converter

AotModel API

The AotConverter class is responsible for converting a model into an AOT inference C module.

Classes

Name	Description
`AotConverter`	Performs the AOT conversion process.
`ConversionResult`	What a finished conversion wrote.

Machine-readable model

ConversionResultclassWhat a finished conversion wrote, as AotConverter.convert() returns it.
AotConverterclassConverts a trained model into a self-contained AOT inference C module.

ConversionResult(
    output: Path,
    prefix: str,
    plan: OptimizationPlan,
    report: OptimizationReport,
    work_dir: Path | None,
    operators: tuple[OperatorView, ...] | None = None,
    tensors: tuple[TensorView, ...] | None = None,
    regions: tuple[RegionView, ...] | None = None,
    placements: Mapping[str, PlacementView] | None = None,
) -> None

dataclass

What a finished conversion wrote, as AotConverter.convert() returns it.

A result compares and hashes by identity, so it can key a dict or join a set. To compare two conversions, compare their plan and report, which compare by value.

output: Path

The module directory, module.path joined with module.name, or the .zip or .pack archive that module.path names.

prefix: str

The module.prefix the module’s C symbols and file names start with.

plan: OptimizationPlan

The optimization plan, which the module holds as <prefix>_plan.json.

report: OptimizationReport

The optimization report, which the module holds as <prefix>_report.json.

work_dir: Path | None

The temporary work directory, which holds the generated sources under module.name, when keep_work_dir kept it; otherwise None, as the conversion removed it.

operators: tuple[OperatorView, ...] | None = None

With result_views, the generated operators in execution order, which is the order the generated operator callback reports them in; otherwise None.

tensors: tuple[TensorView, ...] | None = None

With result_views, the module’s tensors in the order of its tensor enumeration; otherwise None.

regions: tuple[RegionView, ...] | None = None

With result_views, the module’s arena regions, where regions[i].id == i; otherwise None.

placements: Mapping[str, PlacementView] | None = None

With result_views, each tensor’s placement, keyed by tensor id; otherwise None.

AotConverter(config: ConvertArgs, registry_context: RegistryContext | None = None)

Converts a trained model into a self-contained AOT inference C module.

The converter owns the six-stage pipeline (Load, Transform, Resolve, Plan, Emit, Export) described on convert(), which is the single entry point; the CLI is a thin wrapper around it. All behavior is driven by the ConvertArgs config supplied at construction, and every dispatch (parsers, operator classes, transforms) goes through the registry context, so a caller can extend the compiler without patching it.

Parameters

Parameters of AotConverter
Name	Type	Default	Description
`config`	`ConvertArgs`	Required	Conversion parameters.
`registry_context`	`RegistryContext \| None`	None	Optional registry context for parser/operator/transform dispatch. If omitted, a fresh default frozen context is created.

config: ConvertArgs = config

Conversion parameters.

registry_context = registry_context or build_default_registry_context()

Registry context used for parser, operator and transform dispatch.

convert() -> ConversionResult

Convert the model to an AOT module.

Code is generated in a temporary work directory that is removed once the conversion ends, whether it succeeds or fails, unless keep_work_dir is set.

Returns

Returns of convert
Value	Type	Description
`ConversionResult`	`ConversionResult`	Where the module was written, its prefix, its
	`ConversionResult`	optimization plan and report, the kept work directory, and, with
	`ConversionResult`	``result_views``, its operators, tensors, regions and placements.

Raises

Errors raised by convert
Type	Description
`InvalidConfigError`	Before the model loads, if an ``operators[]`` attribute key or value is one that no operator the rule reaches in the registry context takes, or if ``module.path`` cannot be written: a directory path that names a file, an archive path that names a directory, a symbolic link to nothing, a path under a file or under a directory it cannot search, a location without write permission, or a file name the file system cannot hold.
`ModelLoadError`	If the model file is missing, unreadable or not a LiteRT model.
`HeliaAotError`	Another subclass naming what stopped the conversion.