Requirements
1.NVIDIA CUDA GPU
1.1 安装GPU驱动
lspci | grep -i vga
将查询的GPU型号与这里的进行对比,如果在里面的话即可安装驱动。
有三种方式可以在Ubuntu 20.04上进行安装驱动
我使用的是:
-
在ubuntu的应用程序搜索界面上搜索Software and Updates
-
选择最上面的驱动,然后点击应用更改
-
下载完毕之后重启电脑
nvidia-smi #发现没有安装这个应用
sudo apt install {nvidia-utils-version}
nvidia-smi
以上成功说明GPU驱动安装成功
1.2 安装CUDA工具包
打开CUDA下载链接:https://developer.nvidia.com/cuda-downloads
选择适合自己的包,查询架构可以使用uname -a
,最好不要选择最新的版本,因为到后面可能没有和是的cupy版本与之适配,笔者就是因为选择最新的11.6版本,然后到后面发现没有适配的cupy版本,回来卸载重装。。。。
wget https://developer.download.nvidia.com/compute/cuda/11.5.0/local_installers/cuda_11.5.0_495.29.05_linux.run
sudo sh cuda_11.5.0_495.29.05_linux.run
如果觉得下载很慢的话,可以手动下载下来;
因为前面安装好了驱动,在执行安装的时候会报错:Existing package manager installation of the driver found. It is strongly recommended that you remove this before continuing.
,然后在选项中不选择驱动安装,但是之后会显示没有选择驱动,不过没关系。
===========
= Summary =
===========
Driver: Not Selected
Toolkit: Installed in /usr/local/cuda-11.5/
Samples: Installed in /home/a307/, but missing recommended libraries
Please make sure that
- PATH includes /usr/local/cuda-11.5/bin
- LD_LIBRARY_PATH includes /usr/local/cuda-11.5/lib64, or, add /usr/local/cuda-11.5/lib64 to /etc/ld.so.conf and run ldconfig as root
To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-11.5/bin
vim .bashrc
export PATH=$PATH:/usr/local/cuda-11.5/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-11.5/lib64
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda-11.5
保存三个环境变量之后,让变量生效:
source ~/.bashrc
nvcc -V # 检查是否安装完成
测试在目录/usr/local/cuda-11.5/samples
cd <sample_dir>
sudo make
sudo ./<sample_programme>
可以使用watch -n 10 nvidia-smi
来查看GPU的使用情况,每n秒刷新一次
2.Python Dependencies
python 版本需要在v3.7.0+
以上才可以
2.1 numpy
numpy版本要在v1.18
以上
pip show numpy
pip install -U numpy
python3 -m pip install -U setuptools pip
3.Additional CUDA Libraries
cuTENSOR: v1.3 / v1.4
The library to accelerate tensor operations. See Environment variables for the details.
NCCL: v2.8 / v2.9 / v2.10 / v2.11
The library to perform collective multi-GPU / multi-node computations.
cuDNN: v7.6 / v8.0 / v8.1 / v8.2 / v8.3
The library to accelerate deep neural network computations.
cuSPARSELt: v0.1.0
The library to accelerate sparse matrix-matrix multiplication.
Installing CuPy
分为三种安装方式,笔者使用的官方第一种方式Installing CuPy from PyPI
pip install cupy-cuda115
python3 -m cupyx.tools.install_library --cuda 11.2 --library cutensor
pip freeze | grep cupy
Upgrading Cupy
pip install -U cupy
Uninstalling Cupy
pip uninstall cupy