Hugging Face libraries now run natively on Google TPUs through TorchTPU

Hugging Face says TorchTPU, a native PyTorch backend for Google’s TPUs, now lets Transformers, Diffusers and Kernels run on TPU without porting; loading a model on a TPU becomes a matter of passing the device name. The older route, PyTorch/XLA, worked but came with lazy tensors, surprise recompilations and hard debugging, and Hugging Face says its own Optimum TPU layer is no longer required for these libraries. Google has designed TPUs for about a decade and uses them to train and serve its models, including Gemini, but the TPU software stack grew up around TensorFlow and JAX, leaving PyTorch users on a harder path.

Early numbers are encouraging and candid about gaps. Of the Transformers tests that run on TPU, 96.5% pass. Hugging Face has filed 34 issues against TorchTPU, 19 of them fixed, and is still working on a nightly TPU test run. A ragged-dot kernel from the Hub made batch-64 decoding 21.5x faster on Qwen3-30B-A3B and 12.8x faster on Gemma 4 26B-A4B, compared with the default implementation on a TPU v6e-8.

Hardware competition in AI has been held back less by chips than by software. Most models and code are written in PyTorch, which has felt most at home on NVIDIA GPUs, and when another accelerator requires rewriting code or learning JAX, teams stay put. Making the TPU just another device attacks that friction. It arrives alongside Amazon’s TorchNeuron effort for its own chips (TorchNeuron), and we covered how Google and Amazon are both pressing on NVIDIA’s developer stronghold (Google joins Amazon). Hugging Face is a natural partner because its libraries sit between most developers and the models they use, though NVIDIA has also agreed to buy Hugging Face (Nvidia and Hugging Face).

More real choice in accelerators puts pressure on prices and supply and gives researchers and startups more places to find compute.

TPUs are available only on Google Cloud, so this widens choice among chips while deepening reliance on one cloud. A 96.5% pass rate is good for early work but means some things will break, which Hugging Face says plainly. The headline speedups compare a tuned kernel with the default TPU path, not TPU with GPU, so they say nothing yet about which hardware is cheaper for a given job. Key signals are pass rates climbing toward parity, independent cost comparisons, and similar native backends for other non-NVIDIA accelerators. More in our Google coverage, Nvidia coverage and tools coverage.