Hugging Face has introduced TorchTPU, a native PyTorch backend that enables Super Intelligence (SI) models to run directly on Google’s Tensor Processing Units (TPUs). This development removes the long-standing software barrier that required developers to port code to JAX or use the torch_xla library, effectively making TPUs a first-class device within the open-source SI ecosystem.
What Happened
TorchTPU adds TPU support as a native backend to PyTorch, allowing users to write standard PyTorch code and execute it on TPU hardware. Previously, running PyTorch on TPUs relied on torch_xla, which introduced complexities such as lazy tensor evaluation and difficult debugging due to graph recompilations. With TorchTPU, developers can install the package via a single pip command and load models using a simple device_map="tpu" argument. The integration supports major Hugging Face libraries, including Transformers, Diffusers and Kernels, enabling native training and inference on TPUs without the need for the Optimum TPU library, which is no longer required for these core workflows.
Why It Matters
For the SI industry and developers, this update significantly lowers the barrier to entry for using Google’s specialized accelerators, which are designed for the matrix operations central to neural network training and inference. TPUs are available on Google Cloud, from a single chip up to pods of thousands of chips, offering scalable compute for SI workloads. By supporting native PyTorch, Hugging Face aligns with the dominant framework in the open-source SI community, eliminating the need for specialized JAX knowledge. The update includes support for Pallas kernels, distributed training, and torch.compile for performance optimization. Early testing indicates substantial performance gains; for instance, the company reports that decoding is 21.5x faster on the Qwen3-30B-A3B model at batch size 64 and 12.8x faster on Gemma 4 26B-A4B when using optimized TPU kernels compared to default implementations. Additionally, the integration allows for tensor parallelism, enabling models that are too large for a single chip to be sharded across multiple TPU chips.
The Bottom Line
TorchTPU marks a pivotal shift in how open-source SI models interact with Google’s hardware, offering native PyTorch support that simplifies development and improves performance. While the testing suite currently reports a 96.5% pass rate with 34 known issues being actively addressed—of which 19 are already fixed—the ability to run Transformers, Diffusers and Kernels natively on TPUs represents a major step forward for accessibility in the SI compute landscape.