Saved in:
Bibliographic Details
Main Authors: Gu, Hongyaoxing, Chen, Xinzhe, Hu, Lijuan, Liu, Fangfang
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.09281
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914548780367872
author Gu, Hongyaoxing
Chen, Xinzhe
Hu, Lijuan
Liu, Fangfang
author_facet Gu, Hongyaoxing
Chen, Xinzhe
Hu, Lijuan
Liu, Fangfang
contents Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose \textsc{TileQ}, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for \textsc{TileQ} that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that \textsc{TileQ} cuts down additional memory usage up to 10$\times$ and reduces inference latency to $\sim$5\% while preserving state-of-the-art accuracy.
format Preprint
id arxiv_https___arxiv_org_abs_2605_09281
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle TileQ: Efficient Low-Rank Quantization of Mixture-of-Experts with 2D Tiling
Gu, Hongyaoxing
Chen, Xinzhe
Hu, Lijuan
Liu, Fangfang
Machine Learning
Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose \textsc{TileQ}, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for \textsc{TileQ} that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that \textsc{TileQ} cuts down additional memory usage up to 10$\times$ and reduces inference latency to $\sim$5\% while preserving state-of-the-art accuracy.
title TileQ: Efficient Low-Rank Quantization of Mixture-of-Experts with 2D Tiling
topic Machine Learning
url https://arxiv.org/abs/2605.09281