Skip to content

v0.1.0

Latest

Choose a tag to compare

@Bobliuuu Bobliuuu released this 13 Nov 07:55
· 1 commit to main since this release

recursive-dflash2 v0.1.0

Head-to-head speculative-decoding benchmark (DFlash, DFlash2, DSpine, EAGLE-3) on one SGLang 0.5.22 build, plus a tree-verification upgrade for DSpine that needs no retraining.

Headline (Qwen3-8B, 1x A100-80GB, batch 1, greedy, mean of 5 datasets; recomputed from bench/results/)

  • DSpine + 48-node tree: 7.20x autoregressive (651 tok/s)
  • DSpine chain: 5.75x, DFlash 4.50x, EAGLE-3 (3-step chain) 2.03x
  • Tree gain over DSpine's own chain, same weights: +25%
  • The tree gain exists at batch 1; at concurrency 8 wider trees lose.

Video

Explainer: recursive_dflash2_explainer.mp4 (attached), also in the repo under video/.

What is not released

No model weights. Drafter weights come from Tencent/DSpine and are not redistributed here.

Reproduce

git clone https://github.com/Bobliuuu/recursive-dflash2
bash recursive-dflash2/bench/setup_box.sh      # pinned stack, downloads Qwen3-8B + DSpine-b16
RDF2_TABLE=factored FIMASK=skip bash bench/p4_serve.sh fast 48 31

Raw per-request results: bench-results.tar.gz (attached) and on Hugging Face: https://huggingface.co/datasets/Bobliu/recursive-dflash2