How to delete the Range operator node in ONNX export

[Buer-Chen]

Description:

We noticed that in the process of exporting the PyTorch model to the ONNX format, the generated ONNX model contains a Range operator. The locations of these nodes are basically like /encoder//encoder//self _ attn _ weights/Range. What part of the code may cause this problem, and how can I modify it to prevent the range operator from appearing in the exported onnx model?
The model export code used is: icefall/egs/librispeech/ASR/zipformer/export-onnx.py
We would appreciate any insights or suggestions on this matter.

[csukuangfj]

like /encoder//encoder//self _ attn _ weights/Range.

Can you post screenshots for it?

What is the issue of having such an operator in ONNX?

[Buer-Chen]

like /encoder//encoder//self _ attn _ weights/Range.

Can you post screenshots for it?

What is the issue of having such an operator in ONNX?
Sure, here is the screenshots:

I want to convert it to rknn model. But rknn-tookit2 do not support the range operator

[csukuangfj]

which version of rknn-toolkit2 are you using?

have you tried 2.1.0 or above?

[csukuangfj]

or are you still using 1.6.0?

[Buer-Chen]

or are you still using 1.6.0?

I tried 2.3.0 and 2.1.0 versions of the tool chain, and both of them encountered this problem. I just looked up the given encoder model in rknn_model_zoo and found that their fp32 model is fixed-length and the derived model has no Range operator, but the int8 model is dynamic-length and contains Range operator. Converting their fp32 model can be converted normally, but the int8 model will encounter the same problem.
The model I converted is our own fine-tuned English model, and the example given by RK is Chinese model. The model given is zipformerv1, and I need to use zipformerv2.