1
DynaTrain: Fast Online Parallelism Switching for Elastic LLM Training
一种快速在线切换并行策略的新方法，让大模型训练更弹性高效。
arXiv:2605.18815v1 Announce Type: new Abstract: Modern large language model (LLM) training is inherently dynamic: resource fluctuations, RLHF phase sh…