SwapMoE: Efficient Memory-Constrained Serving of Large Sparse MoE Models via Dynamic Expert Pruning and Swapping

Mixture of experts (MoE) is a popular technique to improve capacity of large models with conditionally-activated parallel neural network modules (experts). Due to its remarkable scaling performance with sparse computation, it is widely used in modern Large Language Models (LLMs) and Large Vision Models (LVMs). However, serving such large models on edge devices is challenging due to memory constraints. Typical solutions like memory swapping or weight pruning may lead to significantly higher latency or severe accuracy loss. In this paper, we introduce SwapMoE, a framework for efficient continuous MoE-based large models serving with tunable memory budgets. The main idea of SwapMoE is to keep a small dynamic set of important experts, namely Virtual Experts, in the main memory for inference, while seamlessly maintaining how the Virtual Experts map to the actual experts. We use a profiling-guided planner to allocate the resources for SwapMoE that can fully utilize the memory budgets and bandwidth, and an importance-aware scheduler to efficiently identify, update, and use the Virtual Experts for accurate inference. To evaluate SwapMoE, we conduct experiments on multiple edge devices with state-of-the-art MoE-based Large Language Models and Large Vision Models. The results demonstrate remarkable performance of SwapMoE under various memory constraints. Specifically, SwapMoE can enable running large MoE models under tight memory budgets with similar latency to pruned compact models, while with significantly higher accuracy.