unslothai/unsloth

[Feature] Draft model / speculative decoding

Open

#4,753 创建于 2026年4月1日

在 GitHub 查看
 (4 评论) (2 反应) (0 负责人)Python (64,271 star) (5,658 fork)batch import
feature requestgood first issuehelp wanted

描述

Can we have the possibility to select draft model in ui? Seems like an important feature, I wonder how fast would Qwen3.5 27b be if I used Qwen3.5 0.8b as draft model.

贡献者指南

技术栈
python
领域
frontend
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
fresh
清晰度议题是否清楚说明期望改动、验收标准和下一步。
mostly clear
前置要求
Understanding of speculative decodingFamiliarity with unsloth UI codePython basics
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
45
研究方向
Examine the unsloth UI codebase for model selection patterns, likely in the Gradio or Streamlit frontend. Check how the main model is loaded and configured, and identify where a draft model dropdown could be added. Look at the inference pipeline to understand how to pass a draft model. Review comments in issue #4753 for any additional context or design decisions from maintainers.
[Feature] Draft model / speculative decoding · unslothai/unsloth#4753 | Good First Issue