Hugging Face ❤️ TorchTPU: TPUs, Natively

Hugging Face has announced native support for Google's Tensor Processing Units (TPUs) through a new PyTorch backend called TorchTPU. This integration simplifies running PyTorch on TPUs by treating them as just another device, removing the need for the more complex torch_xla library. Key Hugging Face libraries like Transformers and Diffusers now have first-class TPU support, allowing models to be loaded with simple commands like `device_map="tpu"`. The new backend also supports `torch.compile` for performance optimization and allows for custom Pallas kernels to accelerate specific operations. To ensure reliability, Hugging Face is also implementing continuous integration testing for models on TPUs.