REAM: Merging Improves Pruning of Experts in LLMs

TL;DR

REAM is a novel method that merges experts in Mixture-of-Experts LLMs to reduce memory usage while maintaining performance, outperforming pruning-based methods on various benchmarks.

Contribution

Introduces REAM, a merging technique for experts in MoE LLMs that better preserves performance compared to pruning methods.

Findings

REAM often outperforms pruning baselines.

REAM maintains performance close to original models.

Trade-offs exist between different task performances based on data mix.

Abstract

Mixture-of-Experts (MoE) large language models (LLMs) are among the top-performing architectures. The largest models, often with hundreds of billions of parameters, pose significant memory challenges for deployment. Traditional approaches to reduce memory requirements include weight pruning and quantization. Motivated by the Router-weighted Expert Activation Pruning (REAP) that prunes experts, we propose a novel method, Router-weighted Expert Activation Merging (REAM). Instead of removing experts, REAM groups them and merges their weights, better preserving original performance. We evaluate REAM against REAP and other baselines across multiple MoE LLMs on diverse multiple-choice (MC) question answering and generative (GEN) benchmarks. Our results reveal a trade-off between MC and GEN performance that depends on the mix of calibration data. By controlling the mix of general, math and coding data, we examine the Pareto frontier of this trade-off and show that REAM often outperforms the baselines and in many cases is comparable to the original uncompressed models.