Linear Self-Attention

题面

实现线性注意力：LinearAttn(Q,K,V) = φ(Q) (φ(K)^T V) / (φ(Q)(Σ_j φ(K_j)))，φ 为逐元素特征映射（如 ELU(x)+1）。

Implementation Requirements

• Use only native features (external libraries are not permitted)
• The solve function signature must remain unchanged
• 输出写入 output

Examples

见页面两组示例（2×4 与 2×2）。

Constraints

• Q,K,V: M×d；1 ≤ M ≤ 10000；1 ≤ d ≤ 128；float32
• Performance: M = 10,000