`nexuml.core.export`¶

nexuml.core.export ¶

Export, reload, and selective checkpoint loading for trained pipelines.

LoadReport `dataclass` ¶

Selective load result for package/state reuse.

TrainingReload `dataclass` ¶

Reloaded package prepared for current-codebase training.

export_package ¶

export_package(
    pipeline: CompiledPipeline,
    path: Path,
    metadata: dict[str, Any] | None = None,
    lightning_module: Any | None = None,
    trainer: Any | None = None,
    checkpoint_path: str | Path | None = None,
    include_modules: list[str] | None = None,
    source_metadata: dict[str, Any] | None = None,
) -> Path

Export a trained pipeline as a rich package-backed artifact directory.

Parameters:

Name	Type	Description	Default
`pipeline`	`CompiledPipeline`	Compiled pipeline to export.	required
`path`	`Path`	Destination directory.	required
`metadata`	`dict[str, Any] \| None`	Optional provenance metadata merged into the artifact.	`None`
`lightning_module`	`Any \| None`	Optional Lightning module for checkpoint sidecars.	`None`
`trainer`	`Any \| None`	Optional Lightning trainer for training-state sidecars.	`None`
`checkpoint_path`	`str \| Path \| None`	Optional source Lightning checkpoint to preserve.	`None`
`include_modules`	`list[str] \| None`	Optional glob patterns for additional source modules to intern (useful for dynamic imports invisible to torch.package).	`None`
`source_metadata`	`dict[str, Any] \| None`	Optional metadata describing the export source (e.g. CLI checkpoint path). Merged into metadata.	`None`

Returns:

Type	Description
`Path`	Path to the created export directory.

load_weights ¶

load_weights(
    pipeline: CompiledPipeline,
    source: str | Path,
    checkpoint: CheckpointLoadSpec | None = None,
    *,
    include: list[str] | None = None,
    exclude: list[str] | None = None,
    allow_missing: bool | None = None,
    allow_shape_mismatch: bool | None = None,
    freeze_loaded: bool | None = None,
) -> LoadReport

Selectively load weights into an already-compiled pipeline.

Returns:

Type	Description
`LoadReport`	LoadReport summarising matched, missing, and excluded keys.

Raises:

Type	Description
`ValueError`	On shape mismatch or unexpected missing keys when not allowed.

load_package ¶

load_package(
    path: Path,
) -> tuple[
    CompiledPipeline, ResolvedConfig, dict[str, Any]
]

Reload an exported pipeline into the current codebase.

Returns:

Type	Description
`tuple[CompiledPipeline, ResolvedConfig, dict[str, Any]]`	Tuple of `(pipeline, resolved_config, metadata)`.

Raises:

Type	Description
`ValueError`	If no scenario config is found in the artifact.

load_inference_package ¶

load_inference_package(
    path: Path,
) -> tuple[
    CompiledPipeline, ResolvedConfig, dict[str, Any]
]

Load the packaged pipeline object directly from the torch.package artifact.

Returns:

Type	Description
`tuple[CompiledPipeline, ResolvedConfig, dict[str, Any]]`	Tuple of `(pipeline, resolved_config, metadata)`.

Raises:

Type	Description
`FileNotFoundError`	If no package artifact exists at path.

load_package_for_training ¶

load_package_for_training(
    path: Path,
    scenario: ScenarioSpec | None = None,
    checkpoint: CheckpointLoadSpec | None = None,
) -> TrainingReload

Reload a package into the current codebase for resume or fine-tuning.

Returns:

Type	Description
`TrainingReload`	TrainingReload with pipeline, lightning module, scenario, and load report.

Raises:

Type	Description
`ValueError`	If no scenario is provided and the artifact has no packaged config.

export_safetensors ¶

export_safetensors(
    pipeline: CompiledPipeline,
    path: Path,
    include: list[str] | None = None,
    exclude: list[str] | None = None,
    metadata: dict[str, Any] | None = None,
) -> Path

Export pipeline weights as SafeTensors plus a JSON manifest.

Returns:

Type	Description
`Path`	Path to the created `.safetensors` file.

export_onnx ¶

export_onnx(
    pipeline: CompiledPipeline,
    path: Path,
    input_key: str | None = None,
    output_key: str = "reconstructed",
    opset_version: int = 18,
) -> Path

Export an inference-only ONNX graph for single-input pipelines.

Returns:

Type	Description
`Path`	Path to the created `.onnx` file.

Raises:

Type	Description
`ImportError`	If the `onnxscript` package is not installed.

infer ¶

infer(
    pipeline: CompiledPipeline,
    x: TensorDict,
    y: TensorDict | None = None,
) -> TensorDict

Run inference on a pipeline (eval mode, no grad).

Returns:

Type	Description
`TensorDict`	Output TensorDict from the pipeline.

nexuml.core.export¶