Ion Stoica 在 PyTorchCon 2026 讲述 Ray 如何成为 AI 工作负载的分布式计算层
Ray 作者亲自讲它怎么从 RL 项目变成现在的分布式底座，还预告要用 Rust 重写核心组件，做分布式训练的都该看看。
Ray 起源于强化学习场景，Ion Stoica（UC Berkeley / Anyscale）在 PyTorchCon 2026 演讲中回顾其演进过程。演讲涵盖数据处理、训练、推理服务和后训练的复杂度变化，并介绍了面向 TPU pod 和 NVIDIA Blackwell GPU 的拓扑感知调度、基于 NIXL 的 GPU 间传输以及原生 TorchFT 集成。未来规划包括用 Rust 重写 Ray 主要组件、更紧密集成 Kubernetes，并与硬件厂商改进加速器支持。
Ray began with reinforcement learning. At #PyTorchCon 2026, Ion Stoica @istoica05 (@UCBerkeley / @anyscalecompute traces its evolution into the distributed compute layer behind today’s RL frameworks.
In “Ray’s Evolution to Support AI Workloads,” Ion examines how data processing, training, serving, and post-training have grown in complexity and scale.
He highlights topology-aware scheduling for TPU pods and NVIDIA Blackwell GPUs, GPU-to-GPU transfers using NIXL, and native TorchFT integration.
Looking ahead, he outlines plans to rewrite major Ray components in Rust, integrate more closely with Kubernetes, and work with hardware providers to improve accelerator support.
