Metadata-Version: 2.4
Name: exionos-vllm
Version: 0.1.0
Summary: vLLM architecture plugin for Exionos X1
Author: Exionos
License-Expression: Apache-2.0
Project-URL: Homepage, https://huggingface.co/Exionos/Exionos-X1-27B
Keywords: vllm,exionos,llm,inference
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: Programming Language :: Python :: 3
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
Dynamic: license-file

# exionos-vllm

vLLM architecture plugin for [Exionos X1](https://huggingface.co/Exionos/Exionos-X1-27B).

Exionos X1 ships its modeling code as remote code for Transformers. vLLM dispatches
on `config.architectures[0]`, so it needs `ExionosX1ForConditionalGeneration`
registered before it can serve the model. Installing this package is enough — it
registers itself through vLLM's `vllm.general_plugins` entry point at startup.

## Install

```bash
pip install exionos-vllm
```

## Use

No extra flags beyond `--trust-remote-code`:

```bash
vllm serve Exionos/Exionos-X1-27B \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --kv-cache-dtype fp8
```

## Compatibility

Verified against vLLM 0.27.1 and Transformers 5.15. The plugin registers the Exionos
architecture and config types with vLLM; the config registration touches a private
vLLM API and degrades gracefully (it is skipped, not fatal) if that API moves.

## License

Apache-2.0
