Google and Meta open-source TorchTPU for PyTorch on TPUs

Google and Meta have open-sourced TorchTPU, a native PyTorch backend that allows machine learning code to run directly on Google Cloud TPUs as a standard device alongside GPUs and CPUs. TorchTPU replaces an older tool called PyTorch/XLA and was announced at PyTorch Conference 2026 on October 20, 2026. The integration enables popular libraries like Transformers, Diffusers, vLLM, and SGLang to work natively on TPUs without requiring separate forks or workarounds.

• Hugging Face ↗: The tighter integration could lower barriers for researchers training large models by allowing them to load models from Hugging Face's repository and send them to TPU with simple code commands.
• MarkTechPost ↗: TorchTPU offers multiple execution modes including eager mode for debugging and compiled mode for production speed, with reported performance reaching 52% model FLOPs utilization on a training task, and has been tested to scale across thousands of TPU chips.

TorchTPU simplifies how researchers and developers can use TPUs for training and running large language models by using familiar PyTorch code without requiring specialized compilation steps or rewrites. Previously, running PyTorch on TPUs required using an intermediary tool that had limitations like lazy tensor evaluation and difficult debugging.