MemoryArgs

planner: MemoryPlannerType = Field(default=MemoryPlannerType.HILL_CLIMB, description='Memory planner strategy', json_schema_extra=stability(experimental_values=[MemoryPlannerType.GREEDY_BY_SIZE]))

Memory planner strategy. Defaults to hill_climb, which keeps greedy’s plan unless its own uses fewer scratch bytes in total and no more in any memory; greedy_by_size is experimental and opt-in.

planner_options: dict[str, Any] = Field(default_factory=dict, description="Planner tunables as JSON, validated against the selected planner's option schema (hill_climb: iterations, seed, max_stall_iterations)")

Tunables forwarded to the selected planner, validated against that planner’s option schema. Unknown keys are rejected with a did-you-mean hint. hill_climb accepts iterations, seed, and max_stall_iterations; greedy and greedy_by_size accept none. On the CLI this is a JSON object: --memory.planner-options '{"iterations": 2000}'.

constraints: list[MemoryConstraint] | None = Field(None, description='Memory constraints (ordered by preference)')

Memory constraints (ordered by preference). Defaults to None.

tensors: list[AttributeRuleset] = Field(..., description='Tensor attributes', default_factory=list)

Per-tensor attribute rulesets.

allocate_arenas: bool = Field(True, description='If true, the module will use internal, statically allocated arenas. If false, the caller must bind every region via <prefix>_bind_arena() / <prefix>_bind_arenas() before model_init.')

If true, generated module declares its own arena buffers.

dump_residency_json: bool = Field(False, description='If true, write a machine-readable residency report (``<prefix>_residency.json``) alongside the emitted module. Mirrors the verbose log summary as JSON for tooling that needs to introspect arena layout, staged-vs-cold residency, and per-tensor placement post-planning.')

If true, write <prefix>_residency.json alongside the emitted module.