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

The project is an open source repository under the google-pytorch organization, and the team says it is already in production for some of Google Cloud's largest customers across inference and training. The backend supports eager and compiled execution modes and covers about 450 core ATen operators, including forward and backward passes. TPU v8i support is included at launch. The release arrives one day after Amazon opened TorchNeuron for its Trainium chips, another effort aimed at running PyTorch workloads on Amazon's own accelerators. Both moves target the dependence on Nvidia's CUDA stack that has kept many PyTorch teams on one vendor's hardware. For founders, the practical question is whether training and serving code can move between accelerators without a rewrite. If the TPU backend holds up on the models a team already runs, TPU capacity becomes easier to price against Nvidia GPUs and gives buyers more leverage in negotiations.