SCOUT: Symmetric Consensus Outlier Detection for Failure Localization in LLM Pre-Training
大模型预训练遇卡顿不知根源?SCOUT用共识异常检测精准锁定故障节点,摆脱同步困局。
arXiv:2608.11034v1 Announce Type: cross Abstract: In LLM pre-training, synchronization propagates rank-local stalls, slowdowns, and numerical errors i…