Hugging Face presents TorchTPU for native PyTorch on TPUs

Hugging Face presents TorchTPU, a new PyTorch backend for TPUs, Google’s AI accelerators. It lets developers use regular PyTorch code and send models and inputs to the “tpu” device, with support for Transformers, Diffusers and Kernels; the package also includes Pallas kernels, distributed training and torch.compile integration.

According to Hugging Face, TorchTPU removes the need for Optimum TPU to run Transformers and Diffusers models on TPUs. Transformers can load models directly onto the chips, while Diffusers offers eager and compiled execution modes, as well as sharding large models across multiple chips through TensorParallelConfig. Installation on a TPU VM requires a single pip command.