Skip to content

MQA (shapes)

Attention by shapes · 10 of 11

Head 1 keeps K and V. Heads 2…H only project Q and borrow that drawer. Cache size drops by about the head count.

The film-backed teaching deck — drawers shrinking by H — lives on MQA. This page is the matrix-flow view of the same idea.

Next: GQA (shapes)