this will be eventually performed in ModelWrapper automatically
when needed

let's try to stick with ONNX Tensor defs and see what happens