Google releases open-source TorchTPU for PyTorch on TPUs

Google released TorchTPU as an open-source backend on October 20, giving developers a way to run PyTorch workloads on the company’s Tensor Processing Units (TPUs). The project is intended to make it easier to move existing applications from graphics processing units (GPUs) to Google’s AI chips without extensive code changes. TorchTPU was announced at the PyTorch Conference. Google says developers can transfer workloads with minimal adjustments, including changing the device setting that directs a model to a particular chip. The company describes the tool as a way to bring PyTorch applications to TPUs and Google Cloud without rewriting them. The extent to which that ease of migration holds across different applications has not been independently established in the supplied reporting. Google said it worked with Meta, Hugging Face, Inferact and RadixArk to support compatibility with their developer tools. The release also includes integrations with software used for model training and serving, including TorchTitan, vLLM and SGLang. Google announced a TPU agent suite to assist developers with migration and troubleshooting; one report described the suite as still in preview. Google...