huggingface/peft

PEFT, short for Parameter-Efficient Fine-Tuning, is a Python library from Hugging Face for adapting very large pretrained AI models to new tasks without retraining the whole thing. Fine-tuning a big model the normal way means updating every one of its billions of parameters, which is slow and demands a lot of memory and disk. PEFT freezes the original model and only trains a small extra set of parameters layered on top, so the computational and storage costs drop dramatically while, according to the README, the quality stays comparable to fully fine-tuned models. The library packages several specific techniques, including LoRA, soft prompts, and IA3, and the README also describes combining it with quantization (representing weights in lower precision) through approaches like QLoRA to fit training onto smaller GPUs. To use it you install with pip install peft, wrap a base model and a configuration object such as LoraConfig with get_peft_model, and then train it like any other model. In the README's example only about 0.19% of the parameters end up being trained, and the saved adapter can be a few megabytes instead of gigabytes. People reach for PEFT when they want to customise a large language model or image-generation model on their own data but don't have the GPU budget for full fine-tuning, or when they want many small task-specific adapters they can swap in and out without storing many full copies. It is integrated with Hugging Face's Transformers library for training and inference, with Diffusers for managing adapters on diffusion image models, and with Accelerate for distributing very large training jobs. The full README is longer than what was provided.