Conda 配置 CUDA/CuDNN/PyTorch 环境

计划重新调整服务器 Python / Pytorch / CUDA / CuDNN 版本。旧的环境为了兼容一些旧项目 (点名 TVM) 而使用了较早版本，导致新项目各种兼容性问题。在 Conda 中重配。

安装 Anaconda3

root@pve:~/share/project/.env/Anaconda3# chmod 777 Anaconda3-2024.10-1-Linux-x86_64.sh 
root@pve:~/share/project/.env/Anaconda3# ./Anaconda3-2024.10-1-Linux-x86_64.sh 
...

Do you accept the license terms? [yes|no]
>>> yes
...

You can undo this by running `conda init --reverse $SHELL`? [yes|no]
[no] >>> no
...

Thank you for installing Anaconda3!

环境变量

root@pve:~/share/project/.env/Anaconda3# vim /etc/profile
>>> export PATH=/root/anaconda3/bin:$PATH
root@pve:~/share/project/.env/Anaconda3# source /etc/profile
root@pve:~/share/project/.env/Anaconda3# conda -V
conda 24.9.2

安装 Python

创建 Python 3.11 环境

1	root@pve:~/share/project/.env/Anaconda3# conda create -n torch0 python=3.11

初始化

1 2	root@pve:~/share/project/.env/Anaconda3# conda init root@pve:~/share/project/.env/Anaconda3# source /root/.bashrc```

这里使用 python3 才能看到 Conda 安装的正确版本，因为之前配置服务器时在 .bashrc 中将 Python alias 了

1
2
3

root@pve:~/share/project/.env/Anaconda3# conda activate torch0
(torch0) root@pve:~/share/project/.env/Anaconda3# python3 -V
Python 3.11.11

安装 CUDA

CudaToolkit 安装，首先查看当前 Driver 支持的最高版本，然后执行

1	(torch0) root@pve:~/share/project/.env/Anaconda3# conda install nvidia/label/cuda-12.2.2::cuda-toolkit

查找 CuDNN 并安装

(torch0) root@pve:~/share/project/.env/Anaconda3# conda search cudnn
Loading channels: done
# Name                       Version           Build  Channel             
cudnn                          7.0.5       cuda8.0_0  pkgs/main           
cudnn                       8.9.2.26        cuda11_0  pkgs/main               
cudnn                       9.1.1.17        cuda12_1  pkgs/main
...

(torch0) root@pve:~/share/project/.env/Anaconda3# conda install cudnn=9.1

安装 PyTorch

参考 PyTorch 官网安装 PyTorch (p.s. 现在对 Rocm 支持越来越好了，AMD YES)

1	conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.1 -c pytorch -c nvidia

首次安装卡在 99% 了，Ctrl + C 之后重新安装，出现如下报错

1
2
3

CondaVerificationError: The package for pytorch located at /root/anaconda3/pkgs/pytorch-2.5.1-py3.11_cuda12.1_cudnn9.1.0_0
appears to be corrupted. The path 'lib/python3.11/site-packages/torch/include/ATen/cuda/CUDAApplyUtils.cuh'
specified in the package manifest cannot be found.

清理损坏的包

(torch0) root@pve:~/share/project/.env/Anaconda3# conda clean --packages
Will remove 158 (10.90 GB) package(s).
Proceed ([y]/n)? y
# reinstall

Hint: 这里 Solving environment: done 会卡住很久，耐心等待

测试，注意使用 Python3

print(torch.backends.cudnn.version())  # 90100
print(torch.__version__)               # 2.5.1
print(torch.version.cuda)              # 12.1
print(torch.cuda.is_available())       # True