Skip to content

Latest commit

 

History

History
176 lines (121 loc) · 5.02 KB

local-ubuntu.md

File metadata and controls

176 lines (121 loc) · 5.02 KB

Setting up Ubuntu with GPUs for machine learning

This is our setup for linux machines for machine learning. Installing nvidia drivers became a lot easier with Ubuntu 20.04 onwards and we use miniconda to manage CUDA installations. CUDA versions matter because PyTorch and Tensorflow support only older versions.

We use terminator as the terminal, and oh-my-zsh. It gives indicators about the git branch, uncommitted changes, and conda environment on the shell.

Z Shell

Here's the setup procedure, up to installing PyTorch. This takes about an hour depending on your download speeds.

sudo apt-get update
sudo apt-get upgrade

If you ran this on a fresh Ubuntu 20.04 installation it would upgrade the kernel and you will be required to restart the computer.

Install vim and git.

sudo apt-get install vim-gtk3
sudo apt-get install git gitk

Edit grub file [Optional]

I usually edit grub so that I can see the boot menu and troubleshoot (load ubuntu in recovery mode) if things go bad. This was important before Ubuntu 20.04 when driver installations occasionally failed and I had to recover by going into recovery mode. This step is not required.

sudo vim /etc/default/grub

Update the following settings. This gives a 5-second boot menu, and shows all the ubuntu startup logs instead of the nice splash screen.

GRUB_TIMEOUT=5 # Wait 5 seconds
GRUB_TIMEOUT_STYLE=menu # Display the menu 
GRUB_CMDLINE_LINUX_DEFAULT = '' # remove quiet and splash so that you can see details at bootup

Update grub for changes to take effect.

sudo update-grub

You can restart the computer to see if this worked.

Install Drivers

Next, we want to install the nvidia drivers. You can check if you already have them installed by running nvidia-smi.

sudo apt-get install --no-install-recommends nvidia-driver-450

--no-install-recommends doesn't install packages recommended as dependencies. Only required dependencies will be installed. This does a lean driver installation. You can install a version higher than 450 if available. You should restart the computer for the driver installation to take effect.

Run the following after the restart to make sure drivers are installed. It should show the driver version and a list of GPUs on the computer.

nvidia-smi

Install terminator.

sudo apt-get install terminator

Quit the terminal and start the terminator. You might want to add it to favorites for quick access.

Install Z Shell and oh-my-zsh

sudo apt-get install zsh
sh -c "$(wget https://raw.github.com/ohmyzsh/ohmyzsh/master/tools/install.sh -O -)"

Restart terminator to see if it works.

Install miniconda

Now that shell is setup we install miniconda. Miniconda lets us manage different python environments - even with different versions of python. It also takes care of cuda versions.

Download miniconda.

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh

Install miniconda to ~/miniconda.

bash miniconda.sh -b -p ~/miniconda

Next, we need to set conda paths on oh-my-zsh. We do this by adding an oh-my-zsh plugin. Create a file ~/.oh-my-zsh/conda.zsh and include this in it.

source ~/miniconda/etc/profile.d/conda.sh

Restart terminator again for it to take effect. You can check conda installation by running conda -V.

Create a conda environment

Next, we create a separate conda environment for our deep-learning work. You can have many environments. See miniconda guide for details on how to manage conda environments.

Let's create an environment for our PyTorch work, and name it myenv.

conda create -n myenv "python=3.8"

If you add the option -y it won't ask for confirmation.

Activate our new conda environment.

conda activate myenv

Install PyTorch and other packages

Let's install PyTorch latest version with cuda.

conda install pytorch pytorch-cuda=11.8 -c pytorch -c nvidia

Install other packages like Jupyter Notebooks.

pip install setuptools numpy opt_einsum sentencepiece jupyterlab ninja einsum ipywidgets matplotlib seaborn

You can check torch installation with torch.cuda.device_count(). Run ipython and try the following,

>>> import torch
>>> torch.cuda.is_available()
>>> torch.cuda.device_count()

Create SSH Keys

Create public/private keys with ssh-keygen:

ssh-keygen -t rsa -b 4096 -C "[email protected]"

Your public key will be located at ~/.ssh/id_rsa.pub by default - ssh-keygen will ask you to confirm this path. Copy the contents of the public key file to Github/Bitbucket SSH keys to access your repositories.