Implementation:Onnx Onnx Save Model External

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

Overview

Concrete tool for saving ONNX models with external tensor data using save_model with external data parameters and the internal write_external_data_tensors function.

Description

When save_model is called with save_as_external_data=True, it automatically calls convert_model_to_external_data to mark tensors for externalization, then write_external_data_tensors to write the tensor data to external files, and finally serializes the lightweight model protobuf. The write_external_data_tensors function iterates over all tensors marked as external, writes their raw data to the specified files, and clears the in-memory raw data from the tensors.

Usage

Use save_model with save_as_external_data=True for the simple case. Use write_external_data_tensors directly when you have already called convert_model_to_external_data and need to control the writing step separately.

Code Reference

Source Location

Repository: onnx
File: onnx/__init__.py:295-345 (save_model), onnx/external_data_helper.py:335-356 (write_external_data_tensors)

Signature

# save_model with external data parameters (see Save_Model for full signature)
onnx.save_model(
    proto,
    f,
    save_as_external_data=True,
    all_tensors_to_one_file=True,
    location="weights.bin",
    size_threshold=1024,
)

# Internal function for manual control
def write_external_data_tensors(
    model: ModelProto,
    filepath: str,
) -> ModelProto:
    """Serializes data for tensors with data_location=EXTERNAL.

    Args:
        model: Model with externally-marked tensors.
        filepath: Base directory for writing external data files.

    Returns:
        Modified model with raw_data cleared from externalized tensors.
    """

Import

import onnx
from onnx.external_data_helper import write_external_data_tensors

I/O Contract

Inputs

Name	Type	Required	Description
model	ModelProto	Yes	Model with external data markers
filepath	str	Yes	Base directory for writing external data files

Outputs

Name	Type	Description
return	ModelProto	Model with raw_data cleared (data written to disk)
files	Files on disk	External data file(s) containing tensor data

Usage Examples

Automatic External Data Save

import onnx

# One-step save with external data
onnx.save_model(
    model,
    "output/model.onnx",
    save_as_external_data=True,
    all_tensors_to_one_file=True,
    location="model_weights.bin",
    size_threshold=1024,
)
# Creates: output/model.onnx + output/model_weights.bin

Manual Two-step Process

import onnx
from onnx.external_data_helper import (
    convert_model_to_external_data,
    write_external_data_tensors,
)

# Step 1: Mark tensors for externalization
convert_model_to_external_data(model, location="weights.bin")

# Step 2: Write external data and save model
write_external_data_tensors(model, "output/")
onnx.save_model(model, "output/model.onnx")

Related Pages

Implements Principle

Principle:Onnx_Onnx_External_Data_Saving

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