Hugging Face releases PEFT library

As the size of large language models (LLMs) has grown rapidly, Full Fine-tuning, which updates all parameters, has faced limitations in requiring massive computing resources and storage space.

The 🤗 PEFT(Parameter-Efficient Fine-Tuning) library solves this problem by freezing most parameters and training only a small number of additional parameters. This can dramatically reduce computation and storage costs, and also prevents Catastrophic Forgetting, a phenomenon where the model loses existing knowledge.

The major techniques currently supported are as follows:

• LoRA (Low-Rank Adaptation)
• Prefix Tuning
• Prompt Tuning
• P-Tuning

Using this library, it is possible to fine-tune models with billions of parameters even in consumer GPU environments such as RTX 2080 Ti or 3080. It can also be applied to various modalities, such as image generation models, like Dreambooth training for Stable Diffusion.