Google announced the open-sourcing of TorchTPU at PyTorch Conference 2026, a project jointly developed with Meta's PyTorch team that makes TPU a native hardware option within the PyTorch ecosystem alongside GPU and CPU. The code is now available on GitHub, allowing existing PyTorch workloads to migrate to TPU with minimal modifications.
Major frameworks including vLLM, SGLang, TorchTitan, VeRL, and Ray have already integrated TorchTPU support. In Hugging Face Transformers v5, developers can deploy models to TPU simply by passing `device_map="tpu"`. Benchmark tests showed TorchTitan training Qwen 3 8B achieved a maximum Model FLOPs Utilization (MFU) of 52%, scaling up to 8,192 Ironwood TPUs.
Alongside TorchTPU, Google released llm-d 1.0 in collaboration with Red Hat and IBM, while LiteRT also joined the PyTorch ecosystem, further expanding alternatives to NVIDIA's dominant AI infrastructure stack.
Google Open-Sources TorchTPU at PyTorch Conference, Making TPU Native Hardware Option Alongside GPU
免責事項: Phemexニュースで提供されるコンテンツは、あくまで情報提供を目的としたものであり、第三者の記事から取得した情報の正確性・完全性・信頼性について保証するものではありません。本コンテンツは金融または投資の助言を目的としたものではなく、投資に関する最終判断はご自身での調査と、信頼できる専門家への相談を踏まえて行ってください。