Module attention

Functions

• scaled_dot_product_attention
  Computes (softmax(QK^T*sqrt(d_k)) + M)V. M is the attention mask, and is a bias (0 for unmasked, -inf for masked).