Learn about this first
The PyTorch teams at Google and Meta jointly developed and open-sourced TorchTPU, providing a native PyTorch backend for Google Cloud TPU, as announced at PyTorch Conference 2026. The PyTorch blog says models can run simply by changing .to("cuda") to .to("tpu"), with no JAX dependency required; MarkTechPost says it replaces PyTorch/XLA, and Google recommends it for all new workloads.
Hugging Face announced native TPU support through TorchTPU. Transformers, Diffusers, and Kernels can run natively, and loading a model only requires device_map="tpu"; Optimum TPU is no longer needed. To install it on a TPU VM, run pip install torch-tpu --extra-index-url https://download.pytorch.org/whl/cpu.
TorchTPU is v0.2.0 on PyPI, requires Python 3.11+, uses the Apache 2.0 license, and runs on Google Cloud TPU. Testing covers chips ranging from v5e to 8,192 Ironwood (TPU7x) chips. Rohan Paul said that training Qwen 3 8B used up to 52% of peak TPU compute.
AI-generated from reports · updated 26 days agoLatest developments
Google open-sources TorchTPU, providing a native PyTorch backend for TPUs
Event progress2
TorchTPU open-sourced
PyTorch Blog: Google and Meta release TorchTPU: a native TPU backend for PyTorchHugging Face announces native TorchTPU support
Official Hugging Face team blog: The Hugging Face ecosystem adds native TPU support through TorchTPU