QLoRA in AI answers

QLoRA is an open-source finetuning method that enables efficient training of large language models by keeping the base model quantized to 4-bit while injecting Low Rank Adapters for training. It employs memory-saving techniques such as 4-bit NormalFloat (NF4), double quantization, and paged optimizers to allow finetuning of very large models (e.g., 65B parameters) on a single 48GB GPU without sacrificing performance. The project integrates with Hugging Face's PEFT and transformers, uses bitsandbytes, and releases models, code, CUDA kernels, and demos to democratize LLM research.