Skip to main content

Module nary_engine

Module nary_engine 

Source
Expand description

Engine orchestration for bounded exact n-ary induction.

The n-ary counterpart of the binary crate::induce_exact pipeline: enumerate every canonical pattern (crate::nary::enumerate_patterns), flatten the batch into the device layout (crate::nary_layout::flatten_patterns), score it through the device-resident launcher (CudaKernelProvider::ilp_exact_nary_score_device), and reduce the per-pattern coverage counts deterministically (crate::reduce::reduce_nary).

Transfer budget: relation and example values never visit the host — ingestion is device-to-device column copies inside the launcher; row counts come from cached host-side metadata (a pure struct read); the only device-to-host transfers are the two per-pattern count arrays. The reduction then runs on those counts entirely host-side.

Structs§

InduceExactNaryRequest
Inputs to one induce_exact_nary call.
NaryInductionConfig
Bounds and selection size for one induce_exact_nary call.
NaryInductionResult
Result of one induce_exact_nary call.
ScoredNaryCandidate
One kept pattern with full metadata and diagnostics.

Functions§

induce_exact_nary
Run exact n-ary induction against one request.