TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training
2026-04-10 • Distributed, Parallel, and Cluster Computing
Distributed, Parallel, and Cluster ComputingArtificial Intelligence
AI summaryⓘ
The authors address the challenge of efficiently sharing model weights during reinforcement learning training across different computers. They propose Reference-Oriented Storage (ROS), which avoids making extra copies of weights by tracking which GPUs already have them and serving requests directly from those GPUs. Their system, TensorHub, extends this idea with optimizations for network topology, reliability, and consistency. Experiments show TensorHub reduces delays and speeds up training tasks significantly, and it is already used in real-world RL training setups.
reinforcement learningmodel weightsweight transferGPURDMA bandwidthfault tolerancetopology optimizationconsistencyelastic rolloutcross-datacenter
Authors
Chenhao Ye, Huaizheng Zhang, Mingcong Han, Baoquan Zhong, Xiang Li, Qixiang Chen, Xinyi Zhang, Weidong Zhang, Kaihua Jiang, Wang Zhang, He Sun, Wencong Xiao, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau
Abstract
Modern LLM reinforcement learning (RL) workloads require a highly efficient weight transfer system to scale training across heterogeneous computational resources. However, existing weight transfer approaches either fail to provide flexibility for dynamically scaling clusters or incur fundamental data movement overhead, resulting in poor performance. We introduce Reference-Oriented Storage (ROS), a new storage abstraction for RL weight transfer that exploits the highly replicated model weights in place. ROS presents the illusion that certain versions of the model weights are stored and can be fetched on demand. Underneath, ROS does not physically store any copies of the weights; instead, it tracks the workers that hold these weights on GPUs for inference. Upon request, ROS directly uses them to serve reads. We build TensorHub, a production-quality system that extends the ROS idea with topology-optimized transfer, strong consistency, and fault tolerance. Evaluation shows that TensorHub fully saturates RDMA bandwidth and adapts to three distinct rollout workloads with minimal engineering effort. Specifically, TensorHub reduces total GPU stall time by up to 6.7x for standalone rollouts, accelerates weight update for elastic rollout by 4.8x, and cuts cross-datacenter rollout stall time by 19x. TensorHub has been deployed in production to support cutting-edge RL training.