xla

Why we included this project

PyTorch/XLA is the main way to run PyTorch models on Google Cloud TPUs. It sends tensor operations through the XLA compiler, which fuses and optimizes the computation graph so the TPU's hardware is actually put to work, not treated as a slow compatibility layer. The package installs with pip on TPU VMs, supports distributed training across chips via the PJRT runtime, and has handled demanding workloads like Stable Diffusion. Since TPUs can be far cheaper per flop than GPUs for many training and inference jobs, this is the practical route for teams that want to keep their existing PyTorch code while moving to TPU infrastructure.