> ## Content Index
> Fetch the complete content index at: https://digirolla.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google Open-Sources TorchTPU, Letting PyTorch Run Natively on Its AI Chips
- URL: https://digirolla.com/google-open-sources-torchtpu-letting-pytorch-run-natively-on-its-ai-chips/
- Published: 2026-10-22T07:05:10.000Z
- Updated: 2026-10-22T07:05:10.000Z
- Description: Built with Meta's PyTorch team, TorchTPU lets developers move existing PyTorch code to Google's TPUs by swapping "cuda" for "tpu," a day after Amazon released a similar tool for its Trainium chips.
- Author: Digirolla
- Tags: AI

Google has released **TorchTPU**, its native PyTorch backend for Tensor Processing Units, as an open-source project on GitHub. The company announced it on Tuesday, October 20, at PyTorch Conference 2026.

"With TorchTPU, TPUs join GPUs and CPUs as a native option for PyTorch," Google wrote in its announcement.

## From "cuda" to "tpu"

PyTorch is one of the most widely used toolkits for building, training and running AI models. As TechCrunch explains, many of its features were originally built around Nvidia's GPUs, which have long been the default hardware for AI work.

TorchTPU is designed to change that for Google's chips. According to the PyTorch blog post from the TorchTPU team, the goal is that porting a workload should be as simple as changing `model.to("cuda")` to `model.to("tpu")`. The team notes that this simple change won't reach peak TPU performance, since a model may be tuned for GPUs, but the model should run out of the box.

The project was started at Google in mid-2025 in collaboration with the PyTorch team at Meta. It supports eager mode for debugging, `torch.compile` for production performance and PyTorch's native distributed training tools.

## Built with partners

Google says it worked with several companies so that popular tools work on TPUs:

- **Hugging Face:** In Transformers v5, deploying a model to TPUs is as simple as passing `device_map="tpu"`, according to Google.
- **vLLM and SGLang:** Native TPU backends for the two serving engines, developed with Inferact and RadixArk.
- **TorchTitan, Ray and TorchRec:** Integrations for large-scale training, orchestration and recommendation models.

"We cannot be constrained by a single hardware architecture," said Jana van Greunen, director of PyTorch engineering at Meta, in Google's announcement.

## Google's numbers

The TorchTPU team says it tested 4,917 Hugging Face models and that 98% ran without crashing, while 97% of a set of representative architectures matched GPU results numerically. Google also says TorchTPU training via TorchTitan reached up to 52% model FLOPs utilization on Qwen 3 8B and showed near-linear scaling across up to 8,192 Ironwood TPUs. These are Google's own results.

## More tools in the bundle

Google also announced llm-d 1.0, a distributed inference project released with Red Hat, IBM and a wider community; a TPU agent suite in early preview on GitHub to help developers port and tune models; and LiteRT, its on-device AI runtime formerly known as TensorFlow Lite, joining the PyTorch ecosystem.

## Amazon got there a day earlier

TechCrunch notes that Amazon made a similar move on Monday with **TorchNeuron**, a native PyTorch backend for its Trainium chips. Amazon also said it is contributing code to the main PyTorch project that decouples features from Nvidia hardware. Both companies are trying to make it easier for developers to move AI workloads to their own chips and clouds without rewriting code.

## Sources

- [Google Cloud Blog: TorchTPU is now open source](https://cloud.google.com/blog/topics/ai-infrastructure/torchtpu-is-now-open-source-run-native-pytorch-on-tpus?ref=digirolla.com)
- [PyTorch Blog: Introducing TorchTPU](https://pytorch.org/blog/introducing-torchtpu/?ref=digirolla.com)
- [TechCrunch: Google has joined Amazon in attacking Nvidia's stronghold with AI app developers](https://techcrunch.com/2026/10/20/google-has-joined-amazon-in-attacking-nvidias-stronghold-with-ai-app-developers/?ref=digirolla.com)
- [PyTorch Blog: Introducing TorchNeuron](https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/?ref=digirolla.com)