vLLM-Omni
#2 en Inférence & serving LLMVllm · 3× · vu le 03 sept. 2026
51
Momentum
vLLM-Omni is an Open Source extension of the vLLM serving framework that supports omni-modal and non-autoregressive models (image, audio, video, robotics actions) alongside classical text-LLM Inference. It extends the efficiency techniques known from vLLM (e.g., KV-Cache management) with pipeline-like multi-stage execution for Diffusion Transformers and TTS architectures. The software is available as an open-source project under the Apache-2.0 license and is primarily self-hosted, though it is also used in production by third-party providers such as Baseten. Initially released in November 2025, it has since followed a continuous release cycle in parallel with vLLM major versions.
Historique du momentum
05.06.03.09.
Fonctionnalités
| Deployment (Self-Hosted/Cloud) | Primarily self-hosted on own GPU infrastructure; managed deployment also available via third parties such as Baseten |
| Throughput/Latency | TTS benchmarks measure RTF (realtime factor), TTFP (time to first packet), and Tput (generated audio seconds per wall-clock second); values are model-dependent |
| License | Apache License 2.0 |
| Platform | Supports CUDA (NVIDIA), ROCm (AMD), MUSA, NPU, and XPU as hardware backends |
| Price | Free, open source (Apache 2.0); costs only arise from self-managed GPU infrastructure or third-party managed hosting (e.g. Baseten, RunPod, Modal) |
| Protocol Compatibility | OpenAI-compatible API, including for text-to-speech (TTS) endpoints |
| Release Date | First official release on November 30, 2025 (v0.11.0rc, built on vLLM v0.11.0) |
| Supported Models/Providers | Omni models (Qwen3-Omni, MiniCPM-o 4.5, Cosmos3, HunyuanImage, BAGEL), TTS (Qwen3-TTS, IndexTTS 2.5, CosyVoice3), diffusion (MiniMax H3, Wan2.2, LTX), robotics (π0, GR00T-N1.7) |