Quark quantization inference errors

[heman-CL]

Dear great authors,
I've tried to use quark quantization to generate int8 model under NPU.
However, while inferring the model, it shows the following errors:

The snippet model is like: (Use onnx.load to check the node info and it seems there's attribute generated after quark quantization )

Error Unrecognized attribute: axes for operator Squeeze (Actually the original model is split but is converted into slice)

Could you please help to comment it?
Thanks

[cyndwith]

@heman-CL

• Could you please share few details about the specific model you are using?
• What is the ONNX version of the model are you using?
  Having this information will help me in trying to reproduce the error and better understand the issue. Thank you!

[heman-CL]

Hi cyndwith,

I've tried to convert onnx model with opset=11 again. This issue is gone. (Squeeze-11/Squeeze-13 are with different structures)
However, I've faced another errors while inferring. It shows:

KernelParamGenPass.cpp:2130] xir::Op{name = (Squeeze_output_0_DequantizeLinear_Output), type = transpose}. This order: (0,3,2,

And then it just finished without anything.
The original graph: Transpose (perm = (0,2,1)) -> Squeeze (axes = 0)
Quantized_graph: Transpose (perm = (0,2,1)) -> QuantizeLinear -> DequantizeLinear -> Squeeze (axes = 0) -> QuantizeLinear -> DequantizeLinear -> ...

Thanks

[cyndwith]

Please try exporting/update the ONNX model to use opset=17, which is the recommended version for Quark quantization.

[heman-CL]

Hi cyndwith,
I've tried to export onnx with torch.onnx.export( , opset_version = 17).
The issues still happen.
However, I've found the following node properties still show ai.onnx v13 (Squeeze/Split --> node properties --> module)
Do I need to do further conversions?
Thanks

[cyndwith]

It should function with opset_version=17. Could you provide more information about the network so I can replicate this error?