Quickstart¶
The shortest path is to start from a working vllm-metax environment and then
layer vllm-omni and vllm-omni-metax into it.
Recommended deployment path¶
For most users, the recommended bring-up path is:
- Start from a MetaX image or host where
vllm-metaxalready works. - Install
vllm-omni. - Install
vllm-omni-metax. - Launch your normal
vllm-omniworkflow.
This mirrors the guidance used by vLLM-metax: get the hardware backend stable
first, then add higher-level functionality.
Releases¶
Below is version mapping to released plugin and mcoplib with maca:
| plugin version | maca version | mcoplib version | docker image url |
|---|---|---|---|
v0.18.0 |
maca3.5.3.x |
0.4.3 |
vllm-metax:0.18.0 |
v0.19.0 |
maca3.5.3.x |
0.4.4 |
vllm-metax:0.19.0 |
v0.20.0 |
maca3.5.3.x |
0.4.5 |
-- |
Usage Warning
vLLM-Omni-MetaX is intended to work out of the box with the matching Docker images listed above.
All VLM tests are based on the related maca version. Using an incompatible version of maca for vllm-omni-metax may cause unexpected bugs or errors. This is not guaranteed.
What changes after installation¶
You do not need a separate user-facing launcher from this repository. Once the plugin is installed:
vllm-omnikeeps its normal CLI and workflow.- The adapter activates only when MetaX runtime is detected, unless forced.
- A runtime compatibility patch may be applied during activation for the Omni rotary path.
- Omni AR and generation workers continue to use GPU worker classes.
- Diffusion attention backend selection follows a conservative policy:
FLASH_ATTNwhen both capability and package checks pass, otherwiseTORCH_SDPA.
Next steps¶
- Follow Installation for the full environment setup.
- Read User Guide for activation and behavior details.
- Read Architecture if you plan to extend the plugin.