TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training

A team of researchers, including Chenhao Ye and professors Andrea C. Arpaci-Dusseau and Remzi H. Arpaci-Dusseau, has introduced TensorHub, a system designed to solve a critical bottleneck in large language model (LLM) reinforcement learning (RL) training. The core problem is the inefficient transfer of model weights across a dynamically scaling cluster of heterogeneous GPUs, which creates significant data movement overhead and stalls training. Their solution is a novel storage abstraction called Reference-Oriented Storage (ROS). ROS creates the illusion that specific versions of model weights are stored and available for fetch, but instead of physically copying data, it intelligently tracks which workers already hold those weights in GPU memory for inference and uses those live copies to serve read requests directly.

TensorHub builds upon the ROS concept by adding production-ready features like topology-optimized data transfer, strong consistency guarantees, and fault tolerance. The system's evaluation demonstrates dramatic performance gains: it can fully saturate available RDMA network bandwidth and adapt to three distinct RL rollout workloads with minimal engineering effort. Specifically, TensorHub reduces total GPU stall time by up to 6.7x for standalone rollouts, accelerates weight updates for elastic scaling scenarios by 4.8x, and slashes cross-datacenter rollout stall time by a massive 19x. The paper confirms that TensorHub is already deployed in production to support cutting-edge RL training pipelines, indicating its practical viability beyond academic benchmarks.