Google Releases TorchTPU for Native PyTorch Training on TPUs

On October 21, 2026, the Google Developers Blog published an introduction to TorchTPU, describing it as a way to bring a native torch.device("tpu") backend directly into PyTorch. According to the post, developers can train large-scale models on Google Cloud TPUs without rewriting their code in JAX. The announcement also points to a companion TorchTitan TPU fork, which adds production-grade scaling primitives including SimpleFSDP, Graph Trainer, and custom Pallas kernels. Google says these components achieve high Model Flops Utilization (MFU) across both dense and MoE architectures, and that developers can draw on battle-tested optimization playbooks and recipes as well as Hugging Face material. The post frames TorchTitan as the PyTorch ecosystem's clean-room reference platform for large-scale generative AI training, built entirely on PyTorch-native scaling primitives and presented as a modular showcase for pre-training work. Beyond this introductory post, the available material does not specify a release date, regional availability, or benchmark comparisons against other backends, so the announcement is where coverage currently stands.