For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo module
geometry
Kernel-geometry-derived scheduling constants.
The pipeline framework consumes four kernel-derived counts that live in two different configs today:
PipelineConfig.vm_per_load_a/vm_per_load_b— vmcnt entries consumed by one channel-A or channel-B prefetch (driveswait_vmderivation).ScheduleConfig.lgkm_per_load_a/lgkm_per_load_b— lgkmcnt entries consumed by one channel-A or channel-B fragment load (driveswait_lgkmderivation).
Both are kernel-geometry properties — they fall out of BM/BN/BK,
warp count, MMA shape, dtype size — but every kernel computes them
from its own warp-grid + MMA-pattern assumptions. KernelGeometry
captures the result once and exposes the derived counts as fields;
each kernel layout (4-wave, ping-pong, ...) provides its own factory
that does the layout-specific math and returns a KernelGeometry.
Existing kernels that pass the four values directly continue to work — this struct is purely additive consolidation of the API surface.
Structs
-
KernelGeometry: Bundles kernel-shape inputs and the derived scheduling counts.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!