Skip to content

defines

Defines API

Classes:

Copyright 2025 Ambiq. All Rights Reserved.

Classes

MemoryPlannerType

Memory planner type

Attributes:

  • greedy (str) –

    Greedy memory planner

ModuleType

Exported module type

Attributes:

  • neuralspot (str) –

    NeuralSpot module

  • zephyr (str) –

    Zephyr RTOS module

MemoryPlacement

Memory placement for tensors

Attributes:

  • sram (str) –

    Static RAM

  • mram (str) –

    Magnetoresistive RAM

  • tcm (str) –

    Tightly Coupled Memory

Functions

to_qualifiers 
to_qualifiers() -> str

Convert memory placement to qualifiers

ScratchMemoryPlacement

Scratch memory placement for tensors

Attributes:

  • sram (str) –

    Static RAM

  • tcm (str) –

    Tightly Coupled Memory

Functions

to_qualifiers 
to_qualifiers() -> str

Convert scratch memory placement to qualifiers

OperatorAttributeRule

This class defines a powerful and flexible way to customize the conversion process for specific operators in the LiteRT model.

Multiple attributes can be specified and allow slicing the model by operator type, operator identifier(s), or a specific combination of both.

The type field can be a specific operator type (e.g., CONV_2D) or * for all operators. The ident field can be a specific operator identifier or a list of identifiers.

Since multiple attributes can be applied to the same operator, precedance is given based on the pair: 1. (1, 1) if both type and ident are specified and match 2. (1, 0) if type is specified and ident is not 3. (0, 1) if ident is specified and type is not 4. (0, 0) if neither is specified

Currently, if multiple attributes have the same precedence, the last one in the list will be used.

Attributes:

  • type (str) –

    The operator type (e.g. CONV_2D) or * for all.

  • ident (int | list[int] | None) –

    The operator identifier or list of identifiers.

  • attributes (dict[str, BasicType]) –

    A dictionary of operator attributes.

OperatorAttributes

This class provides the baseline set of operator attribute rules.

Additional rules are allowed due to the extra="allow" setting. A specific operator can optionally define its own subclass with additional rules and validation.

Attributes:

TransformsConfig

Load a YAML like:

transforms: "*": false # default for every transform FuseConvRelu: true # attribute for a single one

Any name not in TRANSFORM_REGISTRY is an error.

ConvertArgs

Conversion Arguments

This class provides the high-level arguments for the AOT conversion process.

Attributes:

  • path (Path) –

    Path to yaml configuration.

  • model_path (Path) –

    Path to liteRT file. Default is "model.tflite".

  • output_path (Path) –

    Base path for output module. Can also be a zip file. Default is "output.zip".

  • module_name (str) –

    Name used for module. Default is "helios_aot_nn".

  • module_type (ModuleType) –

    Module type (neuralspot or zephyr). Default is neuralspot.

  • prefix (str) –

    Prefix added to sources to provide unique namespace. Default is "aot".

  • memory_planner (MemoryPlannerType) –

    Memory planner strategy. Default is greedy.

  • operator_attributes (list[OperatorAttributeRule]) –

    Operator attributes. Default is empty list.

  • transforms (TransformsConfig) –

    Transforms configuration. Default is empty.

  • include_test (bool) –

    Include test case. Default is False.

  • subgraph (int) –

    Subgraph index. Default is 0.

  • model_version (str) –

    Model version. Default is "v1.0.0".

  • verbose (int) –

    Verbosity level. Default is 1.

  • log_file (Path | None) –

    Optional log file path. Default is None.