issue about the exported onnx model

[czy2014hust]

I use the AIMET PTQ to quantize the CLIP text model.

But I encounter this error [KeyError: 'Graph has no buffer /text_model/encoder/layers.0/layer_norm1/Constant_output_0, referred to as input for text_model.encoder.layers.0.layer_norm1#2'] when using the qnn-onnx-converter to convert the onnx model.

I check the node in the exported onnx model with Netron, I find the second input of the Pow is disappear. It is just constant node with value 2.0.

[czy2014hust]

By the way, the AIMET commit: 35e588226f02 with torch-gpu.

[zzh-www]

+1. I got this error with aimet1.29.0 too.

[quic-hitameht]

Hi @czy2014hust Can you please tell us which QNN version are you using to convert ONNX models?

[czy2014hust]

QNN version：2.16.4.231110

[shiyuetianqiang]

+1

[Aaron4Fun]

+1. Any suggestions will be appreciated.

[zzh-www]

I found that the problem sovled by converting it with aimet 1.25.0.

[quic-hitameht]

Thanks for sharing your QNN version. I was under the impression that this issue is resolved with QNN version you are using. But let me check and get back to you on this.

Meanwhile, you can use Pytorch 1.9 variant of AIMET 1.29 release. https://github.com/quic/aimet/releases/tag/1.29.0

Background: The root cause for this issue is that Pytorch 1.13 version applies bunch of optimizations to ONNX graph and provides more optimized graphs when the mode is set to EVAL. This includes constant folding, Conv+BN fusion etc.

In this case, the optimization has found identical values for LayerNorm parameters and replaces them with a single value, so it ends up with only one initializer in the ONNX graph and it messes up the QNN conversion step.

[czy2014hust]

Thank you for your reply. I met the same issue with QNN latest version 2.17.0.231124.

[czy2014hust]

Thanks for sharing your QNN version. I was under the impression that this issue is resolved with QNN version you are using. But let me check and get back to you on this.

Meanwhile, you can use Pytorch 1.9 variant of AIMET 1.29 release. https://github.com/quic/aimet/releases/tag/1.29.0

Background: The root cause for this issue is that Pytorch 1.13 version applies bunch of optimizations to ONNX graph and provides more optimized graphs when the mode is set to EVAL. This includes constant folding, Conv+BN fusion etc.

In this case, the optimization has found identical values for LayerNorm parameters and replaces them with a single value, so it ends up with only one initializer in the ONNX graph and it messes up the QNN conversion step.

Hi,Does this issue exist for version AIMET-1.29?

[theoctopusride]

you need to change the onnx simplifier to onnx_utils.simplify_onnx_model = False

[MisterTab]

May I ask how this problem was ultimately resolved

[MisterTab]

@czy2014hust

[1826133674]

I meet this error in aimet_torch 1.31.0. Any suggestions to solve this problem?

Name: aimet-torch Version: torch-gpu-1.31.0 Summary: AIMET torch Package Home-page: https://github.com/quic/aimet/releases/download/torch-gpu_1.31.0 Author: Qualcomm Innovation Center, Inc. Author-email: aimet.os@quicinc.com License: NOTICE.txt

[1826133674]

you need to change the onnx simplifier to onnx_utils.simplify_onnx_model = False

Thank you for your reply！I'v solve it according your tips! By the way, the code is in file xx/aimet_quantsim.py def export_quantsim_model(qsim, output_path, dummy_input, filename_prefix, verbose=False, opset_version=11, use_external_data_format=False, input_names=None, output_names=None):

onnx_utils.update_all_onnx_nodes_name = False
onnx_utils.simplify_onnx_model = False