wei390/SpecPrefetch - DeepSeek-VL2 future-expert predictor training and recall eval

README titles the repo SpeMoe, a future-expert predictor for a frozen DeepSeek-VL2-Tiny. It trains a head that reads the hidden state at layer l and predicts the teacher router's Top-K set at l+1, with BCE plus a ranking loss, and it ships an LRU-cache simulator (eval/verify_expert_prefetch.py) rather than a runtime that copies experts. The README quotes full-set next-layer recall on DeepSeek-VL2-Tiny (K=6) of recall@3 0.4963, recall@6 0.8961, recall@8 0.9683, exact-match@6 0.4653 on ChartQA_TEST (n=2500), and recall@6 falling to 0.8151 on openai_humaneval (n=164). FUTURE_EXPERT_RANK defaults to 32 and only horizon 1 is supported. The README points at weights at https://huggingface.co/jinwei001/SpecPrefetch_deepseekvl2. The page showed 5 commits, 5 stars, and 1 fork on 2026-10-18. No license file was rendered. The arXiv paper does not cite this repository, so the repo is recorded as a related artifact by the same first author name (jinwei001 / Jinwei Kong), not as the paper's official code release.