Google releases open-source bridge for PyTorch on TPUs
Google open-sourced TorchTPU, giving PyTorch developers a native path to run models on Tensor Processing Units with minimal code changes. Google also added TPU support to vLLM, SGLang, TorchTitan, Ray, and Hugging Face libraries. The company reported near-linear scaling across as many as 8,192 Ironwood TPUs and 52% model FLOPs utilization on Qwen 3 8B.
TorchTPU lets PyTorch code target Google’s TPUs with few edits, preserving eager debugging and torch.compile graph execution. Support also reaches serving stacks vLLM and SGLang, training framework TorchTitan, Ray, and Hugging Face libraries; Transformers v5 can route models via device_map="tpu", while Diffusers gains multi-chip tensor parallelism.
Google measured throughput rising almost proportionally up to 8,192 Ironwood chips and 52% model FLOPs utilization on Qwen 3 8B. Meta has deployed it in live work involving language models, advertising, and recommendations. Separately, llm-d 1.0, released with Red Hat and IBM, helped Zenteiq raise token output roughly fourfold per 1,000 TPU v6e chips. VeRL’s TPU backend and a TPU agent suite remain in preview.
The bridge could lower the cost and effort for teams to try non-GPU hardware, giving PyTorch developers and smaller AI groups more options. If workloads move easily, cloud providers and chip suppliers may face sharper competition, potentially affecting pricing and access to AI compute. Broader effects depend on real-world performance, tool maturity, and whether preview features become stable; users may benefit from faster experimentation, while reliance on a few vendors could still shape costs and availability.