At PyTorch Conference 2026 on October 20, Google announced that TorchTPU, the native PyTorch stack for Google TPUs, is now open source on GitHub. The stack was co-developed with the PyTorch team at Meta. With TorchTPU, TPUs join GPUs and CPUs as a native option for PyTorch, so existing workloads can move onto TPU with minimal code changes and familiar APIs.
Developers can debug in eager mode, then use torch.compile for production-level performance on TPU. The PyTorch blog frames TorchTPU as a native backend that prioritizes eager execution, unlike earlier PyTorch on XLA paths built on Lazy Tensors. Porting often means pointing the device at tpu instead of cuda. Google says TorchTPU covers about 450 Aten ops and has been validated out of the box on about 5,000 of the most downloaded Hugging Face models.
On the ecosystem side, Google highlights serving and model-hub paths that matter most day to day: native PyTorch backends for vLLM and SGLang, and Hugging Face Transformers v5 support for sending models to TPU through device map settings. Diffusers is also listed as running natively with eager mode, torch.compile, and tensor parallelism. The open-source release is meant to keep TPU a first-class option alongside GPUs and CPUs for the PyTorch community.