The model does not have an ir_version set properly.

[ns-ofekshochat]

I have exported a pytorch model (the gpt2 one from the transformers package) with torch.onnx.export . I've loaded the model and the checker complained that The model does not have an ir_version set properly.
I first tried opset_version parameter, and later set it directly with model.ir_version = onnx.IR_VERSION but to no avail.
has anyone encountered this before?

[xadupre]

Did you try to use the paramter opset_version= when calling torch.onnx.export?

[ns-ofekshochat]

I did, yes

[xadupre]

So you got the right opset but the wrong ir_version? It seems a bug of the converter then. onnxruntime only supports opset 18 and ir_version 8 right now.

[ns-ofekshochat]

does seem so. well, that doesn't seem fun

[xadupre]

• @thiagocrepaldi who worked on torch.onnx.

[justinchuby]

Please also share the versions of the packages you are using: onnx, torch, onnxruntime etc.

[ns-ofekshochat]

torch 2.1.0.dev20230714+cu121
onnx 1.14.0
onnxruntime 1.15.1

[justinchuby]

What’s the IR version being used?