Hugging Face Ecosystem Natively Supports TPUs

On October 20, 2026, Hugging Face announced on its official blog a partnership with TorchTPU to natively support TPUs across its ecosystem via the PyTorch native backend TorchTPU. As a result, the three major libraries Transformers, Diffusers, and Kernels can run directly on TPUs without relying on torch_xla or Optimum TPU. The user experience is simplified: models only need to set device_map="tpu", or use pipe.to("tpu") for a pipeline, to specify the device. This change removes the extra adaptation layer previously required to run on TPUs, and is an update to how the Hugging Face ecosystem supports TPUs. As of now, the official blog is the only source, and no third-party hands-on tests or other follow-up reports have been seen.

AI overview of reporting · Updated 53 min前

Not enough continuous observations to show a trend yet.