从4090D到pytorch2.3实践
一、基础软件安装
visual studio下载
下载链接:https://visualstudio.microsoft.com/zh-hans/downloads/
Python下载
下载链接:https://www.python.org/downloads/windows/
二、cuda和cuDNN安装
cuda安装
GPU为4090D 最高支持的cuda版本 cuda12.6
为匹配pytorch安装,安装cuda12.1(实际cuda12.1.1)
下载链接:https://developer.nvidia.com/cuda-toolkit-archive
cuDNN安装
cuda和cndnn的版本关系查询:
https://developer.nvidia.com/rdp/cudnn-archive?login=from_csdn
选择安装cuDNN-v8.9.7,匹配cuda-v12.1
安装步骤:
(1)下载zip文件
(2)放到cuda文件夹中;
(3)添加到系统环境变量中;
(4)cd到安装目录执行检查程序bandwidthTest.exe和deviceQuery.exe
三、pytorch安装
参看:https://pytorch.org/get-started/locally/
实践安装了pytorch2.3成功跑成。
pytorch测试代码:
import torch
x = torch.rand(5, 3)
print(x)
ok = torch.cuda.is_available()
print(ok)
结果:
tensor([[0.9755, 0.3740, 0.4724],
[0.4747, 0.6055, 0.2118],
[0.7834, 0.7734, 0.7064],
[0.3912, 0.9571, 0.4231],
[0.7749, 0.6789, 0.3190]])
True
四、总结
Win11 + GPU4090D + CUDA-v12.1.1+cuDNN-v8.9.7
Python3.10+ pytorch2.3
五、参考:
https://blog.csdn.net/retainenergy/article/details/126183711
https://blog.csdn.net/tian12488421/article/details/135510482
https://blog.csdn.net/dazuibar/article/details/141352245