Anyscale and Google Cloud Announce Native Ray Support for PyTorch on TPUs

On October 20, 2026, compute platform company Anyscale and Google Cloud jointly announced that the open-source distributed computing framework Ray now natively supports running PyTorch workloads on Google Tensor Processing Units (TPUs). This integration builds on Google's TorchTPU project introduced last year and KubeRay on Google Kubernetes Engine (GKE), aiming to provide a unified control plane for platform engineering and machine learning teams to schedule and manage distributed training, vLLM inference, and reinforcement learning tasks across GPUs and TPUs.

Historically, using TPUs with Ray required adapting to a Single-Program, Multiple Data (SPMD) execution model designed primarily for JAX, where a single Ray worker process per host VM managed all attached TPU chips. This update introduces a native PyTorch experience with one worker per TPU chip: each Ray worker requests a single logical TPU device, similar to setting num_gpus=1 on GPUs. Ray automatically isolates and assigns individual TPU devices to workers via TPU_VISIBLE_CHIPS. Developers can migrate existing PyTorch code by simply changing the device string from "cuda" to "tpu" and switching the communication backend from "nccl" to "tpu_dist".

For large-scale training scenarios, Ray Train's TorchTrainer now supports automatic configuration of TPU slices. By specifying use_tpu=True along with the target topology (e.g., "4x4") and accelerator type (e.g., "TPU-V6E"), the system automatically reserves atomic TPU slices, ensuring all training workers are gang-scheduled onto contiguous physical TPU slices over high-speed interconnects, preventing resource fragmentation. Additionally, combined with the vLLM-TorchTPU plugin, developers can deploy OpenAI-compatible online inference services on TPU slices using the same declarative API.