Open Models Almanac
Methodsinference & servingdecoding strategy

single mention · inference & serving

Clustered top-K projection

Replaces full-vocabulary projection with a top-K operation over token clusters to reduce decoding overhead.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See decoding strategy for the methods filed beside it.

Also called top-k operation on clusters of tokens.

Filed under inference & serving › decoding strategy