vLLM-Omni MetaX

vllm-omni-metax is the MetaX adapter plugin for vllm-omni. It reuses the existing vllm-metax hardware backend and adds the glue required for Omni's multi-stage multimodal runtime.

In practice, the responsibilities are split as follows:

Component	Responsibility
vllm-metax	MetaX hardware platform, kernels, runtime integration
vllm-omni	Multi-stage multimodal inference pipeline
vllm-omni-metax	Omni platform plugin that bridges the two

Why this project exists

vllm-omni ships with GPU-oriented execution paths, but MetaX enablement should stay owned by the MetaX backend instead of being maintained as an in-tree fork. This repository keeps that boundary clear:

• It does not reimplement a new hardware backend.
• It detects whether a usable MetaX runtime is present.
• It registers an out-of-tree Omni platform plugin only when activation is appropriate.
• It applies a focused runtime patch only where upstream CUDA-only assumptions block MetaX execution in vllm-omni 0.20.0.
• It mirrors Omni's stage device selection to the MetaX runtime environment.

Recommended workflow

If you are deploying on MetaX hardware, the recommended order is:

1. Prepare a working vllm-metax environment.
2. Install vllm-omni.
3. Install vllm-omni-metax.
4. Start the usual vllm-omni workflow and let this plugin activate automatically.

See:

• Installation
• Quickstart
• User Guide
• Architecture

Design goals

• Reuse the existing MetaX platform implementation from vllm-metax.
• Keep compatibility fixes small, explicit, and runtime-scoped instead of carrying a source fork of vllm-omni.
• Keep operational behavior predictable for users already familiar with vllm-metax.
• Make debugging explicit with environment-controlled enable and disable switches.