Torchtune: PyTorch Native Post-Training Library for LLMs

Torchtune is a library built specifically for PyTorch that aims to enhance the post-training processes of large language models (LLMs), facilitating effective fine-tuning, experimentation, and deployment workflows. In contrast to many current fine-tuning solutions that focus on user-friendliness, tailored recipes, or hardware optimization at the expense of transparency and flexibility, torchtune prioritizes modularity, adaptability, and direct interaction with core PyTorch elements. The paper outlines the foundational design principles of torchtune, illustrates their implementation in its model builders, training recipes, and distributed training framework, and assesses the library's performance in various post-training scenarios, benchmarking it against well-known fine-tuning frameworks.