Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition

ICLR 2025 Conference Submission88 Authors

Keywords: Mixture of Experts, Efficient Large Language Models, Low-Rank Decomposition, Network Sparsity

TL;DR: In this paper, we present a novel approach to compressing Mixture of Experts (MoE) models using Singular Value Decomposition (SVD).

Abstract: Mixture of Experts (MoE) architecture has emerged as a powerful paradigm in the development of Large Language Models (LLMs), offering superior scaling capabilities and reduced computational costs. However, the increased parameter budgets and memory overhead associated with MoE LLMs pose significant challenges to their efficiency and widespread deployment. In this paper, we present MoE-SVD, the first decomposition-based compression framework tailored for MoE LLMs without any extra training. By harnessing the power of Singular Value Decomposition (SVD), MoE-SVD addresses the critical issues of decomposition collapse and matrix redundancy in MoE architectures. Specifically, we first decompose experts into compact low-rank matrices, resulting in accelerated inference and memory optimization. In particular, we propose selective decomposition strategy by measuring sensitivity metrics based on weight singular values and activation statistics to automatically identify decomposable expert layers. Then, we share a single V-matrix across all experts and employ a top-k selection for U-matrices. This low-rank matrix sharing and trimming scheme allows for significant parameter reduction while preserving diversity among experts. Comprehensive experiments conducted on Mixtral-8×7B|22B, Phi-3.5-MoE and DeepSeekMoE across multiple datasets reveal that MoE-SVD consistently outperforms existing compression methods in terms of performance-efficiency tradeoffs. Notably, we achieve a remarkable 60\% compression ratio on Mixtral-7x8B and Phi-3.5-MoE, resulting in a 1.5$\times$ inference acceleration with minimal performance degradation. Codes are available in the supplementary materials.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 88