Google Open-Sources TorchTPU, a Native PyTorch Backend for TPUs Co-Developed With Meta
Google made TorchTPU, its native PyTorch stack for TPUs, open source on Oct. 20 and says it was co-developed with Meta's PyTorch team. A Meta engineering director says it lets Meta deploy PyTorch workloads to Google TPUs. Coverage, performance and cost figures are Google's own or, for one benchmark, Google citing SemiAnalysis.
For PyTorch users, the cost of leaving GPUs has been the rewrite, and Google is now pitching TPUs as a device-string change, with Meta's engineers on the project. On Tuesday, October 20, at PyTorch Conference 2026, Google said TorchTPU, its native PyTorch stack for its TPU chips, is now open source on GitHub. Google says it was "co-developed with the PyTorch team at Meta."
What Google and Meta say it does
Google says TorchTPU runs existing PyTorch workloads on TPUs "with minimal code changes", using eager mode for debugging and torch.compile for production performance. The TorchTPU team's post on the PyTorch blog gives the example of changing .to("cuda") to .to("tpu"). It says the effort started at Google in mid-2025 with Meta's PyTorch team and was built as a native PyTorch backend, in contrast with the earlier PyTorch/XLA, which the team says required intrusive source-code changes and lacked eager execution. The team also says that simple swap will not reach peak TPU performance, because a model's shapes, sharding and operations may be tuned for GPUs, but that the model should run.
Google's post quotes Jana van Greunen, Meta's Director of PyTorch Engineering, saying TorchTPU supports "major LLM, Ads, and Recommendation systems workloads on key production use cases" and "enables us to deploy our PyTorch workloads to Google TPUs." Neither post discloses a TPU purchase or rental agreement, its size or its value. The PyTorch blog says TorchTPU is already used in production for inference and training by "some of the largest Google Cloud customers", without naming them.
The serving and training stack around it
Google announced partner integrations alongside the release:
- A native PyTorch vLLM backend on TorchTPU, built with Inferact, and an SGLang backend, built with RadixArk.
- llm-d 1.0, released with Red Hat, IBM and what Google calls a community of more than 500 contributors and 25 organizations, for routing inference across TPU and GPU fleets on Google Kubernetes Engine.
- TorchTPU integrations for TorchTitan training, a TPU backend for the VeRL reinforcement-learning framework in preview, and native support in Ray.
- Hugging Face Transformers v5 support for deploying models to TPUs with
device_map="tpu". - LiteRT, Google's on-device inference runtime formerly called TensorFlow Lite, joining the PyTorch ecosystem.
The figures are the developers' own
The PyTorch blog says TorchTPU supports about 450 ATen operators. In the team's own tests, 4,821 of 4,917 Hugging Face models (98%) ran without crashing, and 638 of 657 architecture-representative models (97%) matched GPU numerics within the team's thresholds. Citing SemiAnalysis's InferenceX benchmark, the TorchTPU team says TPU v7 with the vLLM-TorchTPU stack reached 2.01 cents per million tokens on the 8K/1K test, down from 3.6 cents at its September debut. We did not open SemiAnalysis's results.
The release came the same day Amazon's TorchNeuron, a native PyTorch backend for Trainium, entered public beta, which we covered separately. No benchmark we have seen compares the two.
Limits
- Model-coverage, performance and cost figures come from Google and the TorchTPU team; the InferenceX figure is Google's citation of SemiAnalysis, not a result we checked.
- The scale of Meta's TPU use is described only in the quote above; no contract, volume or dollar figure is public.
- Neither post gives a time of day for the announcement.