独立运营 · 不出售排名位置  · 我们怎么评估工具 →
Anclum 提交收录
Z

Zihao-Wu06/AutoDSpark

本站为独立目录,非该工具官网;下方信息取自其公开仓库。

One command trains a small helper model that speeds up your LLM. (原文为英文)

2 ★ 本站实测 · 核对 2026-11-19

装之前:命令归属未核实,仅作参考展示 · 许可证 Apache-2.0 · 链接可访问

第一步(仅参考,归属未核实):git clone https://github.com/Zihao-Wu06/AutoDSpark.git && cd AutoDSpark

★ 2·更新 2026-11-16
在 GitHub 查看 ↗链接已于 2026-11-17 核查 · HTTP 200 · 打开的是厂商自己的官网
官方说明Make your own LLM generate up to 2.8x faster with one command: train a DSpark speculative-decoding drafter for any local model and serve it with vLLM. Faster inference, same performance.
★ 数取值
2 ★ · Zihao-Wu06/AutoDSpark · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
来源明细
从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
收录与活跃
收录于 2026-11-16 · 最近见到 2026-11-20 · 共见到 5 次 · 被 2 个来源列出
评分构成
综合分 203 · 2 ★ · 2 个来源 · repo
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: Agent 技能。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库 Zihao-Wu06/AutoDSpark;我们收录于 2026-11-16;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。
README 原文(英文,未翻译) · 在 GitHub 读完整 README →

# AutoDSpark

以下摘录自该仓库的 README(片段)

### Make your own LLM generate up to 2.8x faster.
One command trains a small helper model that speeds up your LLM. Faster inference, same performance.
Results | How it works | Get started | Full guide
Qwen3-4B on 2× RTX 5090: 165 → 303 tokens/s overall, 165 → 467 tokens/s on math.
## ✨ What you get
Faster answers. In our Qwen3-4B test, replies came 1.5x to 2.8x faster, depending on the
task.
Same performance. Nothing about your model changes: same weights, same answers, same quality.
Made for your model. Works for fine-tuned and less common models that have no ready-made
speed-up available.
One command, a few hours, your own GPUs. No cluster, no machine-learning expertise needed.
## 📊 Results
We trained a helper for Qwen3-4B with AutoDSpark on two RTX 5090 graphics cards. It took
under 3 hours from start to finish. Then we asked the model the same questions with and without
the helper and timed it:
| Task | Before (tokens/s) | After (tokens/s) | Faster by |
| Math | 165 | 467 | 2.83x |
| Information extraction | 165 | 306 | 1.86x |
| Code | 165 | 299 | 1.81x |
| Reasoning | 165 | 288 | 1.73x |
| Writing | 165 | 270 | 1.63x |
| Chat | 165 | 250 | 1.51x |
| Overall | 165 | 303 | 1.83x |
A token is roughly three quarters of an English word. Tasks with predictable output, like math,
code and data extraction, speed up the most. Training on examples from your own use case makes
the helper faster still. Full test setup: guide.
## 💡 How it works
Why LLMs are slow. A language model writes one token at a time, and every single token needs
a full pass through the whole model. A 500-token answer means 500 passes, one after another.
The trick: draft, then check. AutoDSpark trains a small, fast drafter that guesses the
next several tokens at once. Your model then checks the whole guess in one pass, the same
cost as writing a single token. Every correct guess is kept; at the first wrong one, your model
puts in its own token instead. So you get several tokens for the price of one, and the final
answer is exactly what your model would have written alone. This technique is called
speculative decoding; AutoDSpark uses DSpark, a recent
drafter design from DeepSeek-AI and Peking University.
Why you need your own drafter. A drafter is only fast if it guesses well, and it only guesses
well if it knows your model's habits. Ready-made drafters exist for a few popular models; for
anything else, including models you fine-tuned yourself, you have to train one. That used to
mean hand-tuning a dozen settings. AutoDSpark does all of it with one command:
mermaid
flowchart LR
A["1. Plan reads your model and GPUs"] --> B["2. Practice data your model answers thousands of prompts"]
B --> C["3. Train the drafter learns to predict those answers"]
C --> D["4. Speed test with vs without, on your hardware"]
D --> E["5. Serve faster ready-to-run vLLM command"]
Under the hood, training is done by vLLM Speculators
and serving by vLLM. The guide
explains each step in detail.
## ⚡ Get started
You need a Linux machine with an NVIDIA GPU and your model in Hugging Face format.
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[![Listed on Anclum](https://anclum.com/badge/listed.svg)](https://anclum.com/zh/s/zihao-wu06-autodspark/)
来源
github · https://github.com/Zihao-Wu06/AutoDSpark
按领域浏览
Agent 技能
Traction
2 ★
收录来源
2 个独立来源
收录于
今天

更多条目

全部 471 →

所属领域与标签

每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。

浏览全部领域 →

← 返回索引