# Anyscale and Google Cloud Add Native PyTorch Support for Ray on TPUs By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/36258 Published: 2026-10-21T02:48:02.000Z Updated: 2026-10-21T02:48:02.000Z Section: Technology > The integration assigns one Ray worker to each TPU chip and supports training and model serving on Google Kubernetes Engine. Anyscale and Google Cloud introduced native PyTorch support for Ray on Google TPUs on Oct. 20, letting developers coordinate AI training and model serving on Google Kubernetes Engine (GKE). Each Ray worker now maps to one TPU chip. Ray can schedule a job’s workers together on the same TPU slice. The integration combines Google’s TorchTPU with [KubeRay, the Kubernetes operator for Ray workloads](). TorchTPU runs PyTorch operations on TPUs, while Ray assigns workers and coordinates jobs. On GKE, teams can manage TPU jobs alongside workloads that use GPUs. For training, Ray Train can coordinate PyTorch jobs across multiple hosts within one TPU slice. An example configuration puts 16 workers on a 4x4 TPU-V6E slice, with one worker assigned to each chip. Ray Train also sets up the distributed environment variables the workers need to run the job together. Ray’s placement tools schedule a job’s workers as a group on the same slice. The worker-to-chip layout isolates each worker’s assigned device, allowing developers to organize training around individual chips while Ray distributes work across the cluster. The described PyTorch configuration spans hosts within a slice. For model serving, Ray Serve LLM can use vLLM-TorchTPU to run OpenAI-compatible endpoints on TPU slices. A vLLM plugin connects to TorchTPU for execution on the chips, while Ray Serve LLM coordinates serving. The integration therefore covers both distributed training and applications that deliver model responses. The latest integration adds native PyTorch support with one worker assigned to each chip. Ray 2.59.0’s PyTorch backend does not support training across multiple TPU slices. ## Links in this article - [KubeRay, the Kubernetes operator for Ray workloads](https://www.tokenpost.com/news/technology/32744)