Papers & Research

Second-level updates for trillion-parameter models: Point-to-point weight transfer in large-scale distributed reinforcement learning

Original title: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

AI summary

For reinforcement learning workloads in SGLang, the LMSYS team proposed an RDMA-based point-to-point weight update mechanism as a supplement to the traditional NCCL broadcast method. The design uses source-side CPU engine replicas and Mooncake TransferEngine for P2P RDMA transfer, reducing weight transfer time for the 1T-parameter Kimi-K2 model from 53 seconds to 7.2 seconds, a 7x speedup. The tradeoff is that each training rank requires an additional 32G inference engine replica in CPU memory. This optimization minimizes network redundancy, allows inference servers to resume rollout faster, and is compatible with all mainstream open-source models.

Key Takeaways

  • For reinforcement learning workloads in SGLang,the LMSYS team proposed an RDMA-based…
  • The design uses source-side CPU engine replicas and Mooncake TransferEngine for P2P RDMA…
  • 2 seconds,a 7x speedup.
  • The tradeoff is that each training rank requires an additional 32G inference engine replica in CPU…
  • This optimization minimizes network redundancy,allows inference servers to resume rollout…

Quick Facts

Published
Updated
Source
LMSYS Blog
Category
Papers & Research
AI score
74
Read the original Read the full article at the original source
Why it matters

LMSYS 把分布式 RL 训练的权重同步从 NCCL 广播改成 RDMA P2P,1T 参数模型传输快了 7 倍,做大规模 RL infra 的团队该认真看看这个工程方案。

Related entities

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Second-level updates for trillion-parameter models: Point-to-point weight transfer in large-scale distributed reinforcement learning. April 29, 2026. https://new.nat6.net/en/news/cmokelt5u014nsl4fegroemhe/

View all →