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。
-
查看已有CUDA环境
 -
从官网下载CUDA 10.0的runfile到服务器上。
 -
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：
- 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。
-
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)。
 -
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
-
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/*
-
查看cuDNN版本，指令为cat /usr/local/cuda-10.0/include/cudnn.h | grep CUDNN_MAJOR -A2。
 -
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
-
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
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
|
