Install CUDA 10 on Ubuntu 18.04

install-cuda-10-bionic.sh

# WARNING: These steps seem to not work anymore!

#!/bin/bash

# Purge existign CUDA first
sudo apt --purge remove "cublas*" "cuda*"
sudo apt --purge remove "nvidia*"

# Install CUDA Toolkit 10
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.0.130-1_amd64.deb
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub && sudo apt update
sudo dpkg -i cuda-repo-ubuntu1804_10.0.130-1_amd64.deb

sudo apt update
sudo apt install -y cuda

# Install CuDNN 7 and NCCL 2
wget https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
sudo dpkg -i nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb

sudo apt update
sudo apt install -y libcudnn7 libcudnn7-dev libnccl2 libc-ares-dev

sudo apt autoremove
sudo apt upgrade

# Link libraries to standard locations
sudo mkdir -p /usr/local/cuda-10.0/nccl/lib
sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/
sudo ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/local/cuda-10.0/lib64/

echo 'If everything worked fine, reboot now.'

It might be useful to add this to the very first line:

sudo apt list > ~/apt_list_backup.txt

...in case people need to undo some of the changes.

Finally, to verify the installation, check

nvidia-smi
nvcc -V

sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/

ln: target '/usr/local/cuda/nccl/lib/' is not a directory: No such file or directory

In what point is nvcc installed ?

@kkore

sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/

ln: target '/usr/local/cuda/nccl/lib/' is not a directory: No such file or directory

In what point is nvcc installed ?

I did "sudo apt-get cuda toolkit" before this script and then it worked out.

Very useful! Thanks

I needed to downgrade CUDA from10.2 to 10.0 version because Pytorch 1.5.1 does not support Tesla 40 GPUs...

I reinstalled Pytorch 1.2.0 with:
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch

Important steps before Pytorch installation:

1. Use sudo apt install cuda=10.0.130-1 instead sudo apt install cuda
2. Don't use sudo apt upgrade
3. Include in ~/.bashrc this line: export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}
4. Warning: nvidia-smi shows 'CUDA Version 11.0' but v10.0 is really installed and working correctly... I don't know why it shows a different version...

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.36.06    Driver Version: 450.36.06    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla K40c          On   | 00000000:03:00.0 Off |                    0 |
| 28%   57C    P0    66W / 235W |   1058MiB / 11441MiB |      0%      Default |
|                               |                      |                  N/A |

Very useful! Thanks

I needed to downgrade CUDA from10.2 to 10.0 version because Pytorch 1.5.1 does not support Tesla 40 GPUs...

I reinstalled Pytorch 1.2.0 with:
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch

Important steps before Pytorch installation:

1. Use `sudo apt install cuda=10.0.130-1` instead `sudo apt install cuda`

2. Don't use `sudo apt upgrade`

3. Include in  ~/.bashrc   this line:   `export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}`

4. Warning: `nvidia-smi` **shows 'CUDA Version 11.0'  but v10.0 is really installed and working correctly**... I don't know why it shows a different version...

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.36.06    Driver Version: 450.36.06    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla K40c          On   | 00000000:03:00.0 Off |                    0 |
| 28%   57C    P0    66W / 235W |   1058MiB / 11441MiB |      0%      Default |
|                               |                      |                  N/A |

Worked thanks!!!

Here is the version that works for me. Credits to @jpison and @bogdan-kulynych.

#!/bin/bash

# Purge existign CUDA first
sudo apt --purge remove "cublas*" "cuda*"
sudo apt --purge remove "nvidia*"

# Install CUDA Toolkit 10
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.0.130-1_amd64.deb
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub && sudo apt update
sudo dpkg -i cuda-repo-ubuntu1804_10.0.130-1_amd64.deb

sudo apt update
sudo apt install -y cuda=10.0.130-1

# Install CuDNN 7 and NCCL 2
wget https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
sudo dpkg -i nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb

sudo apt update
sudo apt install -y libcudnn7 libcudnn7-dev libnccl2 libc-ares-dev

sudo apt autoremove
# sudo apt upgrade

# Link libraries to standard locations
# sudo mkdir -p /usr/local/cuda-10.0/nccl/lib
# sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/
# sudo ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/local/cuda-10.0/lib64/

echo 'export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}' >> ~/.bashrc
echo 'If everything worked fine, reboot now.'

• Since we are using Deb installation method, we don't need to change the LD_LIBRARY_PATH variables (reference).
• After running this script, you need to reboot the system.
• Type nvidia-smi and nvcc --version to verify your installation.

Whole day I had been looking for a way to install cuda 10.0, but every methods I found were ended up with the message, "cuda : Depends: cuda-10-0 (>= 10.0.130) but it is not going to be installed".
I am quite sure above procedures are only way to install cuda 10.0 version that tensorflow-gpu==1.14 requires on "Ubuntu 18.04.5 LTS".
The very very important thing is that never install "nvidia-driver-***" driver by yourself.
Required nvidia drivers are installed while doing sudo apt install -y cuda=10.0.130-1

In addition, for me following commands are not necessary.

# It seems no need following:
sudo mkdir -p /usr/local/cuda-10.0/nccl/lib
sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/
sudo ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/local/cuda-10.0/lib64/

Here shows my driver is running that twice faster than tensorflow CPU does.

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.06    Driver Version: 450.51.06    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 960M    On   | 00000000:01:00.0 Off |                  N/A |
| N/A   55C    P0    N/A /  N/A |   3972MiB /  4046MiB |     66%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1133      G   /usr/lib/xorg/Xorg                139MiB |
|    0   N/A  N/A      1388      G   /usr/bin/gnome-shell              120MiB |
|    0   N/A  N/A     17390      C   python                           3706MiB |
+-----------------------------------------------------------------------------

Whole day I had been looking for a way to install cuda 10.0, but every methods I found were ended up with the message, "cuda : Depends: cuda-10-0 (>= 10.0.130) but it is not going to be installed".
I am quite sure above procedures are only way to install cuda 10.0 version that tensorflow-gpu==1.14 requires on "Ubuntu 18.04.5 LTS".
The very very important thing is that never install "nvidia-driver-***" driver by yourself.
Required nvidia drivers are installed while doing sudo apt install -y cuda=10.0.130-1

In addition, for me following commands are not necessary.

# It seems no need following:
sudo mkdir -p /usr/local/cuda-10.0/nccl/lib
sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/
sudo ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/local/cuda-10.0/lib64/

Here shows my driver is running that twice faster than tensorflow CPU does.

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.06    Driver Version: 450.51.06    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 960M    On   | 00000000:01:00.0 Off |                  N/A |
| N/A   55C    P0    N/A /  N/A |   3972MiB /  4046MiB |     66%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1133      G   /usr/lib/xorg/Xorg                139MiB |
|    0   N/A  N/A      1388      G   /usr/bin/gnome-shell              120MiB |
|    0   N/A  N/A     17390      C   python                           3706MiB |
+-----------------------------------------------------------------------------

I have a Tesla K80 (Linux) GPU and I want to use TF 1.14.
I have a dev GPU in which I have installed driver 435.21 and CUDA 10; I can see the PID's and memory spike here. However, on a different k8 pod which comes with preinstalled nvidia-driver 450, I am not able to see any processes when running nvidia-smi (CUDA 10).

I am using the same Docker image to install the CUDA,cUDNN, and tensorflow_gpu==1.14.0, which worded with driver version 435.21.

Does anyone have any idea what is going wrong?

this crashed my server..

sudo ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/local/cuda/nccl/lib/

ln: target '/usr/local/cuda/nccl/lib/' is not a directory: No such file or directory

In what point is nvcc installed ?

You can have nvcc by install nvdia cuda toolkit: "sudo apt install nvidia-cuda-toolkit". Now you can check the directory again. good luck

While doing sudo apt install -y cuda=10.0.130-1 I ran into a error: cuda : Depends: cuda-10-0 (>= 10.0.130) but it is not going to be installed. Running sudo dpkg --force-all -i cuda-repo-ubuntu1804_10.0.130-1_amd64.deb fixed it for me. Note, though, that that will overwrite any existing versions.