Attention Mechanism詳解。Self-Attention (Q/K/V Dot-Product)・Multi-Head MHA (h heads parallel)・Multi-Query MQA (1 KV head・LLaMA-2)・Grouped-Query GQA (n KV heads・Llama 3)・Multi-Head Latent Attention MLA (DeepSeek V3・Low-rank KV Compression)・Sliding Window Attention SWA (Mistral)・Sparse Attention・Sliding+Global (Gemma 3)・Linear Attention (Mamba代替)・Flash Attention 3・¥0 OSS、2026年MLA+Mamba Hybrid主流。