目录

• 一、前言
• 二、安装CUDA
• 三、安装cuDNN
• 四、切换CUDA版本
• 五、总结
• 六、参考

---

一、前言

• 正如题目所言，最近笔者要跑一个TensorFlow搭建的模型，等我按照要求将对应版本的TensorFlow和Keras安装好之后，发现训练模型巨慢，GPU显存只用了一点点而且利用率一直是零，而且提示找不到一些库，提示如下。

2022-06-10 13:06:14.299058: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcudart.so.10.0'; dlerror: libcudart.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299110: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcublas.so.10.0'; dlerror: libcublas.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299155: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcufft.so.10.0'; dlerror: libcufft.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299198: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcurand.so.10.0'; dlerror: libcurand.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299239: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcusolver.so.10.0'; dlerror: libcusolver.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299281: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcusparse.so.10.0'; dlerror: libcusparse.so.10.0: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299326: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Could not dlopen library 'libcudnn.so.7'; dlerror: libcudnn.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/extras/CUPTI/lib64
2022-06-10 13:06:14.299336: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1663] Cannot dlopen some GPU libraries. Skipping registering GPU devices...
2022-06-10 13:06:14.299421: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:

• 观察提示信息和一些现象，笔者得出结论，是CUDA和cuDNN版本没有装合适，因为该程序会去/usr/local/cuda-10.0/lib64文件夹下找库，但是我就没有装CUDA 10.0。去网上找了一番资料后，笔者发现果然是CUDA和cuDNN的版本问题，TensorFlow版本与CUDA版本居然也有对应关系，这下让我更加觉得TensorFlow不好用了。但是这台机器也不是笔者独占的，而且机器上已经有装好的CUDA 11.2和cuDNN 8.4.0了，这种情况确实让人抓狂，不过在笔者浏览了浩瀚的因特耐特之后，发现居然有一种多版本CUDA共存和自由切换的操作，现将该技术整理如下。
• 任务描述：在一台安装了CUDA 11.2和cuDNN 8.4.0的机器上安装CUDA 10.0和cuDNN 7.4.1，使得两者互不干扰和自由切换。
• CUDA和cuDNN的版本选择参考这篇博客。

---

二、安装CUDA

1. 查看已有CUDA环境
   

2. 从官网下载CUDA 10.0的runfile到服务器上。
   

3. 安装CUDA 10.0
   执行如下指令

   sudo sh cuda_10.0.130_410.48_linux.run
   

• 出现协议说明，可以按q跳过。
  

  - 出现问题`Do you accept the previously read EULA?`
  	- 输入`accept`+回车，继续安装。
  
  - 出现不支持配置的提醒：`You are attempting to install on an unsupported configuration. Do you wish to continue?`
  	- 输入`y`，继续安装。
  
  - 出现是否安装显卡驱动的提醒，我们已经装过了：`Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 410.48？`
  	- 输入`n`，继续安装。
  
  - 出现是否安装CUDA工具包：`Install the CUDA 10.0 Toolkit?`
  	- 输入`y`，开始安装。
  
  - 出现工具包安装地址：`Enter Toolkit Location`
  	- 回车
  
  - 出现是否添加符号链接，现在已经有一个了，为了不影响现有的CUDA环境，选择否：`Do you want to install a symbolic link at /usr/local/cuda?`
  	- 输入`n`，继续安装。
  
  - 出现是否安装样例，选择是：`Install the CUDA 10.0 Samples?`
  	- 输入`y`，继续安装
  
  - 出现安装样例位置，默认即可：`Enter CUDA Samples Location`
  	- 回车
  

  不出意外此时应该安装完成，但如果此时你也出现Error: unsupported compiler: 9.4.0. Use --override to override this check.报错，我们按照他说的加上--override选项跳过检查。
  

  执行新的指令，选项和上图一致：

  sudo sh cuda_10.0.130_410.48_linux.run --override
  

  安装成功会出现以下提示：
  

4. 为了不影响现有的CUDA环境，就不修改环境变量了，下文会详细讲述怎么使用新安装的CUDA 10.0。

三、安装cuDNN

1. 根据安装的CUDA工具包版本在官网选择适合版本的cuDNN，本文安装的CUDA版本是10.0，就选择TensorFlow 1.14.0对应的cuDNN 7.4.1，选择Local Installer for Linux x86_64 (Tar)。
   

2. 复制cuDNN库的链接，使用wget下载或者下载到自己电脑之后再传到服务器上。
   下载下来之后，文件名是cudnn-10.0-linux-x64-v7.4.1.5.solitairetheme8，需要重命名一下，改成cudnn-10.0-linux-x64-v7.4.1.5.tgz：

   	mv cudnn-10.0-linux-x64-v7.4.1.5.solitairetheme8 cudnn-10.0-linux-x64-v7.4.1.5.tgz
   

3. 解压cuDNN文件，并进入解压出的文件夹，拷贝文件到/usr/local/cuda-10.0中。

   	tar -xvf cudnn-10.0-linux-x64-v7.4.1.5.tgz
   	cd cuda
   	sudo cp lib64/* /usr/local/cuda-10.0/lib64/
   	sudo cp include/* /usr/local/cuda-10.0/include/
   	sudo chmod a+r /usr/local/cuda-10.0/lib64/*
   	sudo chmod a+r /usr/local/cuda-10.0/include/*
   

4. 查看cuDNN版本，指令为cat /usr/local/cuda-10.0/include/cudnn.h | grep CUDNN_MAJOR -A2。
   

5. 更新软链接，如果你安装的不是7.4.1记得更新下边命令中的数字。

   	cd /usr/local/cuda-10.0/lib64/
   	sudo rm -rf libcudnn.so libcudnn.so.7
   	sudo ln -s libcudnn.so.7.4.1 libcudnn.so.7
   	sudo ln -s libcudnn.so.7 libcudnn.so
   	sudo ldconfig -v
   

6. 最后避免影响到原来的CUDA环境，再执行一下

   	source /etc/profile
   

   此时另一个版本的CUDA和cuDNN已经“偷偷”安装好了。

   但是此时nvcc -V版本还是11.2，具体怎么实现CUDA版本转换，请看下节。

---

四、切换CUDA版本

• 切换到普通用户，查看CUDA版本，可以看到还是11.2。
  
• 下面我们要用到一个脚本。phohenecker大神写的CUDA版本切换脚本：
  特此将代码附上：

	#!/usr/bin/env bash
	
	# Copyright (c) 2018 Patrick Hohenecker
	#
	# Permission is hereby granted, free of charge, to any person obtaining a copy
	# of this software and associated documentation files (the "Software"), to deal
	# in the Software without restriction, including without limitation the rights
	# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
	# copies of the Software, and to permit persons to whom the Software is
	# furnished to do so, subject to the following conditions:
	#
	# The above copyright notice and this permission notice shall be included in all
	# copies or substantial portions of the Software.
	#
	# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
	# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
	# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
	# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
	# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
	# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
	# SOFTWARE.
	
	# author:   Patrick Hohenecker <mail@paho.at>
	# version:  2018.1
	# date:     May 15, 2018
	
	
	set -e
	
	
	# ensure that the script has been sourced rather than just executed
	if [[ "${BASH_SOURCE[0]}" = "${0}" ]]; then
	    echo "Please use 'source' to execute switch-cuda.sh!"
	    exit 1
	fi
	
	INSTALL_FOLDER="/usr/local"  # the location to look for CUDA installations at
	TARGET_VERSION=${1}          # the target CUDA version to switch to (if provided)
	
	# if no version to switch to has been provided, then just print all available CUDA installations
	if [[ -z ${TARGET_VERSION} ]]; then
	    echo "The following CUDA installations have been found (in '${INSTALL_FOLDER}'):"
	    ls -l "${INSTALL_FOLDER}" | egrep -o "cuda-[0-9]+\\.[0-9]+$" | while read -r line; do
	        echo "* ${line}"
	    done
	    set +e
	    return
	# otherwise, check whether there is an installation of the requested CUDA version
	elif [[ ! -d "${INSTALL_FOLDER}/cuda-${TARGET_VERSION}" ]]; then
	    echo "No installation of CUDA ${TARGET_VERSION} has been found!"
	    set +e
	    return
	fi
	
	# the path of the installation to use
	cuda_path="${INSTALL_FOLDER}/cuda-${TARGET_VERSION}"
	
	# filter out those CUDA entries from the PATH that are not needed anymore
	path_elements=(${PATH//:/ })
	new_path="${cuda_path}/bin"
	for p in "${path_elements[@]}"; do
	    if [[ ! ${p} =~ ^${INSTALL_FOLDER}/cuda ]]; then
	        new_path="${new_path}:${p}"
	    fi
	done
	
	# filter out those CUDA entries from the LD_LIBRARY_PATH that are not needed anymore
	ld_path_elements=(${LD_LIBRARY_PATH//:/ })
	new_ld_path="${cuda_path}/lib64:${cuda_path}/extras/CUPTI/lib64"
	for p in "${ld_path_elements[@]}"; do
	    if [[ ! ${p} =~ ^${INSTALL_FOLDER}/cuda ]]; then
	        new_ld_path="${new_ld_path}:${p}"
	    fi
	done
	
	# update environment variables
	export CUDA_HOME="${cuda_path}"
	export CUDA_ROOT="${cuda_path}"
	export LD_LIBRARY_PATH="${new_ld_path}"
	export PATH="${new_path}"
	
	echo "Switched to CUDA ${TARGET_VERSION}."
	
	set +e
	return

• 新建switch-cuda.sh文件，将上边代码写入；

  	vi switch-cuda.sh
  	source switch-cuda.sh
  	source switch-cuda.sh 10.0
  

  
  可以看到当执行source switch-cuda.sh的时候该脚本会扫描所有已安装的CUDA，并列出，用户只需要选择想用的CUDA版本号就可以轻松切换，例如source switch-cuda.sh 10.0，可以看到上图的nvcc也是成功切换了版本。
  并且该脚本基于export 语句，重启终端后，CUDA环境还是会恢复到默认的11.2，不影响下次使用，无需手动切回CUDA版本，下图为重启终端后的效果。
  

---

五、总结

以上就是今天要讲的内容，本文介绍了如何在一台机器上同时安装多个版本的CUDA，并且介绍了一种简便切换CUDA版本的操作。
如果本文能给你带来帮助的话，点个赞鼓励一下作者吧！

六、参考

• [1] CUDA工具包：https://developer.nvidia.com/cuda-toolkit-archive
• [2] cuDNN库：https://developer.nvidia.com/rdp/cudnn-archive
• [3] CUDA切换脚本：https://github.com/phohenecker/switch-cuda
• [4] 安装多版本CUDA：https://blog.csdn.net/sinat_30545761/article/details/107709468