---
library_name: transformers
tags:
  - speculative-decoding
  - diffusion-drafter
  - adaflash
  - qwen3
license: apache-2.0
base_model: z-lab/Qwen3-8B-DFlash-b16
---

<p align="center">
  <img src="https://raw.githubusercontent.com/ZinYY/AdaFlash/main/asset/imgs/AdaFlash_Logo.jpg" width="50%" alt="AdaFlash" />
</p>

<div align="center">

**AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters**

[![Paper](https://img.shields.io/badge/Paper-arXiv%3A2607.19223-b31b1b.svg)](https://arxiv.org/abs/2607.19223)
[![GitHub](https://img.shields.io/badge/GitHub-AdaFlash-blue.svg)](https://github.com/AdaFlash-project/AdaFlash)

</div>

---

## Overview

This is the **AdaFlash drafter** for [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) — an on-policy distilled diffusion drafter with an adaptive length head, trained using PerfectBlend data and reverse-KL loss. It is initialized from [Qwen3-8B-DFlash-b16](https://huggingface.co/z-lab/Qwen3-8B-DFlash-b16).

AdaFlash accelerates LLM inference through speculative decoding. The adaptive length head dynamically adjusts the number of draft candidates per step, balancing speed and acceptance rate.

## Quick Start

AdaFlash requires the patched [SGLang PR #34171](https://github.com/sgl-project/sglang/pull/34171) for serving.

```bash
# Install
conda create -n adaflash python=3.12 -y && conda activate adaflash
pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git@refs/pull/34171/head#subdirectory=python"

# Serve
export MODEL_PATH=Qwen/Qwen3-8B
export DRAFT_MODEL_PATH=AdaFlash/Qwen3-8B-AdaFlash

# Launch the server & benchmark
bash scripts/tools/benchmark.sh
```

For full setup, training, and benchmark details, see the [GitHub repo](https://github.com/AdaFlash-project/AdaFlash).

## Citation

```bibtex
@article{arxiv'26:adaflash,
  title   = {AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters},
  author  = {Qian, Yu-Yang and Wu, Hao-Cong and Chen, Chen and Sun, Jiacheng and Dong, Zhenhua and Zhao, Peng and Zhou, Zhi-Hua},
  journal = {ArXiv preprint},
  volume  = {arXiv:2607.19223},
  year    = {2026}
}
```