Blog Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL We introduced a RDMA-based, Peer to Peer weight update mechanism for RL workloads in SGLang as a supplement to traditional NCCL broadcast methods, compatible with all major open source models. By util... Jiadong Guo, Xin Ji, Letian Ruan, Teng Ma, Chenyang Zhao, Yueming Yuan, Zhichen Zeng April 29, 2026
LMSYS团队针对SGLang中的强化学习工作负载，提出了一种基于RDMA的点对点权重更新机制，作为传统NCCL广播方法的补充。该设计利用源端CPU引擎副本和Mooncake TransferEngine进行P2P RDMA传输，将拥有1T参数的Kimi-K2模型的权重传输时间从53秒大幅缩短至7.2秒，提速7倍。其代价是每个训练等级需在CPU内存中额外占用一个32G的推理引擎副本。此优化最大限度地减少了网络冗余，允许推理服务器更快恢复rollout过程，且兼容所有主流开源模型。