Comparison
LoRAvsQLoRA
LoRA
instead of editing the whole model, train a small patch that sits on top of it.
Low-rank adaptation: freeze the base weights and train small low-rank matrices that add adjustments on top. Typically well under one percent of parameters are trained, and the output is a small adapter file rather than a new model. The default fine-tuning method almost everywhere.
Full entry →QLoRA
you fine-tuned a mid-sized model on one consumer GPU, because the base underneath the adapter was quantised to four bits.
LoRA applied over a 4-bit quantised base model, shrinking memory enough to fine-tune mid-sized models on a single consumer GPU. The technique that democratised fine-tuning.
Full entry →