Glossary · Architecture

Grouped-Query Attention (GQA)

An attention mechanism variant that shares key and value projections across groups of query heads, reducing memory bandwidth requirements during inference compared to standard multi-head attention.