TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training

2026-04-10

AI Summary

This work addresses the challenges of low weight-transfer efficiency and poor scalability in large-model reinforcement learning training under heterogeneous resources. The authors propose TensorHub, a novel system that introduces reference-oriented storage (ROS) abstraction to enable zero-copy, efficient weight sharing by servicing read requests on-demand using existing weight replicas already present on GPUs. TensorHub integrates RDMA-based high-speed networking, topology-aware transmission, strong consistency protocols, and fault-tolerance mechanisms to significantly enhance transfer efficiency. Empirical results demonstrate substantial reductions in GPU stall time—by 6.7×, 4.8×, and 19×—in single-node, elastic scaling, and cross-datacenter rollback scenarios, respectively. The system has been successfully deployed in production environments.

Technology Category

• Machine Learning: Scalability of ML Systems
• Multiagent Systems: Adversarial Agents
• Intelligent Robots: Learning & Optimization for ROB

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Experiences and lessons learnt from Web-based algorithms and system deployments
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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.

Problem

• weight transfer
• large language models
• reinforcement learning
• elastic scaling
• heterogeneous resources

Innovation

• Reference-Oriented Storage
• weight transfer
• LLM reinforcement learning
• elastic training
• RDMA optimization