MemFine: Memory-Aware Fine-Grained Scheduling for MoE Training

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Zhao, Lu, Shi, Rong, Zhang, Shaoqing, Chen, Yueqiang, He, Baoguo, Sun, Hongfeng, Yin, Ziqing, Su, Shangchao, Cui, Zhiyan, Dong, Liang, Li, Xiyuan, Wang, Lingbin, He, Jianwei, Ma, Jiesong, Huang, Weikang, Tong, Jianglei, Gao, Dongdong, Zhang, Jian, Tian, Hong, Shen, Hui, Luo, Zongtai, Sun, Zhaoqun, Niu, Hongxing, Sun, Yue
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866918285384089600
author Zhao, Lu
Shi, Rong
Zhang, Shaoqing
Chen, Yueqiang
He, Baoguo
Sun, Hongfeng
Yin, Ziqing
Su, Shangchao
Cui, Zhiyan
Dong, Liang
Li, Xiyuan
Wang, Lingbin
He, Jianwei
Ma, Jiesong
Huang, Weikang
Tong, Jianglei
Gao, Dongdong
Zhang, Jian
Tian, Hong
Shen, Hui
Luo, Zongtai
Sun, Zhaoqun
Niu, Hongxing
Sun, Yue
author_facet Zhao, Lu
Shi, Rong
Zhang, Shaoqing
Chen, Yueqiang
He, Baoguo
Sun, Hongfeng
Yin, Ziqing
Su, Shangchao
Cui, Zhiyan
Dong, Liang
Li, Xiyuan
Wang, Lingbin
He, Jianwei
Ma, Jiesong
Huang, Weikang
Tong, Jianglei
Gao, Dongdong
Zhang, Jian
Tian, Hong
Shen, Hui
Luo, Zongtai
Sun, Zhaoqun
Niu, Hongxing
Sun, Yue
contents The training of large-scale Mixture of Experts (MoE) models faces a critical memory bottleneck due to severe load imbalance caused by dynamic token routing. This imbalance leads to memory overflow on GPUs with limited capacity, constraining model scalability. Existing load balancing methods, which cap expert capacity, compromise model accuracy and fail on memory-constrained hardware. To address this, we propose MemFine, a memory-aware fine-grained scheduling framework for MoE training. MemFine decomposes the token distribution and expert computation into manageable chunks and employs a chunked recomputation strategy, dynamically optimized through a theoretical memory model to balance memory efficiency and throughput. Experiments demonstrate that MemFine reduces activation memory by 48.03% and improves throughput by 4.42% compared to full recomputation-based baselines, enabling stable large-scale MoE training on memory-limited GPUs.
format Preprint
id arxiv_https___arxiv_org_abs_2511_21431
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MemFine: Memory-Aware Fine-Grained Scheduling for MoE Training
Zhao, Lu
Shi, Rong
Zhang, Shaoqing
Chen, Yueqiang
He, Baoguo
Sun, Hongfeng
Yin, Ziqing
Su, Shangchao
Cui, Zhiyan
Dong, Liang
Li, Xiyuan
Wang, Lingbin
He, Jianwei
Ma, Jiesong
Huang, Weikang
Tong, Jianglei
Gao, Dongdong
Zhang, Jian
Tian, Hong
Shen, Hui
Luo, Zongtai
Sun, Zhaoqun
Niu, Hongxing
Sun, Yue
Distributed, Parallel, and Cluster Computing
The training of large-scale Mixture of Experts (MoE) models faces a critical memory bottleneck due to severe load imbalance caused by dynamic token routing. This imbalance leads to memory overflow on GPUs with limited capacity, constraining model scalability. Existing load balancing methods, which cap expert capacity, compromise model accuracy and fail on memory-constrained hardware. To address this, we propose MemFine, a memory-aware fine-grained scheduling framework for MoE training. MemFine decomposes the token distribution and expert computation into manageable chunks and employs a chunked recomputation strategy, dynamically optimized through a theoretical memory model to balance memory efficiency and throughput. Experiments demonstrate that MemFine reduces activation memory by 48.03% and improves throughput by 4.42% compared to full recomputation-based baselines, enabling stable large-scale MoE training on memory-limited GPUs.
title MemFine: Memory-Aware Fine-Grained Scheduling for MoE Training
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2511.21431