Attention 之后还缺什么：从 Residual Connection 到 Transformer StackSep 19, 2026·5988 words·12 minsDeep-Learning Transformer Residual-Connection Layernorm 6.S191上一篇结尾我写：Multi-Head Attention 仍然只是一个 information routing / read mechanism，一个完整的 Transformer layer 还包含 FFN、residual connection、layer normalization 等组件。