Google and Meta have open-sourced TorchTPU, a native PyTorch backend designed to run PyTorch models directly on Cloud TPUs. Announced on October 20, 2026, at the PyTorch Conference 2026, the software stack integrates TPUs as native PyTorch devices alongside GPUs and CPUs using the standard torch.device("tpu") interface. Google recommends TorchTPU to replace PyTorch/XLA for all new workloads.
TorchTPU plugs into PyTorch through the PrivateUse1 interface, utilizing ordinary PyTorch tensors without wrappers or subclasses while supporting divergent MPMD execution. The open-source release includes four repositories—TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan—under the Apache 2.0 license for Python 3.11+. The backend ships with three eager modes, supports torch.compile for production-level throughput, and scales up to 8,192 Ironwood TPU chips with integrations for Hugging Face, vLLM, and TorchTitan.
This release matters because it streamlines cloud TPU development by replacing PyTorch/XLA with a native backend that achieves up to 52% model floating-point utilization during training.