// (args, kwargs), where args is an empty tuple, and kwargs is a dictionary of
  // string keys to TensorSpecs.
  if (!canonicalized_input_signature.has_tuple_value()) {
    return absl::FailedPreconditionError(absl::StrCat(
        "SignatureDefFunction's canonicalized_input_signature should be "
        "of form tuple(tuple(), dict()), but was instead: \n",

  const SavedConcreteFunction* saved_concrete_func;

  // This field is only present on TF2 ConcreteFunctions, and is useful for
  // determining the original argument *names* of the function, (since the
  // "canonicalized_input_signature" may append extra uniquifying integers).
  // However, SavedBareConcreteFunctions do not have a FunctionSpec.
  // Note(bmzhao): if function_spec_.has_value(), *function_spec_ is guaranteed
  // to be non-null.

        }
      }
      input_signature {
        tuple_value {
        }
      }
    }
  }
}
concrete_functions {
  key: "__inference_my_func_5"
  value {
    canonicalized_input_signature {
      tuple_value {
        values {
          tuple_value {
          }
        }
        values {
          dict_value {
          }
        }
      }
    }

            concrete_functions.at(fn_name);
        TaggedValue input_signature = TaggedValue::Tuple();
        const RepeatedPtrField<StructuredValue>& args =
            saved_concrete_function.canonicalized_input_signature()
                .tuple_value()
                .values(0)
                .tuple_value()
                .values();
        for (const StructuredValue& arg : args) {

  // conditions 2 and 3 are trivially satisfied.
  // We need to ensure that:
  // flatten(input_signature).size() + captures.size() = fdef.signature().size()
  // A concrete function's serialized "canonicalized_input_signature" comes
  // from encoding its "structured_input_signature" field:
  // https://github.com/tensorflow/tensorflow/blob/1c064ab76064c58e54261b805027474885a1534d/tensorflow/python/saved_model/function_serialization.py#L70-L71

      // Our SavedModel import requires input_signature on the tf.function, so
      // we never need to handle the kwonlyargs.
      auto positional_arg_structure =
          concrete_function.canonicalized_input_signature()
              .tuple_value()
              .values(0);
      StructuredValueLinearizer input_linearizer(positional_arg_structure,
                                                 builder.getContext());

      }
      registered_name: "tf.TrackableConstant"
    }
    concrete_functions {
      key: "__inference_<lambda>_18"
      value {
        bound_inputs: 5
        canonicalized_input_signature {
          tuple_value {
            values {
              tuple_value {
              }
            }
            values {
              dict_value {
              }