Hugging Face ❤️ TorchTPU: TPUs, natively
Hugging Face announces that TorchTPU, a native PyTorch backend, lets its Transformers, Diffusers and Kernels libraries run directly on Google TPUs without the older PyTorch/XLA workarounds. It reports 96.5% of library tests run on TPU pass, and 34 issues reported to the project, 19 of them fixed.
What the article says
On October 20, 2026, Hugging Face introduced TorchTPU, a native PyTorch backend that treats Google's TPUs as just another device. Transformers, Diffusers and Kernels now run directly on TPU: a user loads a model from the Hub and moves it to the "tpu" device.
Until now, PyTorch users had to go through PyTorch/XLA, whose lazy execution and recompilations could be surprising. TorchTPU brings tuned Pallas kernels, distributed training and a torch.compile integration. Benchmarks on a Cloud TPU v6e-4 show, for example, Qwen-Image-2.1 at 3.5 s compiled versus 8.0 s in eager mode, and decoding up to 21.5 times faster on a mixture-of-experts model with TPU kernels. Hugging Face says 96.5% of the Transformers tests that run on TPU pass, that the work is in progress, and that a nightly CI is planned.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.