torchtune is a PyTorch native post-training library for authoring, fine-tuning, and experimenting with large language models.
The library addresses the challenge of efficiently post-training LLMs by providing hackable training recipes and simple PyTorch implementations of popular models. It supports the full post-training lifecycle including supervised fine-tuning, knowledge distillation, DPO, PPO, GRPO, and quantization-aware training. The tool emphasizes memory efficiency and performance through best-in-class implementations that leverage the latest PyTorch APIs, with YAML-based configuration for easy recipe customization across single device, multi-device, and multi-node setups.
Developers should be aware that the project is no longer actively maintained as of 2025, though it previously supported a wide range of models including Llama, Gemma, Mistral, Phi, and Qwen variants. The library is most suitable for teams already invested in the PyTorch ecosystem who need flexible, production-grade fine-tuning infrastructure. Those seeking an actively developed alternative should evaluate other options, as the project has wound down development.
The project attracted a substantial base of adopters who reported real-world use cases through the issue tracker. Maintainers typically responded to new issues and pull requests within a day. Work in the issue tracker centered on community help requests, enhancements, and bug fixes, with most open issues raised by outside users rather than the core team.
