Skip to content

Commit 40061ea

Browse files
nullnonenilNULLnullnonenilNULL
authored
Merge pull request #120 from nullnonenilNULL/master
docs(readme): add architecture diagram and Architecture section
2 parents bbd5579 + ad7be35 commit 40061ea

4 files changed

Lines changed: 313 additions & 13 deletions

File tree

‎README.md‎

Lines changed: 18 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -10,7 +10,7 @@
1010
</picture>
1111
</p>
1212

13-
<h4>A modular, scalable, high-performance training framework for LLMs, VLMs, diffusion, and embodied models.</h4>
13+
<h4>A unified, high-performance framework for training LLMs, VLMs, diffusion, and embodied models.</h4>
1414

1515
<p align="center">
1616

@@ -25,11 +25,11 @@
2525
</p>
2626

2727
<p align="center">
28-
<a href="#quickstart"><b>📖 Quick Start</b></a>
28+
<a href="#quickstart"><b>⚡ Quick Start</b></a>
2929
&nbsp;·&nbsp;
3030
<a href="#performance"><b>📊 Performance</b></a>
3131
&nbsp;·&nbsp;
32-
<a href="#models"><b>🤖 Supported Models</b></a>
32+
<a href="#models"><b>🏛️ Supported Models</b></a>
3333
&nbsp;·&nbsp;
3434
<a href="https://github.com/baidu-baige/LoongForge/issues/74"><b>🚀 Roadmap</b></a>
3535
</p>
@@ -42,9 +42,20 @@
4242
4343
**LoongForge** is a unified training framework for **LLMs, VLMs, diffusion, and embodied models**, covering **pre-training**, **continued pre-training**, and **SFT**. Its primary goal is to provide broad coverage of mainstream open-source models while delivering efficient, high-throughput training.
4444

45+
Before going open-source, LoongForge was developed as **AIAK-Training-LLM**, Baidu Baige's training acceleration stack. It has supported production training for enterprise customers across **Education**, **Computer Vision**, and **Embodied AI**, typically delivering **30%~50% speedup over customer baselines**, with the largest production runs reaching **5,000+ XPUs**.
46+
47+
<a id="architecture"></a>
48+
## 🏗️ Architecture
49+
4550
Since training requirements vary across model scenarios, LoongForge builds on multiple distributed backends. LLM/VLM/diffusion run on **Megatron-LM**, while embodied models use a **torch-native DDP/FSDP** stack. Each is deeply optimized to outperform mainstream open-source baselines.
4651

47-
Before going open-source, LoongForge was developed as **AIAK-Training-LLM**, Baidu Baige's training acceleration stack. It has supported production training for enterprise customers across **Education**, **Computer Vision**, and **Embodied AI**, typically delivering **30%~50% speedup over customer baselines**, with the largest production runs reaching **5,000+ XPUs**.
52+
<p align="center">
53+
<picture>
54+
<source media="(prefers-color-scheme: dark)" srcset="./docs/assets/images/architecture/loongforge-architecture-dark.svg">
55+
<source media="(prefers-color-scheme: light)" srcset="./docs/assets/images/architecture/loongforge-architecture.svg">
56+
<img alt="LoongForge Architecture" src="./docs/assets/images/architecture/loongforge-architecture.svg" width="100%">
57+
</picture>
58+
</p>
4859

4960
## 🔥 Latest News
5061

@@ -171,7 +182,7 @@ LoongForge supports a broad range of [state-of-the-art models](https://loongforg
171182
| | Cosmos3 | cosmos3_nano |
172183
| | DreamZero | dreamzero_lora_wan22_5b, dreamzero_full_wan22_5b, dreamzero_lora_wan21_14b, dreamzero_full_wan21_14b, dreamzero_libero_wan22_5b, dreamzero_agibot_wan21_14b, dreamzero_yam_wan21_14b |
173184

174-
## 🏗️ Repository Layout
185+
## 📂 Repository Layout
175186

176187
<details>
177188
<summary><b>📁 Directory tree</b></summary>
@@ -220,7 +231,7 @@ LoongForge is released under the [Apache License 2.0](https://github.com/baidu-b
220231

221232
```bibtex
222233
@software{LoongForge2026,
223-
title = {LoongForge: A modular, scalable, high-performance training framework for LLMs, VLMs, diffusion, and embodied models},
234+
title = {LoongForge: A unified, high-performance framework for training LLMs, VLMs, diffusion, and embodied models},
224235
author = {{The LoongForge Authors}},
225236
year = {2026},
226237
url = {https://github.com/baidu-baige/LoongForge}
@@ -238,3 +249,4 @@ Open a GitHub issue for questions, feedback, or feature requests. You can also j
238249

239250
- **WeChat** — [Scan QR code to join](https://github.com/baidu-baige/LoongForge/issues/80#issue-4594463290)
240251
- **Slack** — [Join here](https://join.slack.com/t/baiduloongforge/shared_invite/zt-3ys3kaq2p-cmdw0nDoaHGOcKibgys5Yw)
252+

‎README_zh.md‎

Lines changed: 19 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -10,7 +10,7 @@
1010
</picture>
1111
</p>
1212

13-
<h4>面向 LLM、VLM、Diffusion 与 Embodied 模型的模块化、可扩展、高性能训练框架。</h4>
13+
<h4>一个统一、高性能的框架，用于训练 LLM、VLM、Diffusion 与 Embodied 模型。</h4>
1414

1515
<p align="center">
1616

@@ -25,11 +25,11 @@
2525
</p>
2626

2727
<p align="center">
28-
<a href="#quickstart"><b>📖 快速开始</b></a>
28+
<a href="#quickstart"><b>⚡ 快速开始</b></a>
2929
&nbsp;·&nbsp;
3030
<a href="#performance"><b>📊 性能表现</b></a>
3131
&nbsp;·&nbsp;
32-
<a href="#models"><b>🤖 支持模型</b></a>
32+
<a href="#models"><b>🏛️ 支持模型</b></a>
3333
&nbsp;·&nbsp;
3434
<a href="https://github.com/baidu-baige/LoongForge/issues/74"><b>🚀 路线图</b></a>
3535
</p>
@@ -42,9 +42,20 @@
4242
4343
**LoongForge** 是面向 **LLM、VLM、Diffusion 与 Embodied 模型** 的统一训练框架，覆盖 **预训练（Pre-training）**、**持续预训练（Continued Pre-training）** 和 **SFT**。其核心目标是覆盖主流开源模型，并提供高效的训练性能。
4444

45+
在开源之前，LoongForge 的前身是 **AIAK-Training-LLM** —— 百度百舸的训练加速栈，已在 **教育**、**计算机视觉** 和 **Embodied AI** 等多家企业客户的生产训练中落地，相对客户原有方案通常带来 **30%~50% 加速**，最大规模的生产训练任务达到 **5,000+ XPU**。
46+
47+
<a id="architecture"></a>
48+
## 🏗️ 架构
49+
4550
由于不同模型场景的训练需求各异，LoongForge 构建在多种分布式训练后端之上。其中 LLM/VLM/Diffusion 采用 **Megatron-LM**，具身模型采用 **torch 原生 DDP/FSDP** 技术栈。每条路线都经过深度性能优化，从而超越主流开源方案。
4651

47-
在开源之前，LoongForge 的前身是 **AIAK-Training-LLM** —— 百度百舸的训练加速栈，已在 **教育**、**计算机视觉** 和 **Embodied AI** 等多家企业客户的生产训练中落地，相对客户原有方案通常带来 **30%~50% 加速**，最大规模的生产训练任务达到 **5,000+ XPU**。
52+
<p align="center">
53+
<picture>
54+
<source media="(prefers-color-scheme: dark)" srcset="./docs/assets/images/architecture/loongforge-architecture-dark.svg">
55+
<source media="(prefers-color-scheme: light)" srcset="./docs/assets/images/architecture/loongforge-architecture.svg">
56+
<img alt="LoongForge 架构图" src="./docs/assets/images/architecture/loongforge-architecture.svg" width="100%">
57+
</picture>
58+
</p>
4859

4960
## 🔥 最新动态
5061

@@ -95,6 +106,7 @@
95106
## 📊 性能表现
96107

97108
相对主流开源基线的训练加速。每一行都标注了测量所用的版本，并随各模型的演进单独刷新：
109+
98110
<img alt="LoongForge Benchmark Speedup" src="docs/assets/images/benchmark_speedup.png" />
99111

100112
<details>
@@ -169,7 +181,7 @@ LoongForge 已支持 LLM、VLM、Diffusion 与 Embodied 等多模态的[广泛
169181
| | Cosmos3 | cosmos3_nano |
170182
| | DreamZero | dreamzero_lora_wan22_5b, dreamzero_full_wan22_5b, dreamzero_lora_wan21_14b, dreamzero_full_wan21_14b, dreamzero_libero_wan22_5b, dreamzero_agibot_wan21_14b, dreamzero_yam_wan21_14b |
171183

172-
## 🏗️ 代码结构
184+
## 📂 代码结构
173185

174186
<details>
175187
<summary><b>📁 目录树</b></summary>
@@ -218,7 +230,7 @@ LoongForge 基于 [Apache License 2.0](https://github.com/baidu-baige/LoongForge
218230

219231
```bibtex
220232
@software{LoongForge2026,
221-
title = {LoongForge: A modular, scalable, high-performance training framework for LLMs, VLMs, diffusion, and embodied models},
233+
title = {LoongForge: A unified, high-performance framework for training LLMs, VLMs, diffusion, and embodied models},
222234
author = {{The LoongForge Authors}},
223235
year = {2026},
224236
url = {https://github.com/baidu-baige/LoongForge}
@@ -235,4 +247,4 @@ LoongForge 的构建离不开 NVIDIA 的 Megatron-LM，并从 HuggingFace Transf
235247
欢迎通过 GitHub Issue 提交问题、反馈或功能建议，也可以加入我们的开发者社区：
236248

237249
- **微信群** — [扫码加入](https://github.com/baidu-baige/LoongForge/issues/80#issue-4594463290)
238-
- **Slack** — [点击加入](https://join.slack.com/t/baiduloongforge/shared_invite/zt-3ys3kaq2p-cmdw0nDoaHGOcKibgys5Yw)
250+
- **Slack** — [点击加入](https://join.slack.com/t/baiduloongforge/shared_invite/zt-3ys3kaq2p-cmdw0nDoaHGOcKibgys5Yw)

‎docs/assets/images/architecture/loongforge-architecture-dark.svg‎

Lines changed: 138 additions & 0 deletions
Loading

0 commit comments

Comments
 (0)