DML incorrect results, probably Split node

[jlami]

Describe the issue

Running the htdemucs model on DML gives incorrect results compared to the CPU execution.

To reproduce

The model we are using is the htdemucs v4 model from here: https://github.com/facebookresearch/demucs

You can find a converted onnx model in this zip: https://github.com/intel/openvino-plugins-ai-audacity/releases/download/v3.4.2-R1/openvino-models.zip (2GB zip, htdemucs_v4.onnx itself is 170MB)

If I comment out this line, DML gives correct results:

onnxruntime/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp

Line 782 in 74028e4

{REG_INFO_VER( 11, Split, typeNameListDefault, supportedTypeListAllScalars, DmlGraphSupport::Supported)}, // Adds negative axis.

Urgency

No response

Platform

Windows

OS Version

10

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

74028e4

ONNX Runtime API

C

Architecture

X64

Execution Provider

DirectML

Execution Provider Library Version

1.14

[jlami]

Setting ep.dml.disable_graph_fusion to 1 while enabling the Split op again also gives correct results, so I guess the problem is in the DmlGraphFusionTransformer

[sophies927]

@smk2007

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.

[jlami]

Is there anything I can do to help with this issue? Or do you have any pointers as to where I could look myself. I could not really make any sense of the graph fusion yet, so any pointers are welcome.