DirectML inference produces incorrect results

[tutryx]

Hello,
I trained the D-FINE nano model on a custom dataset and exported it to ONNX using the provided export_onnx.py.

Inference results:

• ✅ CPU execution provider → correct
• ✅ CUDA execution provider → correct
• ❌ DirectML execution provider → extremely low confidences (~0.05), incorrect bounding box sizes and some operations fallback to the CPU

I used the provided C++ ONNX Runtime example (onnxExample.cpp), adding DirectML with:

OrtGetApiBase()->GetApi(ORT_API_VERSION)->GetExecutionProviderApi("DML", ORT_API_VERSION, reinterpret_cast<const void**>(&m_dmlApi));  
m_dmlApi->SessionOptionsAppendExecutionProvider_DML(session_options, 0);

What I tried:

• Disabled ONNX Runtime optimizations → same wrong results.
• Exported with fixed input size (no dynamic axes), opset 17 → model runs fully on GPU (no CPU fallback), still wrong results.
• Exported without postprocessing → same issue.

So far it looks like DirectML produces incorrect outputs even though CPU/CUDA inference works fine.

CPU / CUDA

DirectML

Is this a known issue with the model/export, or with ONNX Runtime + DirectML?
Would other models as RF-DETR or DT-DETR work with DirectML?

Any guidance would be appreciated.

Thank you!