Google and Meta Open-Source Native PyTorch Support for TPUs

Published 21 October 2026

TorchTPU brings eager and compiled PyTorch execution to Google TPUs, enabling near drop-in migration from GPUs, distributed training, custom kernels, and integrations with tools including vLLM, SGLang, TorchTitan, Ray, and TorchRec. The open-source project reports broad Hugging Face model compatibility and competitive inference and training performance.