Google open-sources TorchTPU for native PyTorch workloads on TPUs

Google released TorchTPU as open source on GitHub, saying the stack lets developers train and serve existing PyTorch models on its TPU accelerators with minimal code changes. Co-developed with Meta's PyTorch team, it supports familiar PyTorch interfaces and integrates with serving and training tools including vLLM, SGLang and TorchTitan. Developers can debug models interactively before using torch.compile to optimize execution. Hugging Face Transformers v5 adds TPU deployment through its device selection setting, while Diffusers supports distributing models across accelerators. The accompanying VeRL reinforcement-learning integration remains in preview.