vLLM adds adaptive verification for draft models without confidence heads

AISummary

vLLM says adaptive verification now works with MTP, EAGLE, DFlash, and DSpark draft checkpoints without a trained confidence head or retraining.

The drafter's top-token probability ranks drafts well but overestimates acceptance, so a small logistic fit trained online from the target's accept/reject decisions corrects the calibration.

The feature is available on main behind enable_adaptive_verification for draft models whose attention backends support variable-length verification.

Post on XView on X

vLLM

@vllm_project

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://vllm.ai/blog/2026-10-14-adaptive-verification-online-estimator

Source: vLLM · x.comPublished Oct 14, 5:56 PM ET · added here 6:08 PM ET

vLLM adds adaptive verification for draft models without confidence heads

Topics