FSE 2025
Mon 23 - Fri 27 June 2025 Trondheim, Norway
Tue 24 Jun 2025 11:00 - 11:20 at Cosmos 3C - SE for LLM Chair(s): Hongyu Zhang

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.

This paper presents the first comprehensive empirical study of issues in LLM training systems. We conducted a manual analysis of 300 high-quality issue reports and corresponding fix commits from the GitHub repositories of three prominent LLM training systems: Microsoft DeepSpeed, NVIDIA Megatron-LM, and Hugging Face Transformers. Our analysis identified common symptoms, root causes, typical fixes, and debugging and testing practices in LLM training systems. Our major findings include: (1) LLM training systems exhibit issues and trends that are uncommon in traditional deep learning, such as Concurrency Error and Tensor Management Error occurring in parallel training, which are particularly difficult to diagnose and resolve. (2) The primary root causes of these issues are API Misuse (19.67%), Configuration Error (18.33%), and General Code Error (16.33%), respectively. Such issues often arise from the rapid evolution of the systems, the integration of complex external dependencies, and a configuration-driven development paradigm. (3) Current testing and debugging practices are often insufficient for identifying issues related to parallel training and large-scale numerical computations. Based on our findings, we propose several research topics and tooling improvements that can facilitate the future development of LLMs.

Tue 24 Jun

Displayed time zone: Amsterdam, Berlin, Bern, Rome, Stockholm, Vienna change

10:30 - 12:30
10:30
10m
Talk
Enhancing Code LLM Training with Programmer Attention
Ideas, Visions and Reflections
Yifan Zhang Vanderbilt University, Chen Huang Sichuan University, Zachary Karas Vanderbilt University, Thuy Dung Nguyen Vanderbilt University, Kevin Leach Vanderbilt University, Yu Huang Vanderbilt University
10:40
20m
Talk
Risk Assessment Framework for Code LLMs via Leveraging Internal States
Industry Papers
Yuheng Huang The University of Tokyo, Lei Ma The University of Tokyo & University of Alberta, Keizaburo Nishikino Fujitsu Limited, Takumi Akazaki Fujitsu Limited
11:00
20m
Talk
An Empirical Study of Issues in Large Language Model Training Systems
Industry Papers
Yanjie Gao Microsoft Research, Ruiming Lu Shanghai Jiao Tong University, Haoxiang Lin Microsoft Research, Yueguo Chen Renmin University of China
DOI
11:20
20m
Talk
Look Before You Leap: An Exploratory Study of Uncertainty Analysis for Large Language Models
Journal First
Yuheng Huang The University of Tokyo, Norman Song , Zhijie Wang University of Alberta, Shengming Zhao University of Alberta, Huaming Chen The University of Sydney, Felix Juefei-Xu New York University, Lei Ma The University of Tokyo & University of Alberta
Link to publication DOI Pre-print
11:40
10m
Talk
EvidenceBot: A Privacy-Preserving, Customizable RAG-Based Tool for Enhancing Large Language Model Interactions
Demonstrations
Nafiz Imtiaz Khan Department of Computer Science, University of California, Davis, Vladimir Filkov University of California at Davis, USA
11:50
20m
Talk
OpsEval: A Comprehensive Benchmark Suite for Evaluating Large Language Models’ Capability in IT Operations Domain
Industry Papers
Yuhe Liu Tsinghua University, Changhua Pei Computer Network Information Center at Chinese Academy of Sciences, Longlong Xu Tsinghua University, Bohan Chen Tsinghua University, Mingze Sun Tsinghua University, Zhirui Zhang Beijing University of Posts and Telecommunications, Yongqian Sun Nankai University, Shenglin Zhang Nankai University, Kun Wang Zhejiang University, Haiming Zhang Chinese Academy of Sciences, Jianhui Li Computer Network Information Center at Chinese Academy of Sciences, Gaogang Xie Computer Network Information Center at Chinese Academy of Sciences, Xidao Wen BizSeer, Xiaohui Nie Computer Network Information Center at Chinese Academy of Sciences, Minghua Ma Microsoft, Dan Pei Tsinghua University
12:10
20m
Talk
Hallucination Detection in Large Language Models with Metamorphic Relations
Research Papers
Borui Yang Beijing University of Posts ad Telecommunications, Md Afif Al Mamun University of Calgary, Jie M. Zhang King's College London, Gias Uddin York University, Canada
DOI

Information for Participants
Tue 24 Jun 2025 10:30 - 12:30 at Cosmos 3C - SE for LLM Chair(s): Hongyu Zhang
Info for room Cosmos 3C:

Cosmos 3C is the third room in the Cosmos 3 wing.

When facing the main Cosmos Hall, access to the Cosmos 3 wing is on the left, close to the stairs. The area is accessed through a large door with the number “3”, which will stay open during the event.