Google and Meta open-source TorchTPU, a native PyTorch backend for TPUs

On November 12, 2026, InfoQ reported that Google and Meta have jointly open-sourced TorchTPU, a native PyTorch backend for Google TPU hardware. The project is designed to let existing PyTorch workloads run on TPUs with minimal, low-invasive changes, rather than requiring deep rewrites.

According to the report, migrating a workload from GPU to TPU with TorchTPU involves replacing the cuda device with tpu and swapping the nccl distributed backend for tpu_dist. Developers also get a unified API, torch.accelerator.current_accelerator, to detect the active accelerator, so hardware-specific optimizations can be enabled without rewriting computation logic. The report contrasts this with the path Google has used since 2020 to run PyTorch workloads on TPUs through PyTorch/XLA, which relies on JAX, XLA, and delayed tensors. Google is cited as saying that approach carries clear drawbacks: it requires invasive source-code edits to explicitly mark compilable regions, suffers from heavy compilation latency, does not support PyTorch's native imperative, or eager, execution mode, and has lagged in supporting newer distributed-computing primitives in the PyTorch ecosystem. TorchTPU takes a different approach, treating eager execution as the top priority and dropping explicit compilation regions entirely.

Google states that its hybrid DeferAndFuse mode runs more than 50% faster than Strict Eager. The team has evaluated roughly 5000 Hugging Face models on multiple generations of TPUs, in both eager and compiled modes. The project is now available as an open-source release, with migration and evaluation work being carried out by the Google and Meta teams.