# OXR vLLM plugin

Registers the OXR architecture through `vllm.general_plugins`. Requires
Python 3.12+, vLLM 0.25.1 and Transformers 5.13.1. It does not modify the
vLLM installation. The supported Docker runtime uses PyTorch 2.11.0+cu130.

Install from the model repository root in a Linux CUDA environment:

```bash
python -m pip install ./plugins/vllm
```

Start with the model files and weights in the repository root:

```bash
VLLM_PLUGINS=oxr vllm serve . \
  --trust-remote-code \
  --served-model-name oxr \
  --chat-template ./chat_template.jinja \
  --dtype bfloat16 \
  --attention-config '{"backend":"TRITON_ATTN"}' \
  --max-model-len 16384
```

The LA model in `layout_analysis/` is served separately by the OXR layout
analysis service, not by this plugin.

To roll back to vLLM 0.19.1, use the previous model plugin together with the
matching previous OXR launcher and Docker image. Do not install this plugin
into the old runtime.