OpGuard is a debugging system for large-scale LLM training. When two training runs diverge, it uses bitwise alignment to compare them at tensor boundaries and pinpoint the first divergent operator—turning vague loss-curve anomalies into precise, actionable evidence.
In production use at ByteDance, OpGuard has reduced root-cause localization from days of manual effort to minutes.
Preprint (PDF) · OpGuard will appear at OSDI ’26:
Ziming Zhou, Yinjie Zhao, Hang Zhu, Wenxiao Wang, Zhihao Bai, Yun Zhang, Shuguang Wang, Haibin Lin, Peng Huang.
OpGuard: Bitwise Alignment for Precise and General Debugging of Production LLM Training.
In Proceedings of the 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI ’26), Seattle, WA, USA, July 2026.
BibTeX
@inproceedings{OpGuard2026OSDI,
author = {Zhou, Ziming and Zhao, Yinjie and Zhu, Hang and Wang, Wenxiao and Bai, Zhihao and Zhang, Yun and Wang, Shuguang and Lin, Haibin and Huang, Peng},
title = {{OpGuard}: Bitwise Alignment for Precise and General Debugging of Production {LLM} Training},
booktitle = {Proceedings of the 20th USENIX Symposium on Operating Systems Design and Implementation},
series = {OSDI '26},
month = {July},
year = {2026},
address = {Seattle, WA, USA},
publisher = {USENIX Association},
}| Section | Description |
|---|---|
| Bitwise alignment | Core abstraction for comparing training runs |
| Live trace demo | Explore divergence evidence in a Perfetto-style viewer |
| Workflow | How OpGuard fits into production debugging |
| Impact | Production case studies and timing results |
Source code and internal tooling are not publicly released.
For questions about the paper, demos, or access inquiries, please email the authors listed above.
Founded in 2023, ByteDance Seed Team is dedicated to crafting the industry's most advanced AI foundation models. The team aspires to become a world-class research team and make significant contributions to the advancement of science and society.