Reproduce “DScale: Scaling Block-Diffusion Speculative Decoding with Adaptive Verification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rongjian Chen, Minxian Xu, Zhengxin Fang, Kejiang Ye, Chengzhong Xu
DScale improves serving for block-diffusion speculative decoding by predicting each request’s useful verification prefix and packing variable-length choices into fixed-address GPU workspaces. It retains DFlash’s full-length drafter while routing tiles and allocating a shared budget of 8N candidate slots. On one A100-40GB, Qwen3-8B throughput rose 43.9% over DFlash across four datasets and concurrency 8–32, within the tested configuration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.