Change use_cache to True which significantly speeds up inference (#2)

initial