Module program 

2026-09-28: A compiled circuit program: the straight-line launches of one plan, with every host decision (kernel, grid, sizes, pointers into the model’s fixed buffers and the workspace) made at compile time.

Owner: model-layers circuit executor. Invariants:

• A step reads only StepEnv: the stream, each row’s GDN state pointers and the block-table width uploaded for the step. Everything else a launch uses was fixed when it was compiled, so a CUDA-graph capture of Program::run replays correctly for any step with the same state pointers (the graph caches key on the sequence’s slot).
• run issues no synchronize and no device-to-host copy, so it may be captured.
• A program holds exactly FusionPlan::launches kernel launches and FusionPlan::copies copies.

Structs

DraftPrograms

2026-09-30: The draft head’s compiled programs, one per row count, ascending (one row first).

GdnState

2026-09-28: One sequence’s recurrent state in one GatedDeltaNet layer.

Launch

2026-09-28: One kernel launch or copy.

Program

2026-09-28: The compiled launches of one plan.

StepEnv

2026-09-28: What varies between two runs of one program.

Enums

LaunchKind

2026-09-29: What a launch puts on the stream.

Constants

MAX_VERIFY_STEPS

2026-09-29: The most rollback points an MTP verify writes: one per row but the last, at K = 4.

Traits

DraftRunner

2026-09-29: A program another component runs in place of its own forward (the MTP draft head’s step), without GDN state. 2026-09-30: One program per row count: one row for a single sequence’s draft, n rows for the batched propose’s.