unsloth

Unsloth is the open-source library that delivers up to 30x faster fine-tuning for Large Language Models (LLMs) with up to 74% less VRAM.

Unsloth is an optimized framework for LLM fine-tuning, leveraging custom GPU kernels and manual differentiation to drastically cut training time and memory footprint. We've seen performance gains of up to 30x faster than Flash Attention 2 (FA2) and a memory reduction of up to 74%—for instance, reducing a 23-hour fine-tuning job on a Tesla T4 GPU to just over 2 hours. The platform supports key models like Llama, Gemma, and Qwen, maintains 0% loss in accuracy, and ensures broad hardware compatibility across NVIDIA, AMD, and Intel GPUs, integrating cleanly with the Hugging Face ecosystem.

https://unsloth.ai