# vllm-project/vllm-omni

A framework for efficient model inference with omni-modality models

Repository: https://github.com/vllm-project/vllm-omni
Canonical: https://ross.abutalabs.com/products/vllm-omni
Homepage: https://docs.vllm.ai/projects/vllm-omni
Language: Python
License: Apache-2.0
License Family: permissive
Topics: diffusion, inference, model-serving, pytorch, transformer, audio-generation, image-generation, multimodal, video-generation, world-model
Last push: 2026-08-26T20:19:21+00:00

## Health v2 (maintenance only)
Score: 83/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 96, longevity 25
- inputs: {"age_days": 357, "days_push": 7, "days_rel": 30, "gap_med": 28, "n_releases_24m": 8}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 6369, forks 1554 (observed 2026-08-28T04:09:42.825456+00:00)

## What it is
vLLM-Omni is a Python framework extending vLLM for efficient inference and serving of omni-modality models, including diffusion transformers, TTS, image/video generation, and robot-policy models. It provides pipelined stage execution, distributed parallelism, streaming outputs, and an OpenAI-compatible API server.

## Use cases
- serve a text-to-image diffusion model behind an OpenAI-compatible API
- run offline batched inference with Qwen3-Omni or MiniCPM-o
- deploy TTS models like CosyVoice3 with streaming audio output
- serve video generation models like Wan2.2 or MiniMax H3
- host robot-policy and action models for robotics inference
- run full-duplex realtime voice serving with streaming audio input and output

## When to choose
- you need high-throughput serving of multimodal, diffusion, TTS, or action models on GPU
- you want vLLM-style performance (KV cache, batching, parallelism) for non-autoregressive models
- you need an OpenAI-compatible server with streaming multimodal outputs

## When to avoid
- you only need plain text LLM inference, where core vLLM suffices
- you need Windows or macOS support, since it targets Linux
- you want a lightweight single-model pipeline without distributed serving complexity

## Facets
- artifact type: framework
- maturity: active
- function: llm-inference, machine-learning, image-processing, audio-processing, video-processing, tts, speech-recognition, http-server, api-framework
- domain: large-language-models, machine-learning, artificial-intelligence, image-processing, speech-processing, robotics, gpu-computing
- platform: python, cloud
- tags: diffusion, multimodal, model-serving, openai-compatible-api, dit, world-model, robot-policy, vllm, audio, video, linux, gpu, docker

## Member repositories
- vllm-project/vllm-omni (main) score 83

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:09:42.825456+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:45:12.878419+00:00, confidence not recorded.
  - readme: https://github.com/vllm-project/vllm-omni (fetched 2026-08-28T04:09:42.825456+00:00, sha 21fd672b0165)
  - homepage: https://docs.vllm.ai/projects/vllm-omni (fetched 2026-08-29T08:42:13.857704+00:00, sha 50487ae1953f)
  - site_page: https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/quickstart (fetched 2026-08-29T08:42:13.866648+00:00, sha 46d1086ad515)
  - registry_pypi: https://pypi.org/pypi/vllm-omni/json (fetched 2026-08-29T08:42:13.868730+00:00, sha 8e7ecb34ac22)
- Data as of 2026-08-30T08:39:29.467469+00:00.