Google has open-sourced TorchTPU, a software stack that lets developers run PyTorch models on the company’s Tensor Processing Units, or TPUs. The release is aimed at developers who already use PyTorch to train and serve AI models.
TorchTPU reaches developers
Google announced TorchTPU on October 20, 2026, at PyTorch Conference 2026. The company says the stack supports training, model serving, and deployment through familiar PyTorch APIs, with minimal changes to existing workloads. Google developed it with the PyTorch team at Meta. (Google Cloud)
The code is available in the public GitHub repository. The repository lists support for Python 3.11 and newer, distributed training, model inference, compilation through torch.compile(), and examples involving models such as Llama, Qwen, ResNet, and UNet. (Google PyTorch)
More PyTorch tools support TPUs
Alongside the core stack, Google says TorchTPU works with tools including vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face, and LiteRT. Google also announced native PyTorch backends for vLLM and SGLang, which are designed to serve models on TPUs. (Google Developers Blog)
TorchTPU makes the TPU a native PyTorch device. Google says code can run eagerly or compile through torch.compile, while distributed communication uses torch.distributed. (Google Developers Blog)
Google targets AI developers
These integrations are intended to make Google’s TPU hardware more accessible to developers who already work in PyTorch. Google says users can choose among workload-optimized CPUs, GPUs, and TPUs while using a consistent software and operational model. (Google Cloud)
The project is available for public inspection and contribution through GitHub, giving developers access to the code and its supported interfaces.
References
Lohmeyer, M., & Jia, B. (2026, October 20). TorchTPU is now open source: Run native PyTorch on TPUs. Google Cloud Blog. https://cloud.google.com/blog/topics/ai-infrastructure/torchtpu-is-now-open-source-run-native-pytorch-on-tpus
Zhou, Q., & Chan, C. (2026, October 20). Announcing new vLLM and SGLang inference engines on TPU. Google Developers Blog. https://developers.googleblog.com/announcing-new-vllm-and-sglang-inference-engines-on-tpu/
Google PyTorch. (2026). torch_tpu [Computer software]. GitHub. https://github.com/google-pytorch/torch_tpu