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 function

verify_schedule

def verify_schedule(program: PipelineProgram, config: PipelineConfig, lgkm_per_a: Int, lgkm_per_b: Int, wait_lgkm_first: Int, wait_vm_last: Int)

Verify structural invariants of a finalized pipeline schedule.

Runs at compile time, zero runtime cost. Catches bugs in schedule construction that would otherwise surface as silent GPU miscomputes.

Checks:

Block completeness: every block has exactly 1 MMA op
Wait count bounds: wait values are non-negative and bounded
Distribution invariant: uniform global loads when max_globals > 0
Stage consistency: fragment/global ops use correct stage for half
Prefetch coverage: at least one global is marked prefetch
Completion coverage: each half has a completion load
Warp stagger LDS safety: no cross-block prefetch/fragment race

Note: these are structural checks on the generated program. They do not verify that the execution ordering respects all LDG dependency edges; that is enforced by the scheduler (greedy_schedule / optimal_schedule) which only places ops whose d=0 predecessors are already scheduled.

Args:

program (PipelineProgram): The finalized PipelineProgram to verify.
config (PipelineConfig): Pipeline configuration providing block structure geometry and partition count.
lgkm_per_a (Int): Per-channel lgkm cost for channel 0 loads, used to bound wait_lgkm_first. 0 means unknown (skips the bound check).
lgkm_per_b (Int): Per-channel lgkm cost for channel 1 loads, used to bound wait_lgkm_first. 0 means unknown (skips the bound check).
wait_lgkm_first (Int): The lgkm wait count applied at the first block of each half. 255 means no wait is emitted.
wait_vm_last (Int): The vm wait count applied at the last block of each half.