% Golden example -- a scale rewrite that must be seen, not just aligned. % ../scripts/build.sh rewrite-cancel.tex \documentclass[border=10pt]{standalone} \usepackage[cjk]{superderive} \begin{document} \begin{tikzpicture} \sdstage{D1}{缩放来自方差,不是装饰} \sdrow{R1} \sdstep[role=rewrite]{s1}{QK^{\top}}{QK^{\top}/\sqrt{d_k}} \sdreason{s1}{§3.2.1,避免 softmax 饱和} \sdrowend \sdbbox{all} \sdtopformula{F}{$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(QK^{\top}/\sqrt{d_k})V$} \sdmeaningbox{mb}{120mm}{all} {点积方差随 $d_k$ 线性涨} {缩放是改写,不是新算子} {没有缩放,softmax 进饱和区,梯度消失} \end{tikzpicture} \end{document}