一句话:对 LLM 训练系统中导致崩溃、异常行为及资源浪费的问题进行实证研究。
近年来,大语言模型(LLM)在各类应用中取得显著进展,其训练与评估高度依赖专门的 LLM 训练系统。这类系统通常部署在大量 GPU 上,对 LLM 进行划分并处理大规模数据集。然而,训练系统自身的问题可能导致程序崩溃或出现非预期行为,从而降低开发效率,并浪费 GPU、存储等宝贵资源。本文对 LLM 训练系统中的问题开展实证研究,旨在刻画这些问题的表现与影响。摘要未展开具体的问题分类、数据来源或定量结果。
Large language models (LLMs) have gained significant traction in recent years, driving advancements in various applications. The training and evaluation of these models depend heavily on specialized LLM training systems, which are deployed across numerous GPUs, partition LLMs, and process large datasets. However, issues in LLM training systems can lead to program crashes or unexpected behavior, reducing development productivity and wasting valuable resources such as GPUs and storage.
DOI 原文 ·
@inproceedings{paperbot9172,
title = {An Empirical Study of Issues in Large Language Model Training Systems},
author = {Yanjie Gao and Ruiming Lu and Haoxiang Lin and Yueguo Chen},
booktitle = {FSE},
year = {2025},
doi = {10.1145/3696630.3728538}
}