PyTorch Releases torchtune for Easily Fine-Tuning LLMs

PyTorch has announced the alpha release of torchtune, a new library designed to simplify the fine-tuning of large language models (LLMs) using PyTorch. 

Click here to check out the GitHub repository.

The library is built on PyTorch’s core principles, offering modular building blocks and customisable training recipes for fine-tuning popular LLMs on various GPUs, including consumer-grade and professional ones.

The library provides a comprehensive fine-tuning workflow, from downloading and preparing datasets and model checkpoints to customising training with composable building blocks, logging progress, quantizing models post-tuning, evaluating fine-tuned models, running local inference for testing, and ensuring compatibility with popular production inference systems.

Torchtune aims to address the increasing demand for fine-tuning LLMs by offering flexibility and control. Users can easily add customisations and optimizations to adapt models to specific use cases, including memory-efficient recipes that work on machines with single 24GB gaming GPUs.

The design of torchtune focuses on easy extensibility, democratising fine-tuning for users of varying expertise, and interoperability with the open-source LLM ecosystem. 

The library integrates with popular tools such as Hugging Face Hub, PyTorch FSDP, Weights & Biases, EleutherAI’s LM Evaluation Harness, ExecuTorch, and torchao for various purposes like model and dataset access, distributed training, logging, evaluation, inference, and quantisation.

torchtune currently supports Llama 2, Mistral, and Gemma 7B models and plans to expand with additional models, features, and fine-tuning techniques in the coming weeks including 70 billion parameters and Mixture of Experts models.