Google Opens TorchTPU, a Native PyTorch Backend for Cloud TPU

Google TorchTPU has been available as an open-source project since October 20, 2026. It is a native backend for the PyTorch framework on Google Cloud TPU accelerators. The public google-pytorch/torch_tpu repository uses the Apache 2.0 license, and Google created the project in collaboration with Meta’s PyTorch team.

The release is intended to allow developers to use PyTorch on Cloud TPU directly through a native backend. In its official documentation, Google recommends TorchTPU for new workloads and identifies it as a replacement for the previous PyTorch/XLA project.

Google TorchTPU Brings a Backend for Cloud TPU

TorchTPU adds support for TPU devices to PyTorch and, according to the documentation, enables them to be used through torch.device("tpu"). The project includes ATen kernels, a compilation cache, distributed training, and integration with torch.compile.

Integration with torch.compile is intended to make TorchTPU part of the standard way PyTorch programs are optimized.

The project is publicly available on GitHub under the google-pytorch organization. The repository also lists version 0.2.0 on PyPI, meaning the package can be distributed through the standard Python ecosystem channel.

A Replacement for PyTorch/XLA

Until now, the way to use PyTorch on TPU was PyTorch/XLA, a layer connecting PyTorch with the XLA compiler. Google now identifies TorchTPU in its documentation as a solution that replaces PyTorch/XLA for new deployments.

For teams using PyTorch, the important point is that Google presents the new backend as a native part of working with the framework, rather than as a separate integration layer. The goal is to make training and inference on Cloud TPU available without rewriting projects in JAX or using the older PyTorch/XLA.

The PyTorch Foundation separately described TorchTPU as an open-source collaboration between Google and Meta and as a native backend for Google TPU. This collaboration is particularly relevant because PyTorch is a major ecosystem for developing and operating artificial intelligence models.

Significance for Tools in the PyTorch Ecosystem

Native TPU support could expand hardware options for developers who already build on PyTorch. A factual overview of the announcement cites potentially better interoperability with tools and projects such as vLLM, SGLang, Hugging Face, and TorchTitan.

The Apache 2.0 license is a permissive open-source license, and the repository is publicly available.

Real-World Use Will Test Performance and Compatibility

Google cites benefits related to using TPU in a PyTorch environment, but specific claims about speed, costs, or ease of migration have not yet been independently benchmarked in the available materials. Long-term compatibility with all existing PyTorch models and libraries must also still be confirmed.

When deploying it, it will therefore be important to monitor independent measurements of performance, stability, and compatibility. The pace at which existing projects migrate from PyTorch/XLA and the extent of support in tools intended for inference and distributed training also warrant attention.

Verified and updated: October 21, 2026, 06:20