How are you running Falcon? ggllm.cpp is really slow at tokenizing the input, so it doesn't work for large context.