From Detection to Recovery: Operational Analysis on LLM Pre-training with 504 GPUs
Paper • 2605.09370 • Published
Lablup Technical Report
This repository is associated with the following arXiv paper:
제목: 이상 탐지에서 자동 복구까지 — 504-GPU LLM 학습의 운영 분석 보고 (Lablup Technical Report)
본 기술 보고서는 63개 노드 NVIDIA B200 프로덕션 클러스터(504 GPU)에서 수집한 55일간의 Prometheus 시계열 데이터와, 224회 다중 노드 학습 작업을 포괄하는 73일간의 운영 로그를 바탕으로, 대규모 LLM 사전학습의 장애 탐지부터 자동 복구까지를 실증 분석합니다. SKT·Upstage·Lablup·NVIDIA Korea·VAST Data 5개 조직이 통합 모니터링 파이프라인을 공유한 공동 운영 환경에서 수행되었습니다.
Lablup Inc.
Please cite this work as "Lablup Inc. (2026)". For inquiries, contact Lablup Inc. at https://www.lablup.com/contact.
@misc{arxiv2605.09370,
title = {From Detection to Recovery: Operational Analysis on LLM Pre-training with 504 GPUs},
author = {{Lablup Inc.}},
year = {2026},
eprint = {2605.09370},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
note = {Daemyung Kang, Eunjin Hwang, Hanjeong Lee, HyeokJin Kim, Hyunhoi Koo, Jeongkyu Shin, Jeongseok Kang, Jihyun Kang, Jinho Heo, Joongi Kim, Junbum Lee, Jungseung Yang, Kyujin Cho, and Youngsook Song},
url = {https://arxiv.org/abs/2605.09370}
}