Transformer 注意力机制图解

一张精准的瑞士网格技术图，在白色画布上展示 Transformer 块内的缩放点积注意力机制。核心：三列标注 "Q (Query)"、"K (Key)"、"V (Value)" 的海军蓝垂直列，向下流入以×符号圆圈表示的矩阵乘法节点，「Q (Query)」 框位于 QK 乘积与最终加权和之间。旁边浮动一个 6×6 注意力热力图格，颜色从冰蓝到深红，附说明文字。浅鼠尾草色标注框写道：「Transformer 注意力机制图解」纯几何与字体构成，无插画。

A precise Swiss-grid technical diagram illustrating scaled dot-product attention inside a Transformer block on a white canvas. Central focus: three vertical columns labeled "Q (Query)", "K (Key)", "V (Value)" in navy, each flowing down into a matrix multiplication node shown as an ×-symbol circle. A softmax box labeled "Softmax" sits between the QK product and the final weighted sum. Attention heat tiles — a small 6×6 grid of colour-coded squares from ice-blue to deep crimson — float beside the computation path with a caption: "Attention weights: how much each token attends to every other." All type in a neutral grotesque; numeric annotations in 9pt. Callout box in pale sage: "In self-attention, every word asks: which other words matter most to me?" No illustrations — pure geometry and type.