Hi,
In the old times, there is fold_constant call in relay, so that we could fold a complex constant expr into constant value.
Take pad for example, this op’s pad value seq is different in torch and onnx. So when torch export to onnx, it would create a subgraph to convert into the sequence required by onnx. https://github.com/pytorch/pytorch/issues/29993#issuecomment-557186018
In relay, we could easily call fold_constant this subgraph into a constant, but in relax, seems we cannot do this easily?