For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

splitk_p_ladder

def splitk_p_ladder[sm_count: Int]() -> List[Int]

The rung ladder of split-K partition counts P, shared by the producer and the consumer of a workspace split-K launch.

Single source of truth for two consumers that MUST agree:

• dispatch.mojo's _bucket_ws snaps a desired count UP to a rung (then caps it), so every P the production auto-route picks is a rung of this list or the cap.
• fa4_splitk_combine's launcher compiles one comptime-unrolled combine specialization per rung, and falls back to a generic runtime-P kernel for anything off it.

Keeping them as two hand-copied lists silently costs the specialization whenever one side gains a rung the other does not (which is exactly what happened to the sub-12 rungs).

sm_count is the top rung: _bucket_ws reaches it exactly when raw_grid == 1.

Returns:

List[Int]