1uv pip install vllm
2uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
1vllm serve Qwen/Qwen3.5-27B \
2 --speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.5-27B-DFlash", "num_speculative_tokens": 15}' \
3 --attention-backend flash_attn \
4 --max-num-batched-tokens 32768
1# Optional: enable schedule overlapping (experimental, may not be stable)
2# export SGLANG_ENABLE_SPEC_V2=1
3# export SGLANG_ENABLE_DFLASH_SPEC_V2=1
4# export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
5
6python -m sglang.launch_server \
7 --model-path Qwen/Qwen3.5-27B \
8 --speculative-algorithm DFLASH \
9 --speculative-draft-model-path z-lab/Qwen3.5-27B-DFlash \
10 --speculative-num-draft-tokens 16 \
11 --tp-size 1 \
12 --attention-backend fa3 \
13 --mem-fraction-static 0.75 \
14 --mamba-scheduler-strategy extra_buffer \
15 --trust-remote-code
1from openai import OpenAI
2
3client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
4
5response = client.chat.completions.create(
6 model="Qwen/Qwen3.5-27B",
7 messages=[{"role": "user", "content": "Write a quicksort in Python."}],
8 max_tokens=4096,
9 temperature=0.0
10)
11print(response.choices[0].message.content)
Setup: Single NVIDIA B200, SGLang, thinking enabled, max output length 4096. We report end-to-end throughput, including prefill time. See our
GitHub repository for reproduction scripts.
Special thanks to
David Wang for his outstanding engineering support on this project. We are also grateful to
Modal,
InnoMatrix, and
Yotta Labs for providing the compute resources used to train this draft model.
If you find DFlash useful, please cite our work. To share feedback on DFlash or request new model support, please fill out this form:
DFlash Feedback.
1@article{chen2026dflash,
2 title = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
3 author = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
4 journal = {arXiv preprint arXiv:2602.06036},
5 year = {2026}
6}
