Accelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism
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 G...
Long Zhao, Qinghe Wang, Jiaan Zhu et al.
· Proceedings of the 55th Inte... · 0 citations