| Documentation | Recipe | Slides | DeepWiki | User Forum | Developer Slack | WeChat |
afd-plugin is a vLLM external plugin for Attention-FFN Disaggregation (AFD). It provides plugin-owned worker classes, model runners, model wrappers, connectors, configuration validation, compatibility shims, and hardware-gated integration tests for GPU and Ascend NPU deployments.
Note
This project is still experimental and needs more large-scale testing across different hardware backends.
The plugin extends vLLM through its general-plugin entry point without modifying the vLLM source tree.
- Install AFD for CUDA or Ascend NPU.
- Choose a connector and supported model, then follow its deployment recipe.
- Pass role/topology settings through
--additional-config. Send requests to Attention; FFN serves connector work.
See the support matrix for models, connectors, tested configurations, and known accuracy/shutdown limits.
Browse the documentation index for user guides, module design, validation and troubleshooting.
For CPU-only or macOS development, follow source setup, then run:
uv run pytest
uv run ruff check .Read the contributor rules and test maintenance guide before changing code or tests. See module design for ownership, native builds for backend operators, and the E2E guide for hardware validation.
afd-plugin is licensed under the Apache License 2.0.
If you find afd-plugin helpful in your research or projects, please consider citing it:
@misc{afdplugin2026,
title={afd-plugin: Attention-FFN Disaggregation for vLLM},
author={AFD Plugin Contributors},
year={2026},
howpublished={\url{https://github.com/vllm-project/afd-plugin}},
}