Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation

TL;DR

This paper introduces a low-rank compensation method for mixture-of-experts models that reduces bandwidth usage during inference, maintaining accuracy while improving throughput by combining expert offloading with precision restoration.

Contribution

It proposes a router-guided low-rank compensation technique that enhances bandwidth efficiency in MoE models during inference, integrating with offloading strategies.

Findings

Achieves better bandwidth-accuracy trade-off compared to static quantization.

Improves inference throughput on GPU and GPU-NDP systems.

Maintains expert heterogeneity benefits with reduced data transfer.

Abstract

Mixture-of-Experts (MoE) models scale capacity via sparse activation but stress memory and bandwidth. Offloading alleviates GPU memory by fetching experts on demand, yet token-level routing causes irregular transfers that make inference I/O-bound. Static uniform quantization reduces traffic but degrades accuracy under aggressive compression by ignoring expert heterogeneity. We present Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation, which performs router-guided precision restoration using precomputed low-rank compensators. At inference time, our method transfers compact low-rank factors with Top-n (n<k) experts per token and applies compensation to them, keeping others low-bit. Integrated with offloading on GPU and GPU-NDP systems, our method delivers a superior bandwidth-accuracy trade-off and improved throughput.