OptimizationArgs

OptimizationArgs()

Model-wide optimization settings: the goal, the approximation gate and the knobs.

Every knob defaults to auto, which a static table resolves from goal and allow_approximate; operators[].optimization overrides a knob per operator type or id. The resolved choice of every operator is written to <prefix>_plan.json. Knob fields mirror the knob registry, helia_aot.optimization.KNOBS.

goal: OptimizationGoal = Field(default=OptimizationGoal.BALANCED, description="What auto knob values favor: latency, balanced (today's defaults), size or accuracy")

What auto favors. balanced keeps today’s defaults.

allow_approximate: bool = Field(default=False, description="Allow auto knob values that change numerics relative to the knob's default (e.g. fast accumulation)")

Allow auto to choose knob values that change numerics relative to the knob’s default. An explicit knob value is used as given.

accumulation: Accumulation = Field(default=Accumulation.AUTO, description="FP16 accumulation: precise (standard layout, the library's default accumulation), fast (packed FP16 weights, one FP16 chain per output) or auto")

FP16 accumulation for kernels that offer a choice: precise, fast or auto.

kernel: KernelChoice = Field(default=KernelChoice.AUTO, description='Int8 convolution and depthwise entry: specialized (shape-specialized ns-cmsis-nn direct entries where they apply), generic (the general entries, fewer kernels linked) or auto; both are exact')

Int8 convolution and depthwise entry: specialized, generic or auto.

budget: dict[str, float] | None = Field(None, description='Reserved: constraints for the optimizer, e.g. accuracy_drop or tcm_bytes', json_schema_extra={'cli': False})

Reserved for the optimizer’s constraints (for example accuracy_drop in points or tcm_bytes); setting it is an error.

plan: Path | None = Field(None, description='Reserved: path to a resolved optimization plan to reproduce', json_schema_extra={'cli': False})

Reserved: a resolved plan to reproduce; setting it is an error.