vLLM 自适应验证不再依赖训练置信头，兼容 MTP、EAGLE 等草稿模型
用 vLLM 跑推测解码的可以升级了：MTP、EAGLE 草稿模型不训置信头也能开自适应验证，免重训练。
vLLM 的自适应验证此前要求草稿模型自带训练好的置信头，现在这一前置条件被取消。新版直接利用草稿 logits 中已有的信号：草稿模型自身的 top-token 概率在草稿排序上与训练置信头效果相当，但会系统性高估接受率。一个从目标模型 accept/reject 决策在线训练的小型 logistic 校准可修正该偏差，MTP、EAGLE、DFlash、DSpark checkpoint 无需重训练即可接入。基准测试中，最大草稿长度下的一条统一配置在多数并发水平保持在吞吐-交互性 Pareto 前沿，无需按负载分别调参。功能已在 main 分支通过 enable_adaptive_verification 开关提供，要求草稿模型的注意力后端支持变长验证。
Up until a month ago, adaptive verification in vLLM needed a drafter that shipped with a trained confidence head. Now it works with the draft models you already serve: MTP, EAGLE, DFlash, and DSpark checkpoints without a head. No retraining.
The signal was already in the draft logits. The drafter’s own top-token probability ranks drafts as well as a trained confidence head, but it overestimates how often they’ll be accepted. A tiny logistic fit, trained online from the target’s own accept/reject decisions, corrects the calibration.
In our benchmarks, a single adaptive verification config at the max draft length stays on the throughput–interactivity Pareto frontier at most concurrencies, with no per-load tuning. 📈
Available now on main behind enable_adaptive_verification, for draft models whose attention backends support variable-length verification. Built by Giancarlo Delfin @TheEpicDolphin1 (@inferact), on top of adaptive verification by Lucas Wilkinson (@RedHat_AI) and Ben Chislett (@NVIDIAAI).
Blog link: https://t.co/TxA4M8XUgJ
