MLX-GRPO — Fine-Tune LLMs with GRPO on Apple Silicon

★ 245

MLX-GRPO is a training pipeline written purely in MLX, Apple's machine learning framework, for fine-tuning large language models with GRPO on Apple Silicon. GRPO, Group Relative Policy Optimization, is a reinforcement learning method that compares a model's answers within a group. People reach for it when they want to fine-tune a small language model right on a Mac without cloud GPUs, learn how reinforcement learning for LLMs works from compact, readable code, or test a GRPO idea before moving it to a large cluster. The pipeline is written in Python and uses only MLX with the Metal backend, with no PyTorch or CUDA; according to the project, the process covers dataset preparation and the training loop itself. Unlike general-purpose fine-tuning libraries, MLX-GRPO focuses on one method and one platform. The repository has about 245 stars. It is only for owners of Apple Silicon Macs; model size is limited by the machine's memory, and supported models are worth checking in the repository.