Views
No views yet
Qwen/Qwen3.6-27B.
1uv pip install vllm
2uv pip install -U --torch-backend=auto "vllm @ git+https://github.com/vllm-project/vllm.git@refs/pull/40898/head"uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/23000/head#subdirectory=python"1vllm serve Qwen/Qwen3.6-27B \
2 --speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.6-27B-DFlash", "num_speculative_tokens": 15}' \
3 --attention-backend flash_attn \
4 --max-num-batched-tokens 327681# 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/Qwen3.6-27B \
8 --speculative-algorithm DFLASH \
9 --speculative-draft-model-path z-lab/Qwen3.6-27B-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-code1from openai import OpenAI
2
3client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
4
5response = client.chat.completions.create(
6 model="Qwen/Qwen3.6-27B",
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)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}