Guardado en:
Detalles Bibliográficos
Autores principales: Kang, Xueze, Xiang, Guangyu, Wang, Yuxin, Zhang, Hao, Fang, Yuchu, Zhou, Yuhang, Tang, Zhenheng, Lv, Youhui, Maman, Eliran, Wasserman, Mark, Zameret, Alon, Bian, Zhipeng, Chen, Shushu, Yu, Zhiyou, Wang, Jin, Wu, Xiaoyu, Zheng, Yang, Tian, Chen, Chu, Xiaowen
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2510.00606
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911196826828800
author Kang, Xueze
Xiang, Guangyu
Wang, Yuxin
Zhang, Hao
Fang, Yuchu
Zhou, Yuhang
Tang, Zhenheng
Lv, Youhui
Maman, Eliran
Wasserman, Mark
Zameret, Alon
Bian, Zhipeng
Chen, Shushu
Yu, Zhiyou
Wang, Jin
Wu, Xiaoyu
Zheng, Yang
Tian, Chen
Chu, Xiaowen
author_facet Kang, Xueze
Xiang, Guangyu
Wang, Yuxin
Zhang, Hao
Fang, Yuchu
Zhou, Yuhang
Tang, Zhenheng
Lv, Youhui
Maman, Eliran
Wasserman, Mark
Zameret, Alon
Bian, Zhipeng
Chen, Shushu
Yu, Zhiyou
Wang, Jin
Wu, Xiaoyu
Zheng, Yang
Tian, Chen
Chu, Xiaowen
contents Large-scale LLM pretraining now runs across $10^5$--$10^6$ accelerators, making failures routine and elasticity mandatory. We posit that an elastic-native training system must jointly deliver (i) parameter consistency, (ii) low mean time to recovery (MTTR), (iii) high post-change throughput, and (iv) computation consistency. No prior system achieves all four simultaneously. To achieve these goals, we present ElasWave, which delivers per-step fault tolerance via multi-dimensional scheduling across graph, dataflow, DVFS, and RNG. ElasWave reshapes and reshards micro-batches while preserving the global batch size and gradient scale. It performs online pipeline resharding with asynchronous parameter migration and interleaves ZeRO partitions, reducing parameter recovery processes to disjoint rank-to-rank transfers. It further leverages DVFS to absorb pipeline bubbles and reshards RNG to keep computation consistency. Together, a dynamic communicator enables in-place communication group edits, while per-step in-memory snapshots support online verification and redistribution. We evaluate ElasWave on 96 NPUs and benchmark it against state-of-the-art baselines: throughput improves by $1.35\times$ over ReCycle and $1.60\times$ over TorchFT; communicator recovery completes within one second (up to $82\times/3.6\times$ faster than full/partial rebuilds); migration MTTR drops by as much as $51\%$; and convergence deviation is reduced by approximately $78\%$.
format Preprint
id arxiv_https___arxiv_org_abs_2510_00606
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ElasWave: An Elastic-Native System for Scalable Hybrid-Parallel Training
Kang, Xueze
Xiang, Guangyu
Wang, Yuxin
Zhang, Hao
Fang, Yuchu
Zhou, Yuhang
Tang, Zhenheng
Lv, Youhui
Maman, Eliran
Wasserman, Mark
Zameret, Alon
Bian, Zhipeng
Chen, Shushu
Yu, Zhiyou
Wang, Jin
Wu, Xiaoyu
Zheng, Yang
Tian, Chen
Chu, Xiaowen
Distributed, Parallel, and Cluster Computing
Large-scale LLM pretraining now runs across $10^5$--$10^6$ accelerators, making failures routine and elasticity mandatory. We posit that an elastic-native training system must jointly deliver (i) parameter consistency, (ii) low mean time to recovery (MTTR), (iii) high post-change throughput, and (iv) computation consistency. No prior system achieves all four simultaneously. To achieve these goals, we present ElasWave, which delivers per-step fault tolerance via multi-dimensional scheduling across graph, dataflow, DVFS, and RNG. ElasWave reshapes and reshards micro-batches while preserving the global batch size and gradient scale. It performs online pipeline resharding with asynchronous parameter migration and interleaves ZeRO partitions, reducing parameter recovery processes to disjoint rank-to-rank transfers. It further leverages DVFS to absorb pipeline bubbles and reshards RNG to keep computation consistency. Together, a dynamic communicator enables in-place communication group edits, while per-step in-memory snapshots support online verification and redistribution. We evaluate ElasWave on 96 NPUs and benchmark it against state-of-the-art baselines: throughput improves by $1.35\times$ over ReCycle and $1.60\times$ over TorchFT; communicator recovery completes within one second (up to $82\times/3.6\times$ faster than full/partial rebuilds); migration MTTR drops by as much as $51\%$; and convergence deviation is reduced by approximately $78\%$.
title ElasWave: An Elastic-Native System for Scalable Hybrid-Parallel Training
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2510.00606