Parse, load, and save a model

Parse, load, and save a model#

2026-10-12

2 min read time

Applies to Linux

This tutorial walks through the examples/migraphx/cpp_parse_load_save example in the MIGraphX repository. You learn how to parse an ONNX model into a MIGraphX program, inspect the graph, save it to MessagePack format, and reload it later.

The example source is at examples/migraphx/cpp_parse_load_save in the MIGraphX GitHub repository.

What you will learn#

  • Parse an ONNX file into a migraphx::program.

  • Save a program to MessagePack (.mxr) format.

  • Load a previously saved program from MessagePack or JSON format.

  • Swap in your own ONNX model by changing the input file path.

Prerequisites#

  • MIGraphX installed. See MIGraphX on ROCm installation.

  • A C++ compiler and CMake 3.5 or later.

  • An ONNX model file. The tutorial uses any compatible ONNX file you provide as input_model.onnx.

Build the example#

From the examples/migraphx/cpp_parse_load_save directory:

mkdir build
cd build
cmake ..
make

This produces an executable named parse_load_save.

Parse an ONNX model#

By default, the program parses the input file as ONNX and prints the internal graph structure. Replace input_model.onnx with the path to your model:

./parse_load_save input_model.onnx --parse onnx

The C++ API call used in the example is:

migraphx::program p;
migraphx::onnx_options options;
p = parse_onnx(input_file, options);

Computation graphs saved in a compatible serialized format, such as ONNX, can be read by MIGraphX to create a runnable program.

Save a program#

Save the parsed program to MessagePack format. Replace output_name with your preferred base filename (the .mxr extension is added automatically):

./parse_load_save input_model.onnx --parse onnx --save output_name

The program writes output_name.mxr in the current directory.

Load a saved program#

Load a program that was previously saved in MessagePack format:

./parse_load_save output_name.mxr --load msgpack

To load a program saved in JSON format:

./parse_load_save saved_program.json --load json

Use your own model and data#

To adapt this example for your workflow:

  1. Replace input_model.onnx with the path to your ONNX model file.

  2. If your model has dynamic or unspecified dimensions, configure migraphx::onnx_options before parsing. For example, set a default batch size with options.set_default_dim_value(batch), where batch is your batch size integer.

  3. After parsing, compile the program for your target and run inference. See the C++ MNIST example at examples/vision/cpp_mnist/ for compile and evaluate steps.

Command-line reference#

The parse_load_save executable supports the following usage:

./parse_load_save <input_file> [options]
options:
        --parse onnx
        --load  json/msgpack
        --save  <output_file>

Next steps#

  • Complete an end-to-end inference workflow in the C++ MNIST example documented at examples/vision/cpp_mnist/.

  • Use MIGraphX driver to inspect and verify your model from the command line without writing C++ code.

  • See MIGraphX examples for additional examples by domain.

Contents