QLoRA

Quantized LoRA — fine-tune 65B models on a single 48GB GPU while preserving full 16-bit performance.