float32 or float16 in input
            use_int32_inputs(bool): whether use int32 instead of int64 for some inputs

        Returns:
            T5DecoderInputs: dummy inputs for decoder
        r