Turn DiffusionGemma into a Jev-like model. This brings a new level of precision to structured tasks. By seeding a canvas with a response template, vLLM lets you extract confidence scores and probability distributions for yes/no, multiple-choice, and scored questions, all in a single denoising step. Fast, efficient, and ready to scale!