2022-02-18 08:28:37 +00:00
|
|
|
|
# Colossal-AI
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<div id="top" align="center">
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
[![logo](https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/Colossal-AI_logo.png)](https://www.colossalai.org/)
|
|
|
|
|
|
|
|
|
|
一个整合高效并行技术的AI大模型训练系统。
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
<h3> <a href="https://arxiv.org/abs/2110.14883"> 论文 </a> |
|
|
|
|
|
<a href="https://www.colossalai.org/"> 文档 </a> |
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<a href="https://github.com/hpcaitech/ColossalAI-Examples"> 例程 </a> |
|
2022-02-18 08:28:37 +00:00
|
|
|
|
<a href="https://github.com/hpcaitech/ColossalAI/discussions"> 论坛 </a> |
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<a href="https://medium.com/@hpcaitech"> 博客 </a></h3>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
[![Build](https://github.com/hpcaitech/ColossalAI/actions/workflows/PR_CI.yml/badge.svg)](https://github.com/hpcaitech/ColossalAI/actions/workflows/PR_CI.yml)
|
|
|
|
|
[![Documentation](https://readthedocs.org/projects/colossalai/badge/?version=latest)](https://colossalai.readthedocs.io/en/latest/?badge=latest)
|
|
|
|
|
[![codebeat badge](https://codebeat.co/badges/bfe8f98b-5d61-4256-8ad2-ccd34d9cc156)](https://codebeat.co/projects/github-com-hpcaitech-colossalai-main)
|
2022-03-04 10:04:51 +00:00
|
|
|
|
[![slack badge](https://img.shields.io/badge/Slack-join-blueviolet?logo=slack&)](https://join.slack.com/t/colossalaiworkspace/shared_invite/zt-z7b26eeb-CBp7jouvu~r0~lcFzX832w)
|
2022-03-11 05:53:38 +00:00
|
|
|
|
[![WeChat badge](https://img.shields.io/badge/微信-加入-green?logo=wechat&)](https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/WeChat.png)
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
| [English](README.md) | [中文](README-zh-Hans.md) |
|
2022-03-11 05:53:38 +00:00
|
|
|
|
|
2022-02-18 08:28:37 +00:00
|
|
|
|
</div>
|
2022-03-11 05:53:38 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
## 目录
|
|
|
|
|
<ul>
|
|
|
|
|
<li><a href="#特点">特点</a> </li>
|
|
|
|
|
<li>
|
|
|
|
|
<a href="#展示样例">展示样例</a>
|
|
|
|
|
<ul>
|
|
|
|
|
<li><a href="#ViT">ViT</a></li>
|
|
|
|
|
<li><a href="#GPT-3">GPT-3</a></li>
|
|
|
|
|
<li><a href="#GPT-2">GPT-2</a></li>
|
|
|
|
|
<li><a href="#BERT">BERT</a></li>
|
|
|
|
|
</ul>
|
|
|
|
|
</li>
|
|
|
|
|
|
|
|
|
|
<li>
|
|
|
|
|
<a href="#安装">安装</a>
|
|
|
|
|
<ul>
|
|
|
|
|
<li><a href="#PyPI">PyPI</a></li>
|
|
|
|
|
<li><a href="#从源代码安装">从源代码安装</a></li>
|
|
|
|
|
</ul>
|
|
|
|
|
</li>
|
|
|
|
|
<li><a href="#使用-Docker">使用 Docker</a></li>
|
|
|
|
|
<li><a href="#社区">社区</a></li>
|
|
|
|
|
<li><a href="#做出贡献">做出贡献</a></li>
|
|
|
|
|
<li><a href="#快速预览">快速预览</a></li>
|
|
|
|
|
<ul>
|
|
|
|
|
<li><a href="#几行代码开启分布式训练">几行代码开启分布式训练</a></li>
|
|
|
|
|
<li><a href="#构建一个简单的2维并行模型">构建一个简单的2维并行模型</a></li>
|
|
|
|
|
</ul>
|
|
|
|
|
<li><a href="#引用我们">引用我们</a></li>
|
|
|
|
|
</ul>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
## 特点
|
|
|
|
|
|
|
|
|
|
Colossal-AI为您提供了一系列并行训练组件。我们的目标是让您的分布式AI模型训练像普通的单GPU模型一样简单。我们提供的友好工具可以让您在几行代码内快速开始分布式训练。
|
|
|
|
|
|
|
|
|
|
- 数据并行
|
|
|
|
|
- 流水线并行
|
|
|
|
|
- 1维, 2维, 2.5维, 3维张量并行
|
|
|
|
|
- 序列并行
|
|
|
|
|
- 友好的trainer和engine
|
|
|
|
|
- 可扩展新的并行方式
|
|
|
|
|
- 混合精度
|
|
|
|
|
- 零冗余优化器 (ZeRO)
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|
|
|
|
|
|
|
|
|
|
## 展示样例
|
2022-02-18 08:28:37 +00:00
|
|
|
|
### ViT
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/ViT.png" width="450" />
|
2022-02-28 08:03:13 +00:00
|
|
|
|
|
2022-02-28 09:07:14 +00:00
|
|
|
|
- 14倍批大小和5倍训练速度(张量并行=64)
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
2022-02-28 08:03:13 +00:00
|
|
|
|
### GPT-3
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/GPT3.png" width=700/>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
2022-02-28 09:07:14 +00:00
|
|
|
|
- 释放 50% GPU 资源占用, 或 10.7% 加速
|
2022-02-28 08:03:13 +00:00
|
|
|
|
|
|
|
|
|
### GPT-2
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/GPT2.png" width=800/>
|
2022-02-28 08:03:13 +00:00
|
|
|
|
|
2022-02-28 09:07:14 +00:00
|
|
|
|
- 降低11倍GPU显存占用,或超线性扩展
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
### BERT
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/BERT.png" width=800/>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
2022-02-28 09:07:14 +00:00
|
|
|
|
- 2倍训练速度,或1.5倍序列长度
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
请访问我们的[文档和教程](https://www.colossalai.org/)以了解详情。
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
## 安装
|
|
|
|
|
|
|
|
|
|
### PyPI
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
pip install colossalai
|
|
|
|
|
```
|
|
|
|
|
该命令将会安装CUDA extension,如果你已安装CUDA, NVCC和torch。
|
|
|
|
|
|
|
|
|
|
如果你不想安装CUDA extension, 可在命令中添加`--global-option="--no_cuda_ext"`, 例如:
|
|
|
|
|
```bash
|
|
|
|
|
pip install colossalai --global-option="--no_cuda_ext"
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
如果你想使用`ZeRO`, 你可以使用:
|
|
|
|
|
```bash
|
|
|
|
|
pip install colossalai[zero]
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
### 从源代码安装
|
|
|
|
|
|
|
|
|
|
> Colossal-AI的版本将与该项目的主分支保持一致。欢迎通过issue反馈你遇到的任何问题 :)
|
|
|
|
|
|
|
|
|
|
```shell
|
|
|
|
|
git clone https://github.com/hpcaitech/ColossalAI.git
|
|
|
|
|
cd ColossalAI
|
|
|
|
|
# 安装依赖
|
|
|
|
|
pip install -r requirements/requirements.txt
|
|
|
|
|
|
|
|
|
|
# 安装 colossalai
|
|
|
|
|
pip install .
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
如果你不想安装和使用CUDA kernel fusion (使用fused优化器需安装):
|
|
|
|
|
|
|
|
|
|
```shell
|
|
|
|
|
pip install --global-option="--no_cuda_ext" .
|
|
|
|
|
```
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|
|
|
|
|
|
2022-02-18 08:28:37 +00:00
|
|
|
|
## 使用 Docker
|
|
|
|
|
|
|
|
|
|
运行以下命令从我们提供的docker文件中建立docker镜像。
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
cd ColossalAI
|
|
|
|
|
docker build -t colossalai ./docker
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
运行以下命令从以交互式启动docker镜像.
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
docker run -ti --gpus all --rm --ipc=host colossalai bash
|
|
|
|
|
```
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|
2022-03-04 10:04:51 +00:00
|
|
|
|
|
|
|
|
|
## 社区
|
|
|
|
|
欢迎通过[论坛](https://github.com/hpcaitech/ColossalAI/discussions),
|
|
|
|
|
[Slack](https://join.slack.com/t/colossalaiworkspace/shared_invite/zt-z7b26eeb-CBp7jouvu~r0~lcFzX832w),
|
2022-03-11 05:53:38 +00:00
|
|
|
|
或[微信](https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/WeChat.png "qrcode")加入Colossal-AI社区,与我们分享你的建议和问题。
|
2022-03-04 10:04:51 +00:00
|
|
|
|
|
|
|
|
|
|
2022-02-18 08:28:37 +00:00
|
|
|
|
## 做出贡献
|
|
|
|
|
|
|
|
|
|
欢迎为该项目做出贡献,请参阅[贡献指南](./CONTRIBUTING.md)。
|
|
|
|
|
|
2022-03-04 10:04:51 +00:00
|
|
|
|
真诚感谢所有贡献者!
|
|
|
|
|
|
|
|
|
|
<a href="https://github.com/hpcaitech/ColossalAI/graphs/contributors"><img src="https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/contributor_avatar.png" width="800px"></a>
|
|
|
|
|
|
|
|
|
|
*贡献者头像的展示顺序是随机的。*
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
## 快速预览
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
### 几行代码开启分布式训练
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
```python
|
|
|
|
|
import colossalai
|
|
|
|
|
from colossalai.utils import get_dataloader
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# my_config可以是config文件的路径或字典对象
|
|
|
|
|
# 'localhost' 仅适用于单节点,在多节点时需指明节点名
|
|
|
|
|
colossalai.launch(
|
|
|
|
|
config=my_config,
|
|
|
|
|
rank=rank,
|
|
|
|
|
world_size=world_size,
|
|
|
|
|
backend='nccl',
|
|
|
|
|
port=29500,
|
|
|
|
|
host='localhost'
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
# 构建模型
|
|
|
|
|
model = ...
|
|
|
|
|
|
|
|
|
|
# 构建数据集, dataloader会默认处理分布式数据sampler
|
|
|
|
|
train_dataset = ...
|
|
|
|
|
train_dataloader = get_dataloader(dataset=dataset,
|
|
|
|
|
shuffle=True
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 构建优化器
|
|
|
|
|
optimizer = ...
|
|
|
|
|
|
|
|
|
|
# 构建损失函数
|
|
|
|
|
criterion = ...
|
|
|
|
|
|
|
|
|
|
# 初始化colossalai
|
|
|
|
|
engine, train_dataloader, _, _ = colossalai.initialize(
|
|
|
|
|
model=model,
|
|
|
|
|
optimizer=optimizer,
|
|
|
|
|
criterion=criterion,
|
|
|
|
|
train_dataloader=train_dataloader
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
# 开始训练
|
|
|
|
|
engine.train()
|
|
|
|
|
for epoch in range(NUM_EPOCHS):
|
|
|
|
|
for data, label in train_dataloader:
|
|
|
|
|
engine.zero_grad()
|
|
|
|
|
output = engine(data)
|
|
|
|
|
loss = engine.criterion(output, label)
|
|
|
|
|
engine.backward(loss)
|
|
|
|
|
engine.step()
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
### 构建一个简单的2维并行模型
|
|
|
|
|
|
|
|
|
|
假设我们有一个非常巨大的MLP模型,它巨大的hidden size使得它难以被单个GPU容纳。我们可以将该模型的权重以二维网格的形式分配到多个GPU上,且保持你熟悉的模型构建方式。
|
|
|
|
|
|
|
|
|
|
```python
|
|
|
|
|
from colossalai.nn import Linear2D
|
|
|
|
|
import torch.nn as nn
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class MLP_2D(nn.Module):
|
|
|
|
|
|
|
|
|
|
def __init__(self):
|
|
|
|
|
super().__init__()
|
|
|
|
|
self.linear_1 = Linear2D(in_features=1024, out_features=16384)
|
|
|
|
|
self.linear_2 = Linear2D(in_features=16384, out_features=1024)
|
|
|
|
|
|
|
|
|
|
def forward(self, x):
|
|
|
|
|
x = self.linear_1(x)
|
|
|
|
|
x = self.linear_2(x)
|
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
2022-03-11 05:53:38 +00:00
|
|
|
|
## 引用我们
|
2022-02-18 08:28:37 +00:00
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
@article{bian2021colossal,
|
|
|
|
|
title={Colossal-AI: A Unified Deep Learning System For Large-Scale Parallel Training},
|
|
|
|
|
author={Bian, Zhengda and Liu, Hongxin and Wang, Boxiang and Huang, Haichen and Li, Yongbin and Wang, Chuanrui and Cui, Fan and You, Yang},
|
|
|
|
|
journal={arXiv preprint arXiv:2110.14883},
|
|
|
|
|
year={2021}
|
|
|
|
|
}
|
|
|
|
|
```
|
2022-03-11 05:53:38 +00:00
|
|
|
|
|
|
|
|
|
<p align="right">(<a href="#top">返回顶端</a>)</p>
|