知识卡片
设备选择与精度影响复现
设备优先级解决代码运行位置,不能证明该设备在当前负载下最快。张量与模型应处于兼容设备,TF32 等精度选项也应记入实验条件;浮点差异可能改变接近并列的最大分数,从而改变后续生成。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
Note that if you have a modern NVIDIA GPU (based on the Ampere architecture or newer), the get_device() function automatically enables Tensor Cores for faster matrix multiplications when enable_tensor_cores=True. This can slightly change floating-point rounding but does not noticeably affect results in this book, and the speed advantage on newer cards is worth it. On non-NVIDIA devices, these settings are ignored.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第02章《使用预训练模型生成文本》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-2.html)
- 结论依据:摘录中关于设备选择与精度影响复现的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Note that if you have a modern NVIDIA GPU (based on the Ampere architecture or newer), the get_device() function automatically enables Tensor Cores for faster matrix multiplications when enable_tensor_cores=True. This can slightly change floating-point rounding but does not noticeably affect results in this book, and the speed advantage on newer cards is worth it. On non-NVIDIA devices, these settings are ignored.