In order to serialize and share a model, one can use torch.export and save the ExportedProgram in binary format with torch.export.save(exported_program, 'exported_program.pt2'). However, the serialized object contains all model parameters, that can be heavy for large Vizion or generative models. To reduce the size of serialized objects, it would be helpful to decouple the model architecture from its parameters in order to export the architecture only and save space, especially when the model has not been trained before export. Is there any way to do so with torch.export? Is it planned for the future?