Google opens up TorchTPU, so PyTorch code can run on its AI chips with little rewriting.

Google announced it at PyTorch Conference on Tuesday and built it with Meta's PyTorch team. Hugging Face Transformers v5 can now send a model to a TPU with device_map="tpu". vLLM, SGLang, Ray and TorchTitan all work with it. Google also released a "TPU agent suite" in early preview, which is a set of coding-agent plug-ins that help move projects onto TPUs. And it brought LiteRT, its tool for running models on phones, laptops and browsers, into the PyTorch ecosystem. This comes a day after Amazon's TorchNeuron, which we covered yesterday. What this means for you: TPUs are rented from Google Cloud, so this only matters if you pay for cloud compute. If you do, it's now easy to test Google's or Amazon's chips against Nvidia on your real workload. Google didn't announce any TPU price changes. For hobbyists, the useful piece is LiteRT-Torch, which gets a PyTorch model running on a phone or laptop. Sources: Google Cloud, TechCrunch.