Attention Is All You Need

Attention from the ground up: scaling, softmax weights, weighted sums, multi-head splits, and the encoder block that stacks it all. Twelve problems that walk from a single attention head to a full transformer forward pass.

12 problems