PyTorch code for weight-decomposed low-rank adaptation (DoRA)
Top 44.3% on sourcepulse
DoRA (Weight-Decomposed Low-Rank Adaptation) is a PyTorch implementation for efficient fine-tuning of large language and vision-language models. It targets researchers and practitioners seeking to improve LoRA's performance and stability without increasing inference costs, offering enhanced learning capacity and training stability.
How It Works
DoRA decomposes pre-trained weights into magnitude and direction components. It then applies LoRA specifically to the directional component. This approach aims to improve upon standard LoRA by decoupling the magnitude and direction of weight updates, leading to better fine-tuning results and stability, especially at lower ranks.
Quick Start & Requirements
pip install git+https://github.com/huggingface/peft.git -q
(for HuggingFace PEFT integration).bitsandbytes
for quantization.10 months ago
1 week