LLaMA-MoE v2: Exploring Sparsity of LLaMA from Perspective of Mixture-of-Experts with Post-Training

TL;DR

This paper investigates the sparsity of LLaMA models by constructing Mixture-of-Experts (MoE) modules in transformer blocks, evaluating their performance across domains, and proposing a post-training strategy to mitigate sparsity-induced performance loss.

Contribution

It introduces a comprehensive analysis of sparsifying LLaMA with MoE, explores expert construction methods, and proposes a two-stage post-training strategy to improve performance after sparsification.

Findings

Sparsified LLaMA models maintain competitive performance across domains.

Different expert construction methods impact sparsity and effectiveness.

Post-training strategies can recover performance degraded by sparsity.

Abstract

Recently, inspired by the concept of sparsity, Mixture-of-Experts (MoE) models have gained increasing popularity for scaling model size while keeping the number of activated parameters constant. In this study, we thoroughly investigate the sparsity of the dense LLaMA model by constructing MoE for both the attention (i.e., Attention MoE) and MLP (i.e., MLP MoE) modules in the transformer blocks. Specifically, we investigate different expert construction methods and granularities under the same activation conditions to analyze the impact of sparsifying the model. Additionally, to comprehensively evaluate the model's capabilities across various domains (e.g., conversation, code, math) after sparsification, we apply sparsity to the instructed large language models (LLMs) and construct instructed MoE models. To counteract the performance degradation resulting from increased sparsity, we design a two-stage post-training strategy to enhance model performance. Experiments on the LLaMA3 model demonstrate the potential effectiveness of this approach for future developments of instructed MoE models. The source codes and models are available at: \url{https://github.com/OpenSparseLLMs/LLaMA-MoE-v2}.