腾讯混元发布 Qwen3.8-27B-DSpine-b16 并行草稿投机解码模型

DSpine generates draft tokens in parallel. Each layer passes predicted predecessor features to adjacent positions, carrying causal dependencies through network depth.

This checkpoint is a five-layer draft model for Qwen3.8-27B, with block size 16 and an 8,192-token sliding window.

DSpine architecture

Evaluation

Runtime: SGLang 0.5.22, BF16, FlashAttention 3, concurrency 8.
Draft blocks: DSpine 16; official DFlash2 8.
DSpine inference: dynamic verification referencing DSpark, with ReplaySSM enabled.
General benchmarks: xhigh reasoning, greedy and temperature sampling, at most 4,096 new tokens.
Agent benchmarks: temperature sampling with low and xhigh reasoning; 20 conversations per subset, all recorded tool-decision prefixes replayed in order, prefix caching enabled, at most 32,768 new tokens per request.
Temperature sampling: temperature=1.0, top_p=0.95, top_k=20.

Results

General Benchmark Acceptance Length

Task	DFlash2(T=0)	DSpine(T=0)	DFlash2(T=1)	DSpine(T=1)
GSM8K	5.69	7.20	5.45	6.31
MATH-500	5.49	6.49	5.27	5.66
HumanEval	4.87	5.71	4.34	4.71
MBPP	5.04	6.14	4.79	5.41
MT-Bench	4.11	4.48	3.89	4.03

Agent Acceptance Length

Benchmark	low		xhigh
Benchmark	DFlash2	DSpine	DFlash2	DSpine
WorkBuddy Code	4.19	5.08	4.06	4.92
WorkBuddy Web	3.95	4.91	3.95	5.01
WorkBuddy Office	4.24	5.63	4.20	5.49
SWE-bench Pro	4.19	4.99	4.00	4.77
τ²-bench	4.70	5.53	4.29	4.95

Agent End-to-End Throughput

Throughput is total output tokens divided by end-to-end concurrent replay wall time (tokens/s). Parentheses show speedup over autoregressive decoding (AR).

Benchmark	low			xhigh
Benchmark	AR	DFlash2	DSpine	AR	DFlash2	DSpine
WorkBuddy Code	253.35	590.27 (2.33×)	686.23 (2.71×)	193.00	550.40 (2.85×)	613.55 (3.18×)
WorkBuddy Web	247.40	639.74 (2.59×)	773.19 (3.13×)	252.26	734.04 (2.91×)	840.14 (3.33×)
WorkBuddy Office	210.03	664.77 (3.17×)	805.53 (3.84×)	215.05	662.34 (3.08×)	763.83 (3.55×)
SWE-bench Pro	241.86	657.71 (2.72×)	713.41 (2.95×)	248.17	711.60 (2.87×)	791.30 (3.19×)
τ²-bench	301.39	891.37 (2.96×)	964.30 (3.20×)	340.39	871.98 (2.56×)	978.39 (2.87×)

Installation

git clone https://github.com/Tencent/DSpine.git
cd DSpine
pip install -e .

Usage

sglang serve \
  --model-path Qwen/Qwen3.8-27B \
  --speculative-algorithm DSPINE \
  --speculative-draft-model-path /path/to/Qwen3.8-27B-DSpine-b16 \
  --speculative-num-draft-tokens 16 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --speculative-draft-attention-backend fa3 \
  --linear-attn-backend triton \
  --enable-linear-replayssm-spec \
  --mamba-ssm-dtype float32 \
  --mamba-radix-cache-strategy extra_buffer \
  --max-mamba-cache-size 64 \
  --mem-fraction-static 0.80 \
  --context-length 131072 \
  --max-running-requests 8 \
  --cuda-graph-max-bs-decode 8 \
  --min-free-slots-delay 1 \
  --sampling-backend flashinfer \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --trust-remote-code

Acknowledgements

We thank DFlash, SpecForge, SGLang and FlashInfer. Dynamic verification references DSpark.

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

Apache 2.0