ai-inference.

DSpark Vision Drafters Bring 3x Speedups to Edge VLMs

Speculative decoding for vision-language models achieves significant inference speedups on edge devices and GPUs with minimal parameter overhead and day-one framework support.