All terms
Optimization
QLoRA
Fine-tuning a 4-bit quantized base model with LoRA adapters on a single GPU.
Definition
QLoRA combines LoRA with 4-bit quantization of the frozen base model so large models can be fine-tuned on a single modest GPU. The base weights are stored in a 4-bit format and kept frozen, while small LoRA adapter matrices are trained in full precision, with gradients flowing only into the adapters. This brings the cost of adapting big open-weight models far down, with minimal accuracy loss versus full fine-tuning.