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

In three linesPAT, an adaptive tensor parallelism method, optimizes the generation stage in synchronous RLHF. It dynamically reconfigures parallelization during decoding to compensate for response-length skew. Implemented on SGLang/VeRL, PAT reduces generation latency by up to 34.6% on LLaMA3.1-8B and Qwen3-14B.

Summary generated by Claude — human-verified