Google open-sources TorchTPU, a native PyTorch backend for TPUs

What changes

Google has open-sourced TorchTPU, a backend that runs PyTorch models directly on Cloud TPUs. The project was built with Meta's PyTorch team and announced at PyTorch Conference 2026 on October 20. TPUs now sit alongside GPUs and CPUs as a native PyTorch device.

Google says TorchTPU replaces the earlier PyTorch/XLA and recommends it for all new workloads. Developers can target a TPU by setting torch.device("tpu") in existing code.

How it works

TorchTPU plugs into PyTorch's PrivateUse1 interface, and tensors remain ordinary PyTorch tensors. It offers three eager modes: Strict by default, Debug for troubleshooting and DeferAndFuse to cut memory traffic. For peak throughput, torch.compile(backend="tpu") captures the graph, which is lowered through StableHLO to OpenXLA and compiled to TPU machine code.

The numbers

By Google's own reporting, TorchTitan reached up to 52% MFU training Qwen 3 8B, with near-linear scaling up to 8,192 Ironwood (TPU7x) chips. In one fine-tuning case, step time fell from 11.49 seconds to 2.78 seconds. Google says it continuously tests about 5,000 Hugging Face models.

Ecosystem

The project ships as v0.2.0 on PyPI for Python 3.11 and later under Apache 2.0. Alongside TorchTPU, Google released the vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan repositories. Serving supports FP8, MXFP4 and NVFP4 checkpoints.