Custom patches and utilities for MoERL — an efficient RL fine-tuning framework for Mixture of Experts (MoE) LLMs.

---

✨ What is moerl-zoo?

moerl-zoo is a toolkit inspired by unsloth-zoo, designed to serve the needs of MoERL. It includes lightweight patches, wrappers, and utilities specifically adapted for MoE reinforcement learning fine-tuning pipelines.

🦥 This project is based on unsloth-zoo, created by Daniel Han-Chen and the Unsloth team. All modifications are made under the terms of the LGPL-3.0-or-later license.

🚀 Features

• 🧩 Custom monkey patches and related utilities designed for MoERL compatibility
• 🚀 Supports efficient MoE-based RL fine-tuning
• ⚡ Fully compatible with vllm, bitsandbytes, transformers, trl, and more

📦 Installation

pip install "moerl_zoo @ git+https://github.com/slowfastai/moerl-zoo.git"

🤝 Acknowledgements

Huge thanks to the Unsloth project for their foundational work in efficient LLM optimization ❤️. MoERL builds upon their vision while tailoring functionality toward MoE RL.

📜 License

This project is licensed under the GNU Lesser General Public License v3.0 or later (LGPL-3.0-or-later).

The original base project, 🦥 unsloth-zoo, is licensed under the GNU Lesser General Public License v3.0 or later (LGPL-3.0-or-later).

All original LICENSE and NOTICE files have been preserved where applicable.