LFM2.5-VL DSpark: measure the real speedup on your own request

DSpark is a small draft model, about 280 million parameters, that helps Liquid AI's 3 billion parameter vision model LFM2.5-VL-3B answer faster. The small model guesses a block of words and the big model checks them, so with greedy decoding the answer is exactly what the big model would have written on its own. You get the speedup, not a different model.

The catch is that there are two clocks. In Liquid AI's H100 image-captioning test, decoding was 2.66x faster, but the whole request was 2.27x faster, because the image still has to be processed first. This guide shows you how to see where the time goes, match the two models, and compare the same request on both. The numbers here are Liquid AI's. I haven't timed it myself.