25 lines
1.3 KiB
TeX
25 lines
1.3 KiB
TeX
\begin{figure}[t]
|
||
% \captionsetup[subfigure]{labelformat=simple, labelsep=period}
|
||
% \renewcommand\thesubfigure{\alph{subfigure})} % 将子标题的标签格式改为 "a)"
|
||
\centering
|
||
\begin{subfigure}[b]{0.47\linewidth} % PD:平衡一下图片大小，如果一样的图可以都用0.48
|
||
\hspace{-3px} % PD: 往左挪点防止重心偏右
|
||
\includegraphics[width=\linewidth]{assets/Layer10.pdf}
|
||
% \includegraphics[width=\linewidth]{assets/Layer10_norubost.pdf}
|
||
\caption{Across Head Dimensions} % 子图标题留空即可自动生成 (a)
|
||
\label{fig:attnindim}
|
||
\end{subfigure}
|
||
\hfill % 在两张图之间插入一个弹性空白，使它们左右对齐
|
||
\begin{subfigure}[b]{0.48\linewidth}
|
||
\hspace{-3px} % PD: 往左挪点防止重心偏右
|
||
\includegraphics[width=\linewidth]{assets/AcrossLayer.pdf}
|
||
\caption{Across Layers}
|
||
\label{fig:attninlayer}
|
||
\end{subfigure}
|
||
\caption{Q-state activation strength visualizations in LLaMA-2-7B.
|
||
We compute the average L2 norm per attention head to quantify activation strength.
|
||
Stronger activations are concentrated in high-indexed (\ie low-RoPE frequency) dimensions and vary across layers, highlighting both dimension-wise and layer-wise heterogeneity.
|
||
}
|
||
\label{fig:hotattn}
|
||
\end{figure}
|
||
% \py{font size of figure is too small} |