Announcing Ray support for PyTorch on Google TPUs

Anyscale and Google Cloud have announced native Ray support for PyTorch workloads on Google Cloud TPUs, extending the PARK stack of PyTorch, AI models, Ray, and Kubernetes to TPU infrastructure. Built on Google’s TorchTPU backend and KubeRay on Google Kubernetes Engine, the integration shifts Ray from a host-level TPU model to one worker per TPU chip, mirroring familiar GPU-oriented PyTorch workflows while providing device isolation, TPU-specific distributed communications, and topology-aware gang scheduling. Ray Train’s TorchTrainer can now orchestrate distributed PyTorch training on single- and multi-host TPU slices, automatically reserving contiguous hardware, configuring distributed environments, and supporting DDP, FSDPv2, and DTensor with minimal changes to training code. The release also supports vLLM-TorchTPU through Ray Serve, allowing teams to deploy OpenAI-compatible LLM inference endpoints on TPU slices using the same declarative configuration approach used for GPUs. Google and Anyscale position the integration as a way for platform teams to manage training, inference, and future reinforcement-learning workloads across GPUs and TPUs under one Kubernetes control plane.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.