Google open-sources TorchTPU, a native PyTorch back end for TPUs
Summary
Google opens TorchTPU, a native PyTorch backend for TPUs, enabling PyTorch models to run on TPU devices with custom ATen kernels and a compilation cache. It includes distributed training support and examples for multiple models, plus a user guide and build/install instructions. This is valuable for ML teams adopting TPU-accelerated PyTorch workflows and for open-source contributors.