# Google and Meta open-source TorchTPU, a native PyTorch backend for TPUs

published: 2026-10-25T18:05:39.222Z
topic: infrastructure
source: https://www.infoq.com/news/2026/10/google-meta-torchtpu/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
url: https://aiagentnews.io/stories/google-and-meta-open-source-torchtpu-a-native-pytorch-backend-for-tpus

Google and Meta have open-sourced TorchTPU, a native PyTorch backend for Google's Tensor Processing Units, announced at PyTorch Conference North America 2026 in San Jose. The backend replaces PyTorch/XLA and aims to let developers move a model to TPUs with a one-line device change, though reaching peak performance still requires tuning.

At PyTorch Conference North America 2026 in San Jose, Google and Meta released four repositories: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan. TorchTPU replaces PyTorch/XLA, the lazy-tensor bridge from 2020, and the stated goal is to make porting a model as simple as changing model.to("cuda") to model.to("tpu"). Meta, which created PyTorch, plans to move larger-scale LLM workloads onto TPUs with it.

The speakers said Hugging Face models work with the backend out of the box, and the vLLM and SGLang inference engines run natively. Bill Jia showed near-linear scaling past 8,000 TPU chips. Google's published results include a Qwen3 235B mixture-of-experts run on 8,192 TPU v7x chips at 28% model FLOPs utilization.

The announcement says the one-line change is only a starting point, because model shapes, sharding and ops may be tuned for GPU architectures. Google also previewed a TPU agent suite on GitHub that plugs into coding harnesses to automate onboarding, diagnostics and performance tuning.

## Why it matters if you build agents

Teams running PyTorch agent models can try TPUs with a small code change, but should budget for sharding and kernel work before expecting production-level performance. AWS put a similar PyTorch backend for Trainium, TorchNeuron, into public beta on October 20, and according to The Next Web it does not yet support dynamic input shapes. Confirm TPU capacity before planning a move.