but slightly easier to call from C++.

    Inputs can be specified in one of two ways:

      - The model can define `_generate_bundled_inputs_for_<function_name>`.
        If the user chooses this method inputs[<function>] should map to None

      - The `inputs` argument to this function can be a dictionary mapping functions to a
        list of inputs, of the same form that will be returned by get_all_bundled_inputs_for_<function_name>.
        Alternatively if only bundling inputs for forward the map can be omitted and a singular list of inputs
        can be provided instead.

        The type of the inputs is List[Tuple[Any, ...]]. The outer list corresponds with a
        list of inputs, the inner tuple is the list of args that together make up one input.
        For inputs of functions that take one arg, this will be a tuple of length one. The Any, ...
        is the actual data that makes up the args, e.g. a tensor.

    Info is an optional parameter that maps functions to a list of strings providing extra information about that
    function's bundled inputs. Alternatively if only bundling inputs for forward the map can be omitted and
    a singular list of information can be provided instead. This could be descriptions, expected outputs, etc.
        - Ex: info={model.forward : ['man eating icecream', 'an airplane', 'a dog']}

    This function will attempt to optimize arguments so that (e.g.)
    arguments like `torch.zeros(1000)` will be represented compactly.
    Only top-level arguments will be optimized.
    Tensors in lists or tuples will not.
    ú