Delta Decompression for MoE-based LLMs Compression

TL;DR

MoE LLMs incur large parameter and memory overhead, hindering deployment. The paper proposes D^2-MoE, a delta-based compression framework that decomposes each expert into a shared base weight $W_b$ and an expert-specific delta $\Delta W_i$, merges bases via Fisher information, compresses deltas with truncation-aware SVD, and applies a two-phase semi-dynamical pruning on the base; this pipeline enables high compression ratios with minimal performance loss and no retraining. Empirical results across several state-of-the-art MoE LLMs demonstrate superior compression quality, improved throughput, and robust performance on language modeling and reasoning tasks at 40–60% compression. The approach advances practical deployment of large MoE models by balancing redundancy reduction, diversity preservation, and efficiency, with open-source code available for reproducibility.

Abstract

Mixture-of-Experts (MoE) architectures in large language models (LLMs) achieve exceptional performance, but face prohibitive storage and memory requirements. To address these challenges, we present $D^2$-MoE, a new delta decompression compressor for reducing the parameters of MoE LLMs. Based on observations of expert diversity, we decompose their weights into a shared base weight and unique delta weights. Specifically, our method first merges each expert's weight into the base weight using the Fisher information matrix to capture shared components. Then, we compress delta weights through Singular Value Decomposition (SVD) by exploiting their low-rank properties. Finally, we introduce a semi-dynamical structured pruning strategy for the base weights, combining static and dynamic redundancy analysis to achieve further parameter reduction while maintaining input adaptivity. In this way, our $D^2$-MoE successfully compact MoE LLMs to high compression ratios without additional training. Extensive experiments highlight the superiority of our approach, with over 13% performance gains than other compressors on Mixtral|Phi-3.5|DeepSeek|Qwen2 MoE LLMs at 40$\sim$60% compression rates. Codes are available in https://github.com/lliai/D2MoE.

Figures (6)

• Figure 1: Centered Kernel Alignment (CKA) similarity of experts weights of Mixtral-8x7B, Phi-3.5-MoE, DeepSeekMoE-16B-Base.
• Figure 2: Single values energy retention of experts original weights, merged base weights and delta weights (difference in original weights and merged base weights) from Mixtral-8x7B, Phi-3.5-MoE, DeepSeekMoE-16B-Base.
• Figure 3: Overall Process of $D^2$-MoE. We first merge original expert weights into a shared base weight, weighted according to their Fisher importance. Then, we derive delta weights and compress them using Singular Value Decomposition (SVD). Finally, we apply a two-step pruning strategy: static column-wise pruning followed by dynamic column-wise pruning to further optimize the base weight.
• Figure 4: Expanding $D^2$-MoE via Delta Weights Trimming.
• Figure 5: CKA of Delta V weights and Delta U weights of Mixtral-8x7B