peft
🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.
What is peft?
🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.
Overview of peft
What it is
PEFT (Parameter-Efficient Fine-Tuning) is a library that provides state-of-the-art methods for efficiently adapting large pretrained models to downstream tasks. Instead of fine-tuning all model parameters, PEFT methods only update a small number of extra parameters, significantly reducing computational and storage costs while achieving performance comparable to full fine-tuning.
Key features
- Supports multiple PEFT methods including LoRA, as demonstrated in the Quickstart.
- Integrates with Transformers for easy model training and inference, allowing addition, loading, and switching of adapters.
- Integrates with Diffusers for managing different adapters in diffusion models, reducing memory requirements.
- Integrates with Accelerate for distributed training and inference of large models.
- Enables fine-tuning of large models on consumer hardware; for example, LoRA reduces GPU memory for a 3B parameter model from 47GB to 14.4GB.
- Can be combined with quantization (e.g., QLoRA) to further reduce memory requirements.
- Produces small adapter checkpoints (e.g., 19MB for a 3B model) compared to full model checkpoints (e.g., 11GB).
Use cases
- Fine-tuning large language models (e.g., bigscience/T0_3B, bigscience/mt0-xxl) on downstream tasks with limited GPU memory.
- Training Stable Diffusion models with LoRA, reducing GPU memory from 27.5GB to 15.5GB (or 8.12GB with gradient checkpointing).
- Saving storage by maintaining multiple small adapters for different tasks instead of full model copies.
- Performing inference with PEFT adapters loaded on top of a base model.
- Combining PEFT with quantization to fine-tune models on consumer hardware (e.g., Llama-2-7b on a 16GB GPU).
Getting started
Install PEFT from pip:
pip install peft
For full usage examples, including preparing a model with get_peft_model and loading adapters for inference, see the repository README.
License and pricing
The license is Apache-2.0. The project is open source.
Based on the project's public README and repository metadata.