1 paper
Hanlin Du, Zhiyuan Yan, Yungang Bao +1
RL-based LLM post-training increasingly disaggregates Rollout and Training across separate GPU resources, but static GPU partitioning suffers from severe pipeline bubbles under lon…