Elastic MoE: Unlocking the Inference-Time Scalability of Mixture-of-Experts

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Gu, Naibin, Zhang, Zhenyu, Feng, Yuchen, Chen, Yilong, Fu, Peng, Lin, Zheng, Wang, Shuohuan, Sun, Yu, Wu, Hua, Wang, Weiping, Wang, Haifeng
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917480185724928
author Gu, Naibin
Zhang, Zhenyu
Feng, Yuchen
Chen, Yilong
Fu, Peng
Lin, Zheng
Wang, Shuohuan
Sun, Yu
Wu, Hua
Wang, Weiping
Wang, Haifeng
author_facet Gu, Naibin
Zhang, Zhenyu
Feng, Yuchen
Chen, Yilong
Fu, Peng
Lin, Zheng
Wang, Shuohuan
Sun, Yu
Wu, Hua
Wang, Weiping
Wang, Haifeng
contents Mixture-of-Experts (MoE) models typically fix the number of activated experts $k$ at both training and inference. However, real-world deployments often face heterogeneous hardware, fluctuating workloads, and diverse quality-latency requirements, while training separate models for each scenario is costly. Considering that MoE models already operate with sparse activation, adjusting the number of activated experts offers a natural path to serving diverse budgets with a single model. Yet, we find that activating more experts $k'$ ($> k$) at inference does not yield the expected gains. Instead, performance degrades rapidly after only a slight increase, a phenomenon we term the \textit{inference-time scaling wall}. Further investigation reveals that this degradation stems from a lack of learned collaboration among experts. To address this, we introduce \textbf{Elastic Mixture-of-Experts (EMoE)}, a novel training framework that enables MoE models to elastically vary the number of activated experts at inference. By simultaneously training experts to collaborate in diverse combinations and encouraging the router to make high-quality selections, EMoE ensures robust performance across inference budgets. Extensive experiments across four MoE architectures (7B--21B) and nine benchmarks show that EMoE significantly expands the effective scaling range to 2-3$\times$ the training-time $k$, while also achieving higher peak performance.
format Preprint
id arxiv_https___arxiv_org_abs_2509_21892
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Elastic MoE: Unlocking the Inference-Time Scalability of Mixture-of-Experts
Gu, Naibin
Zhang, Zhenyu
Feng, Yuchen
Chen, Yilong
Fu, Peng
Lin, Zheng
Wang, Shuohuan
Sun, Yu
Wu, Hua
Wang, Weiping
Wang, Haifeng
Computation and Language
Artificial Intelligence
Machine Learning
Mixture-of-Experts (MoE) models typically fix the number of activated experts $k$ at both training and inference. However, real-world deployments often face heterogeneous hardware, fluctuating workloads, and diverse quality-latency requirements, while training separate models for each scenario is costly. Considering that MoE models already operate with sparse activation, adjusting the number of activated experts offers a natural path to serving diverse budgets with a single model. Yet, we find that activating more experts $k'$ ($> k$) at inference does not yield the expected gains. Instead, performance degrades rapidly after only a slight increase, a phenomenon we term the \textit{inference-time scaling wall}. Further investigation reveals that this degradation stems from a lack of learned collaboration among experts. To address this, we introduce \textbf{Elastic Mixture-of-Experts (EMoE)}, a novel training framework that enables MoE models to elastically vary the number of activated experts at inference. By simultaneously training experts to collaborate in diverse combinations and encouraging the router to make high-quality selections, EMoE ensures robust performance across inference budgets. Extensive experiments across four MoE architectures (7B--21B) and nine benchmarks show that EMoE significantly expands the effective scaling range to 2-3$\times$ the training-time $k$, while also achieving higher peak performance.
title Elastic MoE: Unlocking the Inference-Time Scalability of Mixture-of-Experts
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2509.21892