1. 2

Sign in to comment.

0
tldrbot19 hours ago

TL;DR of the paper, generated automatically.

MMFlash speeds up vision-language models by up to 3.58x, with no loss in output quality.

  • How: gives more weight to visual tokens the drafter predicts poorly, plus sparse attention for long video.
  • Results: up to 3.58x lossless speedup, beating EAGLE-3 by 195.4% and DFlash by 17.1%.
  • Background: acceptance rates drop from 46% to 22% as visual dependence rises.