似乎CuDNN安装程序正在寻找错误版本的CUDA。我究竟做错了什么?完整的故事:
Ubuntu 16.04
已安装两个版本的CUDA,分别为9.0和9.1。 / usr / lib / cuda链接到9.1安装,并且LD_LIBRARY_PATH似乎指向该目录:
// Use this or injected parameter
var loader = ServiceLocator.Current.GetInstance<IContentLoader>();
// Remove all children or not
var cascade = true;
ContentReference entryPoint = ...where you want to start
// Get all indexable languages from Find
Languages languages = SearchClient.Instance.Settings.Languages;
// Remove all current instances of all languages below the selected content node
//languages.ForEach(x => ContentIndexer.Instance.RemoveFromIndex(entryPoint, cascade.Checked, x.FieldSuffix));
foreach (var lang in languages)
{
if (cascade)
{
var descendents = loader.GetDescendents(entryPoint);
foreach (ContentReference descendent in descendents)
{
ContentIndexer.Instance.RemoveFromIndex(descendent, false, lang.FieldSuffix);
}
}
// Try delete the entrypoint
var entryTest = loader.Get<IContent>(entryPoint, new CultureInfo(lang.FieldSuffix));
if (entryTest != null)
{
var delRes = ContentIndexer.Instance.Delete(entryTest);
}
}
现在我确实已经安装了CuDNN 7.0,想将其替换为7.1版。 (这可能是一个不明智的决定,但是,我已经开始这样做了!)
下载了用于CuDNN 7.1.3的debian安装程序,以与CUDA 9.1一起使用。试图通过以下方式安装它:
$ echo $LD_LIBRARY_PATH
/usr/local/cuda/lib64:/usr/local/lib
$ ls -l /usr/local
lrwxrwxrwx 1 root root 8 jan 22 2018 cuda -> cuda-9.1
drwxr-xr-x 15 root root 4096 dec 1 2017 cuda-9.0
drwxr-xr-x 15 root root 4096 jan 22 2018 cuda-9.1
如您所见,安装程序在库文件之后查找cuda-9.0目录。 (为什么?)文件存在,但它不是符号链接:
$ sudo dpkg -i libcudnn7_7.1.3.16-1+cuda9.1_amd64.deb
(Reading database ... 261910 files and directories currently installed.)
Preparing to unpack libcudnn7_7.1.3.16-1+cuda9.1_amd64.deb ...
Unpacking libcudnn7 (7.1.3.16-1+cuda9.1) over (7.1.3.16-1+cuda9.1) ...
Setting up libcudnn7 (7.1.3.16-1+cuda9.1) ...
Processing triggers for libc-bin (2.26-0ubuntu2.1) ...
/sbin/ldconfig.real: /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so.7 is not a symbolic link
我也尝试删除libcudnn,但是在第一次重新安装后我这样做了,因此删除会引发相同类型的错误:
$ ls -l /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so
-rwxr-xr-x 1 root root 287624224 jan 16 2018 /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so
因此,更大的图景:我的目标是正确安装CuDNN 7.1。我该如何实现该目标?
答案 0 :(得分:1)
我终于做到了!我进去并手动删除了.so文件:
$ sudo apt-get remove libcudnn7
Reading package lists... Done
Building dependency tree
Reading state information... Done
The following packages were automatically installed and are no longer required:
cuda-9-0 cuda-command-line-tools-9-0 cuda-core-9-0 cuda-cublas-9-0
cuda-cublas-dev-9-0 cuda-cudart-9-0 cuda-cudart-dev-9-0 cuda-cufft-9-0
cuda-cufft-dev-9-0 cuda-curand-9-0 cuda-curand-dev-9-0 cuda-cusolver-9-0
cuda-cusolver-dev-9-0 cuda-cusparse-9-0 cuda-cusparse-dev-9-0
cuda-demo-suite-9-0 cuda-documentation-9-0 cuda-driver-dev-9-0
cuda-libraries-9-0 cuda-libraries-dev-9-0 cuda-license-9-0
cuda-misc-headers-9-0 cuda-npp-9-0 cuda-npp-dev-9-0 cuda-nvgraph-9-0
cuda-nvgraph-dev-9-0 cuda-nvml-dev-9-0 cuda-nvrtc-9-0 cuda-nvrtc-dev-9-0
cuda-runtime-9-0 cuda-samples-9-0 cuda-toolkit-9-0 cuda-visual-tools-9-0
libatk-wrapper-java libatk-wrapper-java-jni python-cliapp python-markdown
python-ttystatus
Use 'sudo apt autoremove' to remove them.
The following packages will be REMOVED:
libcudnn7 libcudnn7-dev
0 upgraded, 0 newly installed, 2 to remove and 18 not upgraded.
After this operation, 680 MB disk space will be freed.
Do you want to continue? [Y/n]
(Reading database ... 261909 files and directories currently installed.)
Removing libcudnn7-dev (7.1.3.16-1+cuda9.1) ...
update-alternatives: removing manually selected alternative - switching libcudnn to auto mode
Removing libcudnn7 (7.1.3.16-1+cuda9.1) ...
Processing triggers for libc-bin (2.26-0ubuntu2.1) ...
/sbin/ldconfig.real: /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so.7 is not a symbolic link
然后使用dpkg进行安装,然后重新安装我的tensorflow-gpu,从而确认一切正常:
$ cd /usr/local/cuda-9.0/targets/x86_64-linux/lib/
$ sudo mv libcudnn.so bak_libcudnn.so
$ sudo mv libcudnn.so.7 bak_libcudnn.so.7
$ sudo mv libcudnn.so.7.0.5 bak_libcudnn.so.7.0.5