谷歌联手Meta奇袭英伟达，开源TorchTPU，给PyTorch代码搬家

Zhidixi (Public account: zhidxcom) Compiler | Yang Jingli Editor | Li Shuiqing Zhidixi reported on October 21 that today, Google announced the open source TorchTPU, which is a native TPU software stack jointly developed by Google and the Meta PyTorch team. Developers can deploy existing PyTorch models to Google TPU with less code changes. According to TechCrunch, most AI applications currently run on NVIDIA GPUs. Google's open source may be an attempt to break NVIDIA's dominance in the AI ​​industry. PyTorch is an open source deep learning framework for developing and training AI models. It was developed by the Facebook AI Research (now Meta AI) team in 2016 and officially released as open source in 2017. Today, PyTorch has become one of the most widely used frameworks in global AI research and application development, covering areas such as large language models, computer vision, and generative AI. Around TorchTPU, Google has also connected TPU to tools such as vLLM, SGLang, llm-d, TorchTitan, VeRL, Ray, Hugging Face, and LiteRT, covering scenarios such as large model training, inference, reinforcement learning, cluster scheduling, and device-side deployment. Google said that the training solution based on TorchTPU has been verified on up to 8,000 Ironwood TPUs, achieving a maximum model floating-point operation utilization of 52% in the Qwen3-8B training test; on the inference side, the Indian AI model company Zenteiq.ai increased the overall inference throughput by more than 4 times after using the relevant solution. For developers, existing PyTorch models can be more easily migrated to TPU, and commonly used training, inference, and task scheduling tools can continue to be used. Developers can also flexibly choose between TPU, GPU and CPU based on model size, running speed and cost. Open source address: ht