Google and Meta release TorchTPU to operate PyTorch on TPU processors
A change of one line of code can move a PyTorch model to Google TPU processors, but the model must have tuning to be fast.
On 21 October 2026, Google and Meta released TorchTPU as open source. It lets PyTorch, the AI software that Meta made in 2016, operate on TPUs, the AI processors of Google. It replaces a layer from 2020 that made large changes to the source code necessary. Meta will move large language model work onto TPUs with it.
How it works
TorchTPU connects to the low-level dispatcher of PyTorch. The speakers tell that Hugging Face models operated with it with no change. The vLLM and SGLang engines, which give model results to users, also operate with it.
The numbers
Bill Jia of Google showed near-linear scaling through 8,000 TPU processors. Google also gives a test of the Qwen3 235B model on 8,192 processors at 28% FLOPs use. This measures how much of the maximum capacity the test uses.
What is not known
The change of one line makes a model operate, not operate fast. Teams must have time for sharding and kernel tuning work. Google gave no test results for its new TPU agent suite, which is in preview. On Hacker News, users wanted to know where to get a TPU.
Insight: Cloud firms race to make their chips native in PyTorch
One day before this release, AWS put TorchNeuron, a PyTorch backend for its Trainium processors, into beta. Analyst Blair Hayes tells that lock-in has moved and does not end. It looks as if each cloud wants to make its processor a usual PyTorch device. Then software is not a cause to stay with Nvidia. Buyers get more selection, but tuning tools can become the new lock-in.
Sources
Posted
