腾讯混元发布 DSpine 投机解码并行草稿模型,Qwen3-8B 上推理代码已开源
tencent/Qwen3-8B-DSpine-b16
腾讯混元发布投机解码并行草稿器 DSpine,在 Qwen3-8B 上平均接受长度达 7.83(T=0)和 6.98(T=1),高于 DFlash2 的 7.55 和 6.53。该模型在每个草稿层让各位置接收前一位置的预测特征,使因果链沿网络深度展开,同时所有位置仍并行草稿。2026-10 已发布推理代码,采用 Apache 2.0 许可,训练代码与 Qwen3.8-27B 版本仍在路线图中。
Draft in Parallel, Condition Through Depth: Adjacent Causal Injection for Speculative Decoding
DSpine is a parallel drafter for speculative decoding. At every draft layer, each position receives the predicted feature of its predecessor, so the causal chain unfolds over network depth while all positions are still drafted in parallel.
News
- [2026-10] Inference code released.
- [2026-09] Paper released on arXiv.
Results
Acceptance length on Qwen3-8B. DSpine, DSpark and DFlash2 are all retrained on Open-PerfectBlend with the same configuration.
| Temp. | Method | GSM8K | MATH-500 | HumanEval | MBPP | LCB | MT-Bench | Arena-Hard | Avg. |
|---|---|---|---|---|---|---|---|---|---|
| T = 0 | DFlash† | 6.32 | 7.99 | 6.56 | 6.29 | 6.11 | 4.26 | 3.91 | 5.92 |
| Domino† | 9.68 | 9.40 | 7.23 | 6.89 | 6.40 | 5.12 | 4.65 | 7.05 | |
| DSpark | 10.00 | 9.88 | 7.78 | 7.35 | 6.81 | 5.39 | 4.87 | 7.44 | |
| DFlash2 | 10.06 | 10.04 | 7.99 | 7.33 | 6.96 | 5.40 | 5.06 | 7.55 | |
| DSpine | 10.55 | 10.30 | 8.27 | 7.66 | 7.25 | 5.63 | 5.13 | 7.83 | |
| T = 1 | DFlash† | 5.75 | 6.51 | 5.52 | 5.23 | 4.53 | 3.77 | 3.25 | 4.94 |
| Domino† | 8.01 | 7.13 | 5.89 | 5.56 | 4.59 | 4.36 | 3.61 | 5.59 | |
| DSpark | 9.33 | 8.64 | 6.91 | 6.66 | 5.78 | 4.99 | 4.25 | 6.65 | |
| DFlash2 | 9.07 | 8.42 | 6.99 | 6.43 | 5.67 | 4.90 | 4.25 | 6.53 | |
| DSpine | 9.83 | 8.99 | 7.39 | 6.97 | 6.11 | 5.09 | 4.49 | 6.98 |
† Official checkpoints: z-lab/Qwen3-8B-DFlash-b16 and Huang2020/Qwen3-8B-Domino-b16.
Installation
git clone https://github.com/tencent/dspine.git
cd dspine
pip install -e .
Usage
The server settings below are the ones behind the acceptance lengths and throughput reported in the paper.
sglang serve \
--model-path Qwen/Qwen3-8B \
--speculative-algorithm DSPINE \
--speculative-draft-model-path <dspine-model-path> \
--speculative-num-draft-tokens 16 \
--dtype bfloat16 \
--attention-backend fa3 \
--page-size 1 \
--mem-fraction-static 0.65 \
--context-length 8192 \
--max-running-requests 64 \
--cuda-graph-bs-decode 1 2 4 8 16 32 64 \
--trust-remote-code
Roadmap
☐ Training code for DSpine
☐ DSpine for Qwen3.8-27B
Acknowledgements
We thank DFlash, SpecForge, SGLang and FlashInfer for their open-source work.
Citation
@article{zhang2026dspine,
title = {Draft in Parallel, Condition Through Depth: Adjacent Causal Injection for Speculative Decoding},
author = {Zhang, Haohui and Chen, Keyu and Sun, Haocheng and Gu, Weibo and Qiao, Ruizhi and Sun, Xing and Jiang, Bo},
journal = {arXiv preprint arXiv:2609.36173},
year = {2026}
}
License
来源:腾讯混元模型发布 · hf-mirror.com