An Examination of the Compositionality of Large Generative Vision-Language Models