Accelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism 事件

Accelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism arXiv:2605.23945v1 Announce Type: new Abstract: Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long r