LoRA
Also: low-rank adaptation, qlora
Low-Rank Adaptation: a way to fine-tune a model by training a small set of extra weights instead of the whole model, cutting the memory and cost dramatically.
LoRA, short for Low-Rank Adaptation, makes fine-tuning practical on ordinary hardware. Instead of updating all of a model’s weights, it freezes them and trains a small set of additional weights alongside. Those extra weights, the LoRA adapter, are tiny, often a few megabytes, and can be shared and swapped independently of the base model.
Because you are training so few parameters, the memory and compute needed drop enormously compared with a full fine-tune, while the results are often close.
QLoRA is LoRA applied on top of a quantised base model, which lowers the memory further and lets you fine-tune surprisingly large models on a single consumer GPU. Between them, LoRA and QLoRA are why fine-tuning is no longer the preserve of well-funded labs.