QLoRA

Also known as: Quantized LoRA

QLoRA fine-tunes LoRA adapters on a base model stored in low-bit quantization (commonly 4-bit, e.g. NF4). The frozen base stays memory-cheap; adapters and optimizer states use higher precision so gradients remain usable.

Why it exists

Full fine-tunes of 7B–70B models need a lot of VRAM. QLoRA makes PEFT feasible on a single consumer or mid-tier GPU by stacking quantization + LoRA (plus paging tricks in the original recipe).

Knobs

Quantization type/bits, LoRA rank and target modules, and data quality dominate outcomes. Expect some quality gap vs full-precision LoRA; measure on your eval set rather than assuming “4-bit is free.”

← Back to AI Glossary