Implementation:Huggingface Optimum TasksManager Get Exporter Config

Field	Value
Page Type	Implementation
Source Repository	https://github.com/huggingface/optimum
Source File	`optimum/exporters/tasks.py`
Domains	NLP, Computer_Vision, Export
Last Updated	2026-02-15 00:00 GMT

Overview

This implementation provides the concrete APIs for constructing backend-specific export configurations. It includes the TasksManager.get_exporter_config_constructor class method for registry lookup and the make_backend_config_constructor_for_task module-level helper for creating partially-applied configuration constructors.

API Reference

`TasksManager.get_exporter_config_constructor`

Source location: optimum/exporters/tasks.py, lines 1227-1312

Purpose: Retrieves the ExportConfigConstructor for a given model (or model type) and task combination, scoped to a specific export backend.

Signature:

@staticmethod
def get_exporter_config_constructor(
    exporter: str,
    model: Optional["PreTrainedModel"] = None,
    task: str = "feature-extraction",
    model_type: Optional[str] = None,
    model_name: Optional[str] = None,
    exporter_config_kwargs: Optional[Dict[str, Any]] = None,
    library_name: Optional[str] = None,
) -> "ExportConfigConstructor":

Parameters:

Parameter	Type	Default	Description
`exporter`	`str`	required	The exporter backend to use (e.g., `"onnx"`, `"tflite"`).
`model`	`Optional[PreTrainedModel]`	`None`	The model instance. Used to infer `model_type` if not provided.
`task`	`str`	`"feature-extraction"`	The task to retrieve the config for.
`model_type`	`Optional[str]`	`None`	The model type string (e.g., `"bert"`, `"gpt2"`). Inferred from model if not provided.
`model_name`	`Optional[str]`	`None`	The name of the model, used only for exception messages.
`exporter_config_kwargs`	`Optional[Dict[str, Any]]`	`None`	Additional arguments to pass to the exporter config class constructor.
`library_name`	`Optional[str]`	`None`	The library name (`"transformers"`, `"timm"`, `"diffusers"`, `"sentence_transformers"`).

Returns: ExportConfigConstructor -- A callable that takes a PretrainedConfig and returns an ExporterConfig.

Internal logic:

If library_name is None, defaults to "transformers" and merges all library-specific supported model type dictionaries
If model_type is None, extracts it from model.config.model_type (or model.config.export_model_type for custom libraries)
Calls get_supported_tasks_for_model_type to validate the task is supported
Handles task synonym resolution (e.g., "default" may be a synonym for "feature-extraction")
Looks up the constructor from the nested registry: supported_model_type[model_type][exporter][task]
If exporter_config_kwargs is provided, wraps the constructor with functools.partial

`make_backend_config_constructor_for_task`

Source location: optimum/exporters/tasks.py, lines 52-59

Purpose: Creates a partially-applied ExportConfigConstructor from a config class and task string. Handles the special -with-past task suffix by setting use_past=True.

Signature:

def make_backend_config_constructor_for_task(
    config_cls: Type,
    task: str,
) -> "ExportConfigConstructor":

Parameters:

Parameter	Type	Default	Description
`config_cls`	`Type`	required	The `ExporterConfig` subclass to create a constructor for.
`task`	`str`	required	The task string (e.g., `"text-classification"`, `"text-generation-with-past"`).

Returns: ExportConfigConstructor -- A functools.partial object that, when called with a PretrainedConfig, returns an ExporterConfig instance.

Internal logic:

def make_backend_config_constructor_for_task(config_cls: Type, task: str) -> "ExportConfigConstructor":
    if "-with-past" in task:
        if not getattr(config_cls, "SUPPORTS_PAST", False):
            raise ValueError(f"{config_cls} does not support tasks with past.")
        constructor = partial(config_cls, use_past=True, task=task.replace("-with-past", ""))
    else:
        constructor = partial(config_cls, task=task)
    return constructor

Import

from optimum.exporters import TasksManager
from optimum.exporters.tasks import make_backend_config_constructor_for_task

Input/Output Summary

API	Input	Output
`get_exporter_config_constructor`	Exporter name, model/model_type, task	`ExportConfigConstructor` callable
`make_backend_config_constructor_for_task`	Config class, task string	`ExportConfigConstructor` (partial)

Usage Example

from optimum.exporters import TasksManager
from transformers import AutoConfig

# Get the ONNX export config constructor for a BERT model
config_constructor = TasksManager.get_exporter_config_constructor(
    exporter="onnx",
    model_type="bert",
    task="text-classification",
    library_name="transformers",
)

# Build the export configuration
model_config = AutoConfig.from_pretrained("bert-base-uncased")
export_config = config_constructor(model_config)

# The export_config now contains input/output specs, dynamic axes, etc.
print(export_config.inputs)   # e.g., {"input_ids": {0: "batch", 1: "sequence"}, ...}
print(export_config.outputs)  # e.g., {"logits": {0: "batch"}}

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment