Qwen3-4B Domino Draft Model
This repository contains a Domino/DFlash draft model for speculative decoding with Qwen/Qwen3-4B. The draft model is not intended to be used as a standalone language model; it should be paired with the target model during generation.
Domino keeps draft generation block-parallel while adding a lightweight causal correction head. This preserves the low drafting cost of parallel speculative decoding and improves draft-token acceptance.
Model Details
- Target model:
Qwen/Qwen3-4B - Draft model:
Huang2020/Qwen3-4B-Domino-b16 - Block size:
16 - Recommended usage: one GPU for the direct
spec_generatepath - Qwen3 chat setting: thinking mode disabled
Installation
uv pip install "git+https://github.com/jianuo-huang/sglang.git@feat/domino-tensor-parallel#subdirectory=python"
Quick Usage
SGLang
sglang serve \
--model-path Qwen/Qwen3-4B \
--speculative-algorithm DFLASH \
--speculative-draft-model-path Huang2020/Qwen3-4B-Domino-b16
Transformers
from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer
draft = AutoModel.from_pretrained(
"Huang2020/Qwen3-4B-Domino-b16",
trust_remote_code=True,
torch_dtype="auto",
device_map="cuda:0",
).eval()
target = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3-4B",
torch_dtype="auto",
device_map="cuda:0",
).eval()
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
messages = [{
"role": "user",
"content": "How many positive whole-number divisors does 196 have?",
}]
input_ids = tokenizer.apply_chat_template(
messages,
return_tensors="pt",
return_dict=False,
add_generation_prompt=True,
enable_thinking=False,
).to(draft.device)
output = draft.spec_generate(
input_ids=input_ids,
target=target,
max_new_tokens=2048,
temperature=0.0,
stop_token_ids=[tokenizer.eos_token_id],
)
generated = output[:, input_ids.shape[1]:]
print(tokenizer.decode(generated[0], skip_special_tokens=True))
Benchmark Figure
The source PDF is available at assets/speedup.pdf.
Citation
@article{huang2026domino,
title={Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding},
author={Huang, Jianuo and Zhang, Yaojie and Zhang, Qituan and Lin, Hao and Xu, Hanlin and Zhang, Linfeng},
journal={arXiv preprint arXiv:2605.29707},
year={2026}
}