Google published TorchTPU’s code Oct. 20, making its Tensor Processing Units easier to try with existing PyTorch software as cloud providers compete to attract AI workloads.
Google introduced TorchTPU on April 7 and listed a public code repository among its goals for 2026. The company says developers can move PyTorch workloads to TPUs without extensive rewrites, although some models may need design changes or custom kernels to get the most from the chips. Developers use the open-source PyTorch framework to build, train and run AI models.
Google built TorchTPU with Meta’s PyTorch team. It supports eager mode for debugging models and torch.compile for optimizing them before deployment, giving developers access to tools they already use. Google said its earlier PyTorch/XLA approach could require separate code for distributed workloads.
Meta Director of PyTorch Engineering Jana van Greunen said in Google’s announcement that TorchTPU lets Meta run PyTorch workloads on TPUs and provide feedback to the engineering teams.
Google also announced support for Hugging Face model libraries and for vLLM and SGLang, tools used to serve AI models. Anyscale and Google Cloud introduced native PyTorch support for Ray on TPUs on Oct. 20. Ray manages workers used for training and model serving; TorchTPU handles PyTorch computations on the chips.
The release coincided with Amazon Web Services’ Oct. 20 public beta of TorchNeuron, an open-source backend for running PyTorch code on Trainium2 and Trainium3. AWS had announced its TorchNeuron beta plans a day earlier. Both projects address the engineering work involved in trying PyTorch programs on the companies’ AI processors.
Google also released its TPU agent suite in early preview. The tools are intended to help developers set up workloads, diagnose problems and tune performance on TPUs.
