Hugging Face has unveiled the 🤗 PEFT (Parameter-Efficient Fine-Tuning) library, designed to enable fine-tuning of billion-scale models on low-resource hardware. The library integrates with 🤗 Transformers and 🤗 Accelerate, supporting techniques like LoRA, Prefix Tuning, Prompt Tuning, and P-Tuning.
PEFT addresses the high computational and storage costs of full fine-tuning by only updating a small number of extra parameters while freezing most of the pretrained model. This approach reduces memory usage and storage requirements—fine-tuned checkpoints are just megabytes instead of gigabytes—while maintaining performance comparable to full fine-tuning. It also mitigates catastrophic forgetting and improves generalization in low-data regimes.
Key use cases include tuning a 3-billion-parameter model on a consumer GPU with 11GB RAM, INT8 tuning of a 6.7-billion-parameter model in Google Colab, and Stable Diffusion Dreambooth training on limited hardware. The library is open-source and available on GitHub.