Megadose The AI curve, compressed.

MMFlash: Faster Speculative Decoding for Vision-Language Models

· HF Daily Papers ·
MMFlash reports up to 3.58× faster decoding without changing the target model’s output distribution.

The researchers found that draft-token acceptance fell from 46% for tokens least dependent on visual input to 22% for those most dependent on it, even after multimodal training. MMFlash adaptively gives more weight to visually dependent tokens the drafter predicts poorly. For long videos, it also uses sparse attention to cut redundant computation and improve drafting quality. Across the tested multimodal benchmarks, reported speedups exceeded EAGLE-3’s by up to 195.4% and DFlash’s by up to 17.1%.

HF Daily Papers' note

score 5

Categories: Research