This repository pins a SGLang serving recipe for RadixArk/Qwen3.8-27B-NVFP4 with DFlash2 speculative decoding on a single NVIDIA DGX Spark (GB10, 128 GB).

A paired 27-request benchmark measured 57.11 tok/s for DFlash2 versus 12.34 tok/s for no-spec (4.63x median ratio), with DFlash2 as the default and no-spec retained as the fallback. The repo includes pinned checkpoints, build and serve scripts, and a frozen protocol, scoped to single-host evidence without claiming universal speedup or exact output parity.