PANews October 21 news, according to the official Google Cloud blog, Google announced the open-sourcing of TorchTPU at the PyTorch Conference 2026. The project was jointly developed by Google and the Meta PyTorch team, and the code has been released on GitHub. TPU has thus become a PyTorch native hardware option on par with GPU and CPU, and existing code can be migrated with only minor modifications. Frameworks such as vLLM, SGLang, TorchTitan, VeRL, and Ray have already integrated it. In Hugging Face Transformers v5, passing device_map="tpu" allows models to be deployed to TPU.
PANews October 21 news, according to the official Google Cloud blog, Google announced the open-sourcing of TorchTPU at the PyTorch Conference 2026. The project was jointly developed by Google and the Meta PyTorch team, and the code has been released on GitHub. TPU has thus become a PyTorch native hardware option on par with GPU and CPU, and existing code can be migrated with only minor modifications. Frameworks such as vLLM, SGLang, TorchTitan, VeRL, and Ray have already integrated it. In Hugging Face Transformers v5, passing device_map="tpu" allows a model to be deployed to TPU. Tests show that TorchTitan training Qwen 3 8B achieves a maximum MFU (Model FLOPs Utilization) of 52%, and can scale to 8,192 Ironwood TPUs. Google also released llm-d 1.0 with Red Hat and IBM, and LiteRT has also joined the PyTorch ecosystem.