LoRA is an economical fine-tuning technique: the base model's weights are frozen and small low-rank matrices added to the layers are trained, representing the modification. It sharply reduces memory and time, and these adapters are light to store. PEFT refers to the family of these economical methods.
Example
LoraConfig(r=8, lora_alpha=16, target_modules=['q_proj', 'v_proj'])