Implementation:Onnx Onnx Load External Data For Model

Knowledge Sources	ONNX ONNX External Data
Domains	External_Data, Large_Model_Support
Last Updated	2026-02-10 00:00 GMT

Overview

Concrete tool for loading externally stored tensor data into an ONNX model provided by the ONNX external_data_helper module.

Description

The load_external_data_for_model function iterates over all tensors in the model that use external data storage, reads the data from the external files, populates the tensor's raw_data field, resets data_location to DEFAULT, and clears the external_data reference entries. This function modifies the model in-place.

Usage

Import this function when you have loaded a model with load_external_data=False and need to load the external data from a different directory than the model file. This is common in deployment scenarios where model structure and weights are stored separately.

Code Reference

Source Location

Repository: onnx
File: onnx/external_data_helper.py
Lines: 68-81

Signature

def load_external_data_for_model(
    model: ModelProto,
    base_dir: str,
) -> None:
    """Loads external tensors into model.

    Args:
        model: ModelProto to load external data into (modified in-place).
        base_dir: Directory containing the external data files.
    """

Import

from onnx.external_data_helper import load_external_data_for_model

I/O Contract

Inputs

Name	Type	Required	Description
model	ModelProto	Yes	Model with external data references (modified in-place)
base_dir	str	Yes	Directory containing the external data files

Outputs

Name	Type	Description
return	None	Model is modified in-place with tensor data loaded

Usage Examples

Load External Data from Custom Directory

import onnx
from onnx.external_data_helper import load_external_data_for_model

# Load model structure without data
model = onnx.load_model("models/model.onnx", load_external_data=False)

# Load data from a different location
load_external_data_for_model(model, "/data/weights/")

# Now all tensors have raw_data populated
for init in model.graph.initializer:
    print(f"{init.name}: {len(init.raw_data)} bytes")

Related Pages

Implements Principle

Principle:Onnx_Onnx_External_Data_Loading

Requires Environment

Environment:Onnx_Onnx_Python_Runtime_Environment

Uses Heuristic

Page Connections

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

Principle

Implementation

Heuristic

Environment