Accelerating vision-language models with LFM2.5-VL-DSpark