Skip to content

resolved_optimizations

Classes

ConstantInterningStats dataclass 

ConstantInterningStats(groups_interned: int = 0, aliases_created: int = 0, bytes_saved: int = 0)

Summary of exact duplicate constant interning.

Attributes:

  • groups_interned (int) –

    Number of duplicate groups that collapsed to one root.

  • aliases_created (int) –

    Number of alias tensors created across all groups.

  • bytes_saved (int) –

    Total destination-arena bytes avoided by aliasing duplicates.

ResolvedOptimizationStats dataclass 

ResolvedOptimizationStats(constant_interning: ConstantInterningStats = ConstantInterningStats())

Summary of resolved-model optimization passes.

Attributes:

Functions

intern_duplicate_constants 

intern_duplicate_constants(model: AirModel, *, tensor_rules: list[AttributeRuleset] | None = None) -> ConstantInterningStats

Alias exact duplicate constant payloads to a shared storage root.

This pass is intended to run after operator resolve() hooks have created backend-only constants such as weight_sum tensors and LUT payloads, and before memory planning assigns offsets. The tensor IDs stay distinct for debuggability; only the underlying storage is shared via alias_of.

Parameters:

  • model

    (AirModel) –

    AIR model to optimize in place.

  • tensor_rules

    (list[AttributeRuleset] | None, default: None ) –

    Optional per-tensor placement rules. Duplicate constants are interned only when these rules resolve to the same source and destination memory intent.

Returns:

optimize_resolved_model 

optimize_resolved_model(model: AirModel, *, tensor_rules: list[AttributeRuleset] | None = None) -> ResolvedOptimizationStats

Run resolved-model optimization passes.

This phase sits between operator resolution and memory planning so future redundancy-reduction passes can operate on the fully materialized AIR graph without interfering with graph-level transforms or emit-time layout.