Open Models Almanac
Methodsoptimizationtraining parallelism

single mention · optimization

Window-bounded KV exchange for sliding-window layers

Under context parallelism, sliding-window layers exchange only the key-value segment reachable by their queries, bounding traffic by the window size.

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

Filed under optimization › training parallelism