PEFT

Hugging Face's PEFT library lets you adapt a large pretrained network by training a small set of added or selected weights while the original parameters stay frozen. Wrapping a Transformers model with get_peft_model and a config object injects the adapter, and the saved result is often just a few megabytes that can sit on top of the base model, be merged into it, or be swapped at runtime. Beyond LoRA and its variants such as DoRA, AdaLoRA, LoHa, LoKr and X-LoRA, the repository implements orthogonal methods like OFT and BOFT, IA3, VeRA, prefix and prompt tuning, P-tuning and dozens of newer research techniques. It plugs into Diffusers for image-model LoRAs, Accelerate and DeepSpeed for multi-GPU runs, TRL for supervised and preference fine-tuning, and bitsandbytes for QLoRA-style 4-bit training, which puts 7B-class models within reach of a single 16 GB card. Setup is pip install peft next to PyTorch. Apache-2.0, so commercial use carries no strings. Anyone fine-tuning LLMs, Whisper or Stable Diffusion on limited hardware will find it the usual starting point, and saved adapters load straight into Transformers and Diffusers.