Onnx checker failed for bertsquad_mha_layernorm.onnx model

[sreedattaSanjay]

• Onnx checker failed for "bertsquad_mha_layernorm.onnx" model
• with the following error "onnx.onnx_cpp2py_export.checker.ValidationError: Nodes in a graph must be topologically sorted, however input 'bert/encoder/Reshape__27:0' of node: name: bert/encoder/Reshape OpType: Reshape"

Steps to reproduce:

1 Run bert_mha_layernorm_fuse() function in benchmark/examples.py with bertsquad-12.onnx model 2 It will output bertsquad_mha_layernorm.onnx model 3 Load the model and check with onnx checker

Code to reproduce : onnx_model=onnx.load(' bertsquad_mha_layernorm.onnx') onnx.checker.check_model(onnx_model)

Can you please let me know how to solve this error?

[ThanatosShinji]

Hi, what's your purpose for performing a check_model? The fused graph has custom-defined operators like layernorm and MHA. It won't pass this check anyway. Nodes in a graph must be topologically sorted this can be done by g.graph_reorder_nodes()

[ThanatosShinji]

update code here

[sreedattaSanjay]

Thank you @ThanatosShinji for your prompt response