Snapdragon X Elite NPU - QNN Execution Provider missing operators for Whisper-Medium

[eelsonn]

Hello everyone,

I’m experimenting with running OpenAI’s Whisper-Medium model via ONNX Runtime on a Windows laptop with a Snapdragon X Elite NPU.
• Tiny/Base models run successfully on the NPU (QNN Execution Provider).
• Medium/Large models always fail “Failed to finalize QNN graph. Error code: 1002” and fall back to GPU (AzureExecutionProvider).

I’ve confirmed:

• onnxruntime-qnn nightly is installed
• QNNExecutionProvider is selected but errors out on graph finalize
• The same hardware can run Whisper-Tiny on NPU

It looks like some ONNX operators used by Whisper-Medium aren’t yet implemented in the QNN plugin. I’m not a deep-learning engineer—just a power user—so I’m not sure which operators are missing or how to work around it.

My questions:

1. Is there a known list of ONNX operators not yet supported by the QNN EP on Snapdragon NPUs?
2. What’s the best way to contribute or track progress on adding those kernels?
3. Are there interim workarounds (e.g. graph partitioning, fallback hooks) to force only supported subgraphs onto the NPU?

I can share logs, minimal repro code and ONNX export steps if helpful. Thank you for any guidance or pointers!

—
Environment:

• Windows 11 on Snapdragon X Elite
• onnxruntime-qnn 1.22.0.dev (nightly)
• optimum-onnxruntime, transformers, faster-whisper
• Python 3.10 venv

[brian-at-pieces]

Here's the list of currently supported ops with the QNN EP. More info directly from Qualcomm here

I ran into similar issues with some of my model ops not being supported and just had to let it fall back to the CPU. It should do this by default I believe unless you set the disable_cpu_ep_fallback session option