Saved in:
Bibliographic Details
Main Authors: Liu, Zhenyu, Liu, Yunzhen, Fan, Zehao, Gagnon, Garrett, Hou, Yayue, Wu, Nan, Kang, Yangwook, Liu, Liu
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.17073
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917156672765952
author Liu, Zhenyu
Liu, Yunzhen
Fan, Zehao
Gagnon, Garrett
Hou, Yayue
Wu, Nan
Kang, Yangwook
Liu, Liu
author_facet Liu, Zhenyu
Liu, Yunzhen
Fan, Zehao
Gagnon, Garrett
Hou, Yayue
Wu, Nan
Kang, Yangwook
Liu, Liu
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2512_17073
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation
Liu, Zhenyu
Liu, Yunzhen
Fan, Zehao
Gagnon, Garrett
Hou, Yayue
Wu, Nan
Kang, Yangwook
Liu, Liu
Machine Learning
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.
title Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation
topic Machine Learning
url https://arxiv.org/abs/2512.17073