Mixture-of-Subspaces in Low-Rank Adaptation

Basic Information

Title: Mixture-of-Subspaces in Low-Rank Adaptation

Authors: Taiqiang Wu; Jiahao Wang; Zhe Zhao; Ngai Wong

Link: https://arxiv.org/pdf/2406.11909

TL, DR: We decompose LoRA into subspaces via structural re-parameterization and propose a simple yet effective MoSLoRA method, employing a learnable mixer to fuse more subspaces and more flexibly.

Quick Start

please follow the README in each folder to reproduce the results.

• TODO: try to insert MoSLoRA into peft package.

Contact

Taiqiang Wu: takiwu@connect.hku.hk

Citation

If you find MoSLoRA useful, please cite it by using the following BibTeX entry.

@article{wu2024moslora,
  title={Mixture-of-Subspaces in Low-Rank Adaptation},
  author={Taiqiang Wu and Jiahao Wang and Zhe Zhao and Ngai Wong},
  booktitle={arXiv:2406.11909},
  url={arxiv.org/abs/2406.11909},
  year={2024}
}