Unsloth
Dramatically speeds up LoRA and QLoRA fine-tuning by rewriting GPU kernels. Compatible with HuggingFace and works with Llama, Mistral, Gemma, and more. No accuracy loss.
Torchtune
Meta's official fine-tuning library. Pure PyTorch — no abstraction layers. Supports LoRA, QLoRA, and full fine-tuning for Llama models. Designed for reproducibility and research.