README.md 857 Bytes
Newer Older
zhanggezhong's avatar
zhanggezhong committed
1
2
# FastMoE

Sugon_ldc's avatar
Sugon_ldc committed
3
4
5
6
一个易于使用和高效的系统,支持PyTorch的混合专家(MoE)模型。

## 安装

zhanggezhong's avatar
zhanggezhong committed
7
方式1:
Sugon_ldc's avatar
Sugon_ldc committed
8

zhanggezhong's avatar
zhanggezhong committed
9
[这里](http://10.0.35.93:8000/customized/fastmoe/22.10/fastmoe-0.3.0%2Bdtk22.10-cp37-cp37m-linux_x86_64.whl)下载whl包,通过pip3 install安装(链接中的whl包为dtk22.10版本)
Sugon_ldc's avatar
Sugon_ldc committed
10

zhanggezhong's avatar
zhanggezhong committed
11
方式2:
Sugon_ldc's avatar
Sugon_ldc committed
12

zhanggezhong's avatar
zhanggezhong committed
13
14
15
16
```
python3 setup.py build
python3 setup.py install 
```
Sugon_ldc's avatar
Sugon_ldc committed
17

zhanggezhong's avatar
zhanggezhong committed
18
## 测试
Sugon_ldc's avatar
Sugon_ldc committed
19

zhanggezhong's avatar
zhanggezhong committed
20
所有测试文件在test文件夹中
Sugon_ldc's avatar
Sugon_ldc committed
21
22

```
zhanggezhong's avatar
zhanggezhong committed
23
pip3 install pytest
Sugon_ldc's avatar
Sugon_ldc committed
24
```
zhanggezhong's avatar
zhanggezhong committed
25
26

```
zhanggezhong's avatar
zhanggezhong committed
27
28
29
30
31
32
33
34
pytest -q test_dp.py
pytest -q test_ddp.py
pytest -q test_local_exchange.py
pytest -q test_numerical.py
pytest -q test_swipe.py
pytest -q test_zero.py
pytest -q test_gates.py
```
Sugon_ldc's avatar
Sugon_ldc committed
35

zhanggezhong's avatar
zhanggezhong committed
36
测试过程中不出现failed,全部pass即为通过
Sugon_ldc's avatar
Sugon_ldc committed
37

zhanggezhong's avatar
zhanggezhong committed
38
tips:Fastmoe系统应用于大规模分布式场景,因此在进行测试的时候,应使用不少于8张计算加速卡,否则会出现skip的问题。