Google releases TorchTPU for running PyTorch workloads on TPUs

Google announced TorchTPU as part of a set of products and integrations for running PyTorch on Google Cloud or in users’ own infrastructure.[1] The PyTorch team describes it as an open-source collaboration between Google and Meta that provides a native PyTorch backend for Google TPUs.[2] Google says TPUs now join GPUs and CPUs as a native option for PyTorch.[1] The project began as an effort at Google in mid-2025, according to Google and Meta.[2] TorchTPU is available as open source on GitHub.[1]

Google says existing PyTorch workloads can run on TPUs with minimal code changes and familiar APIs.[1] The PyTorch team gives changing `.to("cuda")` to `.to("tpu")` as an example of a trivial code adjustment.[2] Google says TorchTPU supports training, serving and deployment using tools including vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face and LiteRT.[1] It also supports eager-mode debugging and `torch.compile` for production-level performance optimization.[1] For models in Transformers v5, Google says developers can deploy to TPUs by setting `device_map="tpu"`.[1]

Google says Hugging Face’s Diffusers library runs natively on TPUs, with eager mode, `torch.compile` and tensor-parallelism support.[1] Google and Hugging Face also partnered to run popular open-weight models from the Hugging Face Hub natively on TorchTPU.[1] The integration supports continuous batching, quantization and multi-chip parallelism on TPU hardware.[1] Google describes its inference setup as combining TorchTPU integrations with Kubernetes-native fleet routing across TPUs and GPUs to optimize throughput per dollar.[1] These capabilities give infrastructure teams options to assess when evaluating TPU deployment alongside their existing hardware and workflows.

The TorchTPU team says the project is already in production for inference and training workloads at some of Google Cloud’s largest customers.[2] The team also says it validated out-of-the-box execution on the 5,000 most-downloaded Hugging Face models.[2] Separately, Google and Meta reported that their TorchTitan TPU path reached more than 29% model FLOPs utilization on Qwen3 235B and 51–52% on Qwen3-8B.[3] Those figures are reported results for the named models, not a guarantee of performance on other workloads. Google has also said TorchTPU will not automatically lead AI developers to abandon Nvidia hardware.[4]