Attention

Explain scaled dot-product attention, masking, tensor shapes, and the compute and memory costs of materializing attention scores.