Google and Meta Release TorchTPU to Run PyTorch Natively on TPUs, Deprecate PyTorch/XLA

Google and Meta have formalized a significant shift in how developers run PyTorch on Google’s TPU accelerators, pairing a new research paper with an open-source release and a clear replacement plan for older tooling. An arXiv paper posted in the cs.AI category describes TorchTPU as “an open-source collaboration between Google and Meta” and “a native and principled PyTorch backend for Google TPUs,” while Google has also released the software publicly on GitHub. At the same time, Google’s older PyTorch/XLA path is now explicitly marked for deprecation.

That matters because PyTorch, the machine-learning framework originally developed at Meta, has been one of the main ways developers build AI models, while Google’s TPUs are widely used AI chips in Google Cloud. Historically, running PyTorch workloads on TPUs meant going through PyTorch/XLA, a bridge layer that has been in use since around 2020. TorchTPU is meant to reduce that friction by making TPUs look and behave more like a native PyTorch target rather than a separate workflow.

The rollout has unfolded over several months. Google outlined the approach in an engineering post on April 7, saying, “We established an ‘Eager First’ philosophy.” A packaged release, torch_tpu 0.2.0, was uploaded to PyPI on Oct. 19. Google Cloud then announced the open-source release on Oct. 20, saying TorchTPU was available on GitHub in the google-pytorch/torch_tpu repository under the Apache 2.0 license. The corresponding arXiv paper, “TorchTPU - A Hybrid Eager/Compiled Architecture for Native PyTorch Execution on TPU,” is listed as arXiv:2610.24161, with a first version posted Oct. 16 and the latest revision, v3, dated Oct. 22.

In practical terms, the project documentation says TorchTPU exposes TPUs as a standard PyTorch device through torch.device("tpu"). It supports eager execution — meaning code runs in a more direct, interactive way familiar to many PyTorch users — as well as torch.compile, PyTorch’s compilation path for optimizing execution. Google’s earlier engineering write-up said the compiled path lowers through StableHLO to XLA, but the larger point for developers is simpler: TorchTPU is designed to let standard PyTorch workloads run on TPUs with fewer TPU-specific adjustments. The documentation also points to a compilation cache and support for distributed training.

The clearest sign that this is not just a research project is what comes next for PyTorch/XLA. The PyTorch/XLA repository README now says: “Deprecated: This project is deprecated and replaced by TorchTPU. The PyTorch/XLA repo will be archived on November 15, 2026. Please start new PyTorch TPU workloads with TorchTPU and migrate existing PyTorch/XLA workloads to TorchTPU.” For current TPU users, that makes TorchTPU the new default path, not just an alternative.

Google is also framing the effort as ecosystem work, not a standalone backend. Its announcement highlights collaboration with Meta’s PyTorch team and support or integration efforts involving projects such as Transformers, Diffusers, vLLM and SGLang. In Google Cloud’s Oct. 20 post, Meta engineer Jana van Greunen said TorchTPU “enables us to deploy our PyTorch workloads to Google TPUs, provide direct feedback to the engineering teams, and build the ecosystem flexibility everyone needs.” For now, however, readers should treat any performance claims cautiously: the available claims come from project materials and Google posts, not independent third-party benchmarks.