QLoRA 4-bit finetuning of LLMs (github.com)

An efficient finetuning approach that reduces memory usage enough to finetune a 65B parameter model on a single 48GB GPU while preserving full 16-bit finetuning task performance!