Open Models Almanac
Methodsinference & servingdecoding strategy

specific method · filed under inference & serving

LiLiCorr

LiLiCorr is NVIDIA's block drafter for Nemotron 3.5 Super: it keeps the eight most likely tokens at each drafted position and scores the whole candidate lattice with a small transformer that rates how well neighbouring candidates fit, then walks it left to right. Verification is unchanged, so the output stays the target's.

Written for the explainer Speculative decoding in production: drafters, acceptance and load.

Also called LiLiCorr parallel speculative decoding, LiLiCorr speculative drafter.

sources
2
model
1
lab adopt it
1
strongest
optional

Documented in

Evidence

2 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
optional 2

LiLiCorr improves parallel speculative decoding by jointly evaluating token candidates across an entire draft

optionalinference servingin Nemotron 3.5 Super VLNVIDIA
1 more from this model

LiLiCorr for Nemotron 3.5 Super

optionalinference servingin Nemotron 3.5 SuperNVIDIA
↑