TorchTPU: Native PyTorch backend for Google TPUs
In brief
- TorchTPU enables native PyTorch execution on Google’s TPUs with minimal code changes.
- The integration supports eager mode and replaces PyTorch/XLA with improved performance.
- It introduces a compilation cache and bfloat16 optimizations for TPU hardware.
Google and Meta unveil TorchTPU, an open-source backend enabling native PyTorch execution on Google’s TPUs. According to the announcement, the integration allows seamless model execution and TPU migration with industry-competitive training and inference performance.
Developed in mid-2025 through a Google-Meta collaboration, TorchTPU replaces the earlier PyTorch/XLA system, which relied on lazy tensors and faced compilation bottlenecks. Unlike its predecessors, TorchTPU natively supports PyTorch’s eager execution mode, requiring only minor code changes—such as switching devices from .to("cuda") to .to("tpu").
The solution targets large Google Cloud customers already leveraging TPUs for massive workloads. It builds on XLA architecture and TPU hardware, ASICs optimized for tensor operations. TorchTPU also introduces bfloat16 precision optimizations tailored for TPUs and a compilation cache to speed up repeated executions.
This summary gives you the essentials. The full article, its figures and quotes are on PyTorch Blog.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error