Saved in:
Bibliographic Details
Main Authors: Hu, Rizhen, He, Yutong, Yan, Ran, Sun, Mou, Yuan, Binghang, Yuan, Kun
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.16415
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915561540157440
author Hu, Rizhen
He, Yutong
Yan, Ran
Sun, Mou
Yuan, Binghang
Yuan, Kun
author_facet Hu, Rizhen
He, Yutong
Yan, Ran
Sun, Mou
Yuan, Binghang
Yuan, Kun
contents As distributed optimization scales to meet the demands of Large Language Model (LLM) training, hardware failures become increasingly non-negligible. Existing fault-tolerant training methods often introduce significant computational or memory overhead, demanding additional resources. To address this challenge, we propose Memory- and Computation-efficient Fault-tolerant Optimization (MeCeFO), a novel algorithm that ensures robust training with minimal overhead. When a computing node fails, MeCeFO seamlessly transfers its training task to a neighboring node while employing memory- and computation-efficient algorithmic optimizations to minimize the extra workload imposed on the neighboring node handling both tasks. MeCeFO leverages three key algorithmic designs: (i) Skip-connection, which drops the multi-head attention (MHA) module during backpropagation for memory- and computation-efficient approximation; (ii) Recomputation, which reduces activation memory in feedforward networks (FFNs); and (iii) Low-rank gradient approximation, enabling efficient estimation of FFN weight matrix gradients. Theoretically, MeCeFO matches the convergence rate of conventional distributed training, with a rate of $\mathcal{O}(1/\sqrt{nT})$, where n is the data parallelism size and T is the number of iterations. Empirically, MeCeFO maintains robust performance under high failure rates, incurring only a 4.18% drop in throughput, demonstrating 5.0$\times$ to 6.7$\times$ greater resilience than previous SOTA approaches. Codes are available at https://github.com/pkumelon/MeCeFO.
format Preprint
id arxiv_https___arxiv_org_abs_2510_16415
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MeCeFO: Enhancing LLM Training Robustness via Fault-Tolerant Optimization
Hu, Rizhen
He, Yutong
Yan, Ran
Sun, Mou
Yuan, Binghang
Yuan, Kun
Distributed, Parallel, and Cluster Computing
As distributed optimization scales to meet the demands of Large Language Model (LLM) training, hardware failures become increasingly non-negligible. Existing fault-tolerant training methods often introduce significant computational or memory overhead, demanding additional resources. To address this challenge, we propose Memory- and Computation-efficient Fault-tolerant Optimization (MeCeFO), a novel algorithm that ensures robust training with minimal overhead. When a computing node fails, MeCeFO seamlessly transfers its training task to a neighboring node while employing memory- and computation-efficient algorithmic optimizations to minimize the extra workload imposed on the neighboring node handling both tasks. MeCeFO leverages three key algorithmic designs: (i) Skip-connection, which drops the multi-head attention (MHA) module during backpropagation for memory- and computation-efficient approximation; (ii) Recomputation, which reduces activation memory in feedforward networks (FFNs); and (iii) Low-rank gradient approximation, enabling efficient estimation of FFN weight matrix gradients. Theoretically, MeCeFO matches the convergence rate of conventional distributed training, with a rate of $\mathcal{O}(1/\sqrt{nT})$, where n is the data parallelism size and T is the number of iterations. Empirically, MeCeFO maintains robust performance under high failure rates, incurring only a 4.18% drop in throughput, demonstrating 5.0$\times$ to 6.7$\times$ greater resilience than previous SOTA approaches. Codes are available at https://github.com/pkumelon/MeCeFO.
title MeCeFO: Enhancing LLM Training Robustness via Fault-Tolerant Optimization
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2510.16415