Rohan Paul· @rohanpaul_ai · X·· 3 小时前AI 评分62
AI 导读
Google 开源 TorchTPU，为 TPU 提供原生 PyTorch 后端，替代此前的 PyTorch/XLA 层。它按写入顺序逐个运行算子以便像 GPU 一样逐步调试，再切换 torch.compile 获得生产速度；训练 Qwen 3 8B 时最高用到 52% 的峰值 TPU 算力。原文还提到 Amazon 也提供原生 PyTorch 后端，让代码脱离 Nvidia GPU 运行。
正文 · 原文
Google and Amazon now both offer native PyTorch backends that let code leave Nvidia GPUs.
Google just open-sourced TorchTPU that replaces PyTorch/XLA, letting developers debug TPU models step by step.
TorchTPU used up to 52% of peak TPU compute training Qwen 3 8B.
TPUs previously reached PyTorch through PyTorch/XLA, which queued operations for a compiler, so debugging differed from GPU work.
TorchTPU replaces that layer, running each operation as written for debugging, then switching to torch.compile for production speed.
来源：Rohan Paul · x.com