As LLMs get larger fine-tuning to the full extent can become difficult to train on consumer hardware. Storing and deploying these tuned models can also be quite expensive and difficult to store. With PEFT (parameter -efficent fine tuning), it approaches fine-tune on a smaller scale of model parameters while freezing most parameters of the pretrained LLMs. Basically it is providing full performance that which is similar if not better than full fine tuning while only having a small number of trainable parameters. This source explains that as well as going over LORA diagrams and a code walk through.