Highlights
vLLM-Omni v0.34.0rc1 is the first release candidate for v0.34.0. It aligns with vLLM 0.34.0 and requires ModelRunner V2 for AR and generation stages. Changes are relative to v0.33.0rc1.
Runtime and realtime serving
- Add session-aware routing across independently launched duplex stages and nodes. #10422
- Add Mooncake KV and hidden-state caching for Qwen3-Omni duplex sessions. #10321
- Route diffusion TTS requests through model-specific adapters, starting with OmniVoice. #8011
Models and generation
- Add Haru-4B speech-to-speech, MiniMax-H3-Image generation and editing, and PAN2 video generation. #10301, #10120, #8602
- Enable CUDA V2 K-step decoding for MiniCPM-o 4.5 and move MOSS-TTS-Nano and MOSS-TTS-Realtime to V2. #10288, #9905, #9906
Compatibility
Requires Python 3.11–3.13 and matching vLLM 0.34.0. The alignment updates dependencies and CUDA/ROCm/XPU definitions, and retains DDUF loading with Diffusers 0.42. #10365
Deployments that select retired Omni V1 runners are rejected. Affected workflows include:
- Drama 3 / Fish-Omni: complete speech serving, including cloning, streaming, and concurrency.
- LFM3-Omni: speech and text-only profiles. LFM3-Audio is a separate profile.
- PersonaPlex: realtime duplex, including the MRv2 profile whose Talker requires V1.
- YuE2: audio synthesis and export paths requiring V1 sampling hooks.
- Qwen3-TTS: ROCm and MUSA overrides that select V1.
- Qwen3-Omni: ROCm and MUSA overrides that select V1.
V2 port tracking #10584. Kokoro-82M and JoyAI-Image-Edit were also retired. #10393
Native vLLM 0.34 Ascend execution remains unqualified. The retained 0.33 Ascend images stopped at a vLLM API import error before device tests. NPU #11982
Release stabilization
CUDA Ready 50/50 and Merge 25/25 passed on the alignment PR head. The Qwen3-Omni Merge shard passed its automatic retry after its original timeout. The alignment also retains the CosyVoice3 test matrix on the native pytest pin and adds the No Async Chunk baseline gate. Alignment #10365 · CUDA #3336
The release tag points to the alignment merge commit, which passed its upstream CUDA check #20124.
Release artifacts
Source commit: 709dc427ef59b9a1a1e0b092efd0236d3f70f723.
- Wheel:
vllm_omni-0.34.0rc1-py3-none-any.whlbuilt by the official release pipeline. SHA-256:436f7e8e4ceea61895fee7f01a9d50612813fc76b9640fe28df2c3386ddf3bf5. - PyPI publication is pending. The wheel above is available from the official vLLM wheel index.
- CUDA amd64, CUDA arm64, and XPU amd64 images are building in omni-release #4665.
Install with the matching CUDA 13.0 dependencies:
uv venv --python 3.12
source .venv/bin/activate
uv pip install "vllm==0.34.0" \
"https://wheels.vllm.ai/omni/709dc427ef59b9a1a1e0b092efd0236d3f70f723/vllm_omni-0.34.0rc1-py3-none-any.whl" --torch-backend=cu130Full changelog: v0.33.0rc1...v0.34.0rc1