vLLM adds the LiLiCorr speculative drafter, ported from SGLang onto its DFlash path
GitHub·medium signal
PR #57934 (+1,933 lines) adds LiLiCorrDraftModel (arXiv 2608.20530). Its head reranks global candidate lattices using target embeddings, full-vocabulary log-probs and draft hidden states. It reuses DFlash2 convolution layers, supports mixed BF16/NVFP4 checkpoints, and matched SGLang's acceptance length exactly in tests on 4x GB200. It follows DSpark and DFlash2 into the same drafting path.