DipSVD: Dual-importance Protected SVD for Efficient LLM Compression

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Ding, Xuan, Sun, Rui, Zhang, Yunjian, Yan, Xiu, Zhou, Yueqi, Huang, Kaihao, Fu, Suzhong, Xie, Chuanlong, Zhu, Yao
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909659926888448
author Ding, Xuan
Sun, Rui
Zhang, Yunjian
Yan, Xiu
Zhou, Yueqi
Huang, Kaihao
Fu, Suzhong
Xie, Chuanlong
Zhu, Yao
author_facet Ding, Xuan
Sun, Rui
Zhang, Yunjian
Yan, Xiu
Zhou, Yueqi
Huang, Kaihao
Fu, Suzhong
Xie, Chuanlong
Zhu, Yao
contents The ever-increasing computational demands and deployment costs of large language models (LLMs) have spurred numerous compressing methods. Compared to quantization and unstructured pruning, SVD compression offers superior hardware compatibility and theoretical guarantees. However, existing SVD-based methods focus on the overall discrepancy between the original and compressed matrices while overlooking the protection of critical components within the matrix, which leads to inferior performance in the compressed models. This paper proposes a dual-level importance protection mechanism to enhance SVD-based compression methods: (1) local importance protection: preserving the most critical singular vectors within each weight matrix through channel-weighted data whitening; and (2) global importance protection: enabling less important layers to bear a greater portion of the compression burden through either a heuristic or optimization-based approach, thereby minimizing the impact of compression on critical layers. Extensive experiments demonstrate that DipSVD outperforms existing SVD-based compression approaches across multiple benchmarks, achieving superior model performance especially at high model compression ratios.
format Preprint
id arxiv_https___arxiv_org_abs_2506_20353
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DipSVD: Dual-importance Protected SVD for Efficient LLM Compression
Ding, Xuan
Sun, Rui
Zhang, Yunjian
Yan, Xiu
Zhou, Yueqi
Huang, Kaihao
Fu, Suzhong
Xie, Chuanlong
Zhu, Yao
Machine Learning
Artificial Intelligence
The ever-increasing computational demands and deployment costs of large language models (LLMs) have spurred numerous compressing methods. Compared to quantization and unstructured pruning, SVD compression offers superior hardware compatibility and theoretical guarantees. However, existing SVD-based methods focus on the overall discrepancy between the original and compressed matrices while overlooking the protection of critical components within the matrix, which leads to inferior performance in the compressed models. This paper proposes a dual-level importance protection mechanism to enhance SVD-based compression methods: (1) local importance protection: preserving the most critical singular vectors within each weight matrix through channel-weighted data whitening; and (2) global importance protection: enabling less important layers to bear a greater portion of the compression burden through either a heuristic or optimization-based approach, thereby minimizing the impact of compression on critical layers. Extensive experiments demonstrate that DipSVD outperforms existing SVD-based compression approaches across multiple benchmarks, achieving superior model performance especially at high model compression ratios.
title DipSVD: Dual-importance Protected SVD for Efficient LLM Compression
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2506.20353