Efficient Streaming Inference of Multimodal Large Language Models on 1 GPU