ORT_ENABLE_ALL works in Python but not in C++

[simon-eisenmann-driveblocks]

Describe the issue

I am using an ONNX model in a C++ ROS2 node. When setting the optimization level to ORT_ENABLE_ALL, I encounter the following error (both CPU and CUDA):

[ml_detection_cpp_ros-1] 2024-10-24 14:05:24.694433895 [I:onnxruntime:, graph_transformer.cc:15 Apply] GraphTransformer GemmActivationFusion modified: 0 with status: OK
[ml_detection_cpp_ros-1] 2024-10-24 14:05:24.694575790 [I:onnxruntime:, graph_transformer.cc:15 Apply] GraphTransformer MatMulIntegerToFloatFusion modified: 0 with status: OK
[ml_detection_cpp_ros-1] 2024-10-24 14:05:24.694695104 [I:onnxruntime:, graph_transformer.cc:15 Apply] GraphTransformer DynamicQuantizeMatMulFusion modified: 0 with status: OK
[ml_detection_cpp_ros-1] 2024-10-24 14:05:24.694768165 [V:onnxruntime:, selector_action_transformer.cc:126 MatchAndProcess] Matched Conv
[ERROR] [ml_detection_cpp_ros-1]: process has died [pid 173062, exit code -6, cmd

This issue does not occur when using ORT_ENABLE_BASIC, and the node works fine. Interestingly, the same model works without issues (can perform inference) in a Python script with ORT_ENABLE_ALL. I would like to understand why there is a discrepancy between the C++ and Python implementations.

Any insights into why this might be occurring or suggestions for additional debugging would be helpful.

To reproduce

When I launch the C++ ROS2 node, I get the above mentioned error when optimizing the model with ORT_ENABLE_ALL:

session_options_ = Ort::SessionOptions();
session_options_.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);

Urgency

The issue is urgent because of a project deadline.

Platform

Linux

OS Version

Ubuntu 22.04.5 LTS

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.18.0

ONNX Runtime API

C++

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

No response