attention的一些关系图

graph TD A[字符串: 'hello'] -->|encode| B[字节: b'hello'] B -->|decode| A C[整数列表: 104,101,108,108,111] -->|bytes| B B -->|list| C D[bytearray] -->|bytes| B B -->|bytearray| D F[字符串: 你好] -->|encode utf-8| G[字节: utf-8编码] G -->|decode utf-8| F

graph TD subgraph "Post-Norm（原始Transformer）" A1[输入x] --> B1[多头注意力] B1 --> C1[Dropout] A1 --> D1[残差连接 +] C1 --> D1 D1 --> E1[LayerNorm] E1 --> F1[FFN前馈网络] F1 --> G1[Dropout] E1 --> H1[残差连接 +] G1 --> H1 H1 --> I1[LayerNorm] I1 --> J1[输出] end subgraph "Pre-Norm（现代变体）" A2[输入x] --> B2[LayerNorm] B2 --> C2[多头注意力] C2 --> D2[Dropout] A2 --> E2[残差连接 +] D2 --> E2 E2 --> F2[LayerNorm] F2 --> G2[FFN前馈网络] G2 --> H2[Dropout] E2 --> I2[残差连接 +] H2 --> I2 I2 --> J2[输出] end