single mention · model architecture
Softplus-based per-head attention gating
An attention implementation detail that applies softplus-based gating per head, rather than specifying a key/value grouping method.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See grouped-query attention for the methods filed beside it.
Also called softplus-based per-head gating.
Filed under model architecture › token mixer › softmax attention › grouped-query attention