single mention · not yet filed
query concatenation
Query concatenation packs multiple query positions, together with their query heads, into a single score MMA.1 It is enabled by KV-outer iteration, where all gathered positions for a tile share the same KV operands, avoiding under-filled MMAs when each KV tile matches only a few queries.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
packs query positions together with their associated query heads, all under the same KV head, into a score MMA
— MiniMax Sparse AttentionUnder KV-outer iteration, however, all gathered positions for a given tile share the same KV operands
— MiniMax Sparse Attention