Operate disaggregated multimodal vLLM-Omni deployments across image, video, and audio endpoints.
npx skills add https://github.com/air-gapped/skills --skill vllm-omni
Please help me install this Agent Skill. Skill: vllm-omni Source: https://github.com/air-gapped/skills/tree/main/.claude/skills/vllm-omni Command: npx skills add https://github.com/air-gapped/skills --skill vllm-omni
vLLM-Omni enables operators to deploy and operate disaggregated, multimodal inference pipelines that support image, video, and audio modalities at scale.
Start a local vLLM-Omni deployment using a prepared stage-config YAML and point clients at the Omni-enabled endpoint.
High-intent search queries and answers about installing and using this skill.
Deploy multimodal serving by using vLLM-Omni stage-config YAML files to disaggregate pipelines across endpoints like /v1/images/generations, /v1/videos, and /v1/audio/speech for scalable inference.
Stage-based disaggregation splits multimodal inference into sequential phases like Thinker, Talker, and Code2Wav, connected by OmniConnector to optimize processing across distinct model architectures.
Yes, vLLM-Omni supports serving multiple architectures including Qwen3-Omni, FLUX, and Wan2.2, allowing operators to configure distinct diffusion and CFG pipelines across the same disaggregated deployment.
Configure diffusion schedulers and LoRA by modifying the stage-config YAML files, which expose operator settings for quantization, frame interpolation, and diffusion pipelines across the supported endpoints.
Yes, vLLM-Omni enforces the official Skill Unit definition by requiring a root SKILL.md with a name and description to validate the deployment and expose metadata for dependencies and components.
Disaggregated multimodal serving requires careful configuration of stage-based pipelines and OmniConnector, meaning operators must manage complex YAML configs to align quantization, diffusion schedulers, and model architectures correctly.
cuiossCentralize project infrastructure configuration in marshal.json with domain defaults and CI settings.
hsliuustc0106Bridge client applications with the vLLM-Omni API using OpenAI-compatible endpoints.
VYRE-StudiosGenerate images from text prompts using HuggingFace Diffusers and Stable Diffusion models.
spacecat2002Integrate diffusion models and Diffusers pipelines into the SGLang runtime.
