Google launches TorchTPU, a native pytorch backend for cloud tpus

Google has launched TorchTPU, a native PyTorch backend for Cloud TPUs, developed in collaboration with Meta's PyTorch team and announced at the PyTorch Conference 2026 on October 20, 2026. The tool replaces PyTorch/XLA and is recommended for all new workloads. TorchTPU allows PyTorch models to run natively on TPUs, with tensors functioning as standard PyTorch tensors without subclasses or wrappers. It supports one process per device and preserves tensor views and in-place mutations. TorchTPU introduces three eager modes, with the peak speed mode using torch.compile(backend="tpu") to capture graphs with TorchDynamo and lower them through StableHLO to OpenXLA, emitting TPU machine code. It also supports custom Pallas kernels and prebuilt Tokamax and Helion tile kernels. The distribution framework uses the tpu_dist process group, supporting DDP, FSDP2, and DTensor with topology-aware mesh for TPU interconnects.