This article will used smart way install GPU environment.
Why so smart ? Because I used a rather stupid way.
This method is stupid method “deepvariant install”.
Why stupid. Because this way sometime will fall.
This is very angry.
Because we want focus deep learning or other professional field not environment issue.
About machine
We are using the D52G provided by Quanta Computer.
Environment inside:
- 80 vCPU
- 8 V100 GPU
- 768 GiB RAM
Environmental configuration
- Ubuntu 18.04 LTS
- Nvidia driver 440.33.01
- CUDA 10.0
- cudnn 7.6.3
- Tensorflow GPU 1.13
1. you need clean your environment.
If your machine is not Re-irrigation system.
You can skip this step.
apt-get remove --purge '^nvidia-.*'
apt autoremove
apt update
2. install Nvidia driver
Add repository to your apt
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.0.130-1_amd64.debdpkg -i cuda-repo-ubuntu1804_10.0.130-1_amd64.debapt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pubapt updatewget http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.debapt install ./nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
3. Check your ubuntu-drivers
If ubuntu-drvers have nvidia drivers you can direct download.
ubuntu-drivers devices
You can choose autoinstall or own want version.
#autoinstall
ubuntu-drivers autoinstall
Reboot your system
reboot
Testing
nvidia-smi
You can see this figure show CUDA version:10.2.
It’s ok. because It’s wrong.
Finally, it will still be based on your installation.
4. Install CUDA 10 and Cudnn 7.6.3
First you need go this page download cudnn.
Then, go to you download folder.
This command will be easy to install cudnn.
apt install --no-install-recommends cuda-10-0
Add environment variables
vim ~/.bashrc
export PATH=$PATH:/usr/local/cuda/bin/;
source ~/.bashrc
Testing
nvcc -V
Install Cudnn
dpkg -i libcudnn7_7.6.3.30-1+cuda10.0_amd64.deb
dpkg -i libcudnn7-dev_7.6.3.30-1+cuda10.0_amd64.deb
dpkg -i libcudnn7-doc_7.6.3.30-1+cuda10.0_amd64.deb
5. install tensorflow-gpu 1.13.1
pip install tensorflow-gpu==1.13.1
pip show tensorflow-gpu
Testing
import tensorflow as tf
tf.test.gpu_device_name()