpt-codex-connector]

Restore stage-level error propagation

The per-batch try/except that previously caught failures in stage_engine.generate and the result-emission loop is now commented out, so any exception during generation or serialization will bubble out of _stage_worker and tear down the stage process without enqueuing {error: ...} responses. When a model throws (e.g., invalid multimodal input or SHM dump failure), the orchestrator will wait indefinitely for results because no downstream error is produced.

Useful? React with 👍 / 👎.

[Gaohan123]

lgtm, thanks!