Google open-sources TorchTPU to bring native PyTorch support to its TPUs

Google has released TorchTPU, an open-source backend that lets PyTorch run natively on its Cloud TPUs. The company unveiled it on October 20, 2026, at the PyTorch Conference.

TorchTPU is a backend, which is the layer that translates PyTorch’s instructions into something a specific piece of hardware can execute.

The project was built jointly with Meta’s PyTorch team. Work on it reportedly started in mid-2025.

Existing workloads can move over with minimal adjustments, the most notable being a swap of the device call from .to(“cuda”) to .to(“tpu”).

TorchTPU also supports eager execution, which is how PyTorch runs by default. In eager mode, code runs line by line as it is written, which makes debugging far less painful. Google is pitching improved debugging as one of the backend’s main advantages.

Under the hood, the release supports approximately 450 ATen operations. ATen is PyTorch’s core library of tensor operations, the basic math building blocks that models are assembled from.

The backend is designed to stretch across thousands of TPU chips. According to the reported figures, TorchTPU hit 52% model FLOPs utilization when training the Qwen 3 8B model on up to 8,192 TPU v7x chips.

The release spans four GitHub repositories under the Apache 2.0 license, a permissive license that allows commercial use. It is compatible with Python 3.11 and newer.

It also plugs into Hugging Face Transformers. Google says thousands of Hugging Face models can run with minimal code changes.

Google says some of its Cloud customers have already started adopting the backend. The company is positioning TorchTPU as production-ready, not an experimental side project.

Before TorchTPU, the main route for running PyTorch on TPUs was PyTorch/XLA. That approach lacked the direct integration with TPU functionality that developers wanted. Google now frames TorchTPU as the production-ready replacement for PyTorch/XLA.