Implement Q-Former

Key Insight

Building a small Q-Former by hand demystifies how a sprawling image gets squeezed into a fixed handful of tokens a language model can read. You create 16 learned query vectors and let them cross-attend to a frozen image encoder's patch features, so each query walks away with one compact summary of what it cares about — 16 notes instead of a whole gallery tour. Training it on COCO captions (predict the caption from those 16 image tokens) teaches the queries to capture caption-relevant content, and the choice of 16 over, say, 256 is the whole point: the image becomes a constant, small number of tokens regardless of resolution, which is what keeps feeding images to an LLM affordable.