qwen36: DFlash speculative decoding path for Thor (drafter window, constant-time rollback, opt-in acceptance modes, serving host)#130

Open

LiangSu8899 wants to merge 11 commits into

mainfrom

feat/qwen36-thor-dflash

Commits on Jul 3, 2026

fix(qwen36_thor): match parent _long_tq_effective_k signature
LiangSu8899
committed
feat(qwen36): DFlash speculative decoding on Thor
LiangSu8899
committed
perf(qwen36): constant-time DFlash partial-accept rollback on Thor
LiangSu8899
committed
feat(qwen36): per-token DFlash drafter window with prompt-tail seeding
LiangSu8899
committed
docs(qwen36): DFlash usage, window knobs, and Thor performance
LiangSu8899
committed
fix(qwen36): commit per-token window rows before the tap shuffle
LiangSu8899
committed
feat(serving): stateless OpenAI-compatible host for the DFlash path
LiangSu8899
committed