Skip to content

planner

AotMemoryPlanner API

This class is the base class for all memory planners in the heliaAOT framework.

Classes:

  • AotMemoryPlanner –

    Base class for memory planners

Functions:

Copyright 2025 Ambiq. All Rights Reserved.

Classes

AirMemoryPlanner 

AirMemoryPlanner(target: SocPlatform)

Functions

fill_memory_constraints 
fill_memory_constraints(_model: AirModel, constraints: list[MemoryConstraint] | None = None) -> dict[MemoryType, MemoryConstraint]

Resolve memory constraints, filling defaults from the target when max_size is None.

fill_tensor_constraints 
fill_tensor_constraints(model: AirModel, attributes: list[AttributeRuleset] | None = None) -> dict[TensorId, TensorAttributes]

Resolve per-tensor placement constraints from attribute rules.

merge_alias_lifetimes_to_fixed_point 
merge_alias_lifetimes_to_fixed_point(model: AirModel, tensor_lifetimes: dict[TensorId, TensorLifetime]) -> None

Merge alias-linked tensor lifetimes until no live range changes remain.

populate_alias_allocations 
populate_alias_allocations(model: AirModel, plan: MemoryPlan, sorted_tensors: list[tuple[TensorId, TensorLifetime]]) -> None

Populate allocation entries for alias tensors by reusing root allocations.

plan 
plan(model: AirModel, memory_constraints: list[MemoryConstraint] | None = None, tensor_constraints: list[AttributeRuleset] | None = None) -> MemoryPlan

Plan tensor memory allocation for the model.

Every tensor is bound to an arena slot. Persistent tensors live in their own per-bank persistent arenas; constant tensors live in per-destination-memory constant arenas (cold by default — the arena buffer is the read-only blob in source memory; staged when a per-tensor constant_destination_memory differs from memory, in which case the arena buffer is a writable runtime copy and a separate source blob lives in the source memory).

Parameters:

  • model
    (AirModel) –

    The target model.

  • memory_constraints
    (list[MemoryConstraint] | None, default: None ) –

    Memory constraints (per-platform writable budget).

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

    Per-tensor attribute rules (memory placement, constant_destination_memory override, etc.).

Returns:

Functions

register_memory_planner 

register_memory_planner(cls: type[AirMemoryPlanner]) -> type[AirMemoryPlanner]

Class decorator: register each memory planner by its NAME.

get_memory_planner 

get_memory_planner(name: str) -> type[AirMemoryPlanner]

Get a memory planner by its name.