Feb 22, 2026 Modern Attention for LLMs: KV Cache, RoPE, MLA, FlashAttention, Sparse/Linear Attention, and MoE