Google releases TorchTPU: PyTorch programs can run on its TPU chips with almost no code rewriting
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Running the same model on chips other than Nvidia's is getting easier, and developers gain a choice.
TorchTPU is an open-source native PyTorch backend in which the TPU works as an ordinary device, without workarounds through torch_xla or a move to JAX; the code is on GitHub. According to its blog, Hugging Face added TPU support to Transformers v5 and Diffusers (for example, device_map="tpu"); the package includes ready-made Pallas kernels, distributed training and torch.compile. Hugging Face's blog gives one measurement: a ragged-dot kernel sped up Qwen3-30B decoding at batch size 64 by 21.5 times. According to TechCrunch, Google worked with Meta, Hugging Face, Inferact and RadixArk, and a day earlier Amazon released a similar TorchNeuron for its Trainium chips; by the outlet's estimate about 85% of SI applications use PyTorch, which tied developers to Nvidia's CUDA.
Sources 2