API reference

Every name below is imported straight from the package (from kpnn2 import ...). Module paths inside kpnn2 are private and may change between releases.

Callables

Names you call or construct.

Name	Description
parse_layered	Parse a source/target edgelist into a LayeredSpec (DAG only; optional widths=, ranks=).
parse_adjacency	Parse a source/target edgelist into an AdjacencySpec (packed layout; cycles allowed).
MaskedLinear	Linear layer that keeps only the connections in a fixed mask.
PackedLinear	Linear layer with one weight per live edge, not a dense matrix.
PackedMultiheadAttention	Attention that scores only live edgelist pairs, not a full matrix.
gather_hop_inputs	Build the input for one hop from the layer activations you saved.
scatter_hop_outputs	Split a hop's concatenated source axis back onto source layers.
align_inputs	Feature names → column index so the wiring matches.
map_node_attributions	Attach spec node names to an attribution tensor you already computed.
aggregate_node_attributions	Reduce mapped node scores with a registered method (default rauter_mangano_2026 is not yet implemented).
list_aggregation_methods	Table of registered attribution aggregation methods and their status.

Specs

Returned by the parsers. Frozen dataclasses; densify with to_mask() when you need a rectangle.

Name	Description
LayeredSpec	A DAG as named layers and packed hops; not a ready-made model.
Hop	Incoming packed wiring of one layer; pass its indices to PackedLinear.
Skip	Record of an edge that jumps layers; already a packed pair of its hop.
AdjacencySpec	Every node in one state vector, with packed edges; the general packed layout.

Errors and version

Name	Description
Kpnn2Error	Error raised by the public API when a call is invalid.
__version__	Package version string.