[Accuracy] MSclap model accuracy issue (CPU vs QNN EP (NPU) )

[mvijayakquic]

Describe the issue

I have exported the MS Clap model into ONNX and run it on both the CPU and NPU. However, there is an accuracy discrepancy between the CPU and NPU results. Here is the setup I used:

• Python: 3.10
• ORT: onnxruntime-qnn version 1.20.0
• MS Clap model: 2023

To reproduce

We have uploaded the script to export and compare the results, the model, as well as testdata.npy.
Please help us debug the issue and let me know if you

MSClap.zip

Urgency

ASAP

Platform

Windows

OS Version

11

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

ONNX Runtime QNN v1.20.0

ONNX Runtime API

Python

Architecture

ARM64

Execution Provider

Other / Unknown

Execution Provider Library Version

ONNX Runtime QNN v1.20.0

[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.