Check the Driver
Note:
- A machine corresponds to only one nvidia driver, and may correspond to a plurality of CUDA nvidia driver. As long as his path in the cuda designated to a specific version of the line.
- Try not to use / usr / local / following default cuda /, ln prevent because the object will change and impact.
- nvidia driver version determines the available cuda range, thereby determining the available tensorflow-gpu version. So, each time before a new machine, first determine the nvidia driver version.
View nvidia driver version:
cat /proc/driver/nvidia/version
Table 1. CUDA Toolkit and Compatible Driver Versions:
Installation Cuda
View native operating system:
cat /etc/issue
Go head loss Sheriff network, download the corresponding version of cuda:
Note:
- The following case study to cuda9.0.
Installation cuda:
sudo sh cuda_9.0.176_384.81_linux.run
Note:
- Do not install driver! Otherwise, the installation fails!
Into the corresponding path:
vim ~/.bashrc
export PATH=/usr/local/cuda-9.0/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64:$LD_LIBRARY_PATH
source ~/.bashrc
View cuda version, verify that the installation was successful:
nvcc -V
Installation Cudnn
Cudnn download the corresponding installation package.
Note:
- The following order cuda9.0, cudnn7.6.2 example.
Decompression:
mkdir cudnn
tar -xvzf cudnn-9.0-linux-x64-v7.6.2.24.tgz -C cudnn/
Copy the file:
sudo cp cudnn/cuda/lib64/lib* /usr/local/cuda-9.0/lib64/
sudo cp cudnn/cuda/include/cudnn.h /usr/local/cuda-9.0/include/
Set soft links:
cd /usr/local/cuda-9.0/lib64
sudo rm -rf libcudnn.so libcudnn.so.7
sudo ln -s libcudnn.so.7.6.2 libcudnn.so.7
sudo ln -s libcudnn.so.7 libcudnn.so
sudo ldconfig