vllm-project/vllm-omni: Omni-Modality Inference Framework Ships v0.16.0 with Diffusion Transformer Support — 3.8K Stars

GitHub·low signal

The vLLM community's vllm-omni framework extends high-throughput LLM inference to omni-modality models processing text, image, video, and audio simultaneously, with v0.16.0 (Feb 28) adding Qwen3-Omni support, audio/TTS generation, diffusion stacks, and multi-platform coverage (CUDA/ROCm/NPU/XPU). The framework extends autoregressive support to Diffusion Transformers (DiT) and other parallel generation models with heterogeneous pipeline abstraction and OpenAI-compatible APIs. A companion vllm-omni-skills repo provides developer-friendly Claude Code and Cursor IDE integrations.