Setup procedure for HAL DGX

Setup

1. After cloning, first migrate the HOME area from the /home area to under /raid/projects

$ bash migrate_home.sh

Note: Before moving onto step 2 you will need to update your current environment. You can do this either by logging out and then logging back in, or by sourcing your /raid/projects/${USER}/.bash_profile.

2. In your /raid/projects/${USER}/.bash_profile make sure that the following comment exists anywhere above the sourcing of ~/.bashrc to set the start location for adding pyenv information to your .bash_profile

# pyenv setup

and then run

$ bash pyenv_setup.sh

3. Source your .bash_profile to make sure everything is setup and then install CPython v3.8.11 from source with optimizations using pyenv

$ . ~/.bash_profile
$ bash build_python.sh

Test

As a simple test that everything is working okay:

• Create a pyenv virtual environment

$ pyenv virtualenv 3.8.11 base

• Activate the base virtual environment

$ pyenv activate base

• Install some dependencies with pip

(base) $ python -m pip install --upgrade pip setuptools wheel
(base) $ python -m pip install matplotlib

• Run a test example Python script

(base) $ python test/example.py

which should produce a plot named mpl_example.png in your current working directory.

Optional Conda Install for MLFlow

MLFlow works best with Conda for managing environments. pyenv has the ability to install Conda distributions as well, so you can pyenv install whatever distribution you'd like (c.f. output of pyenv install --list | grep conda)

$ pyenv install miniconda3-latest

To be able to use Conda for package management with or without pyenv, use the installed miniconda distribution to initialize Conda

$ pyenv shell miniconda3-latest
$ conda init
$ conda config --set auto_activate_base false

and after a shell restart conda will now be found and useable.

You can treat the miniconda version that you've selected as you would any other pyenv version when creating a pyenv virtual environment.

$ pyenv virtualenv miniconda3-latest mlflow-base

These Conda pyenv environments can be activated with either pyenv activate or conda activate, though it is recommended to only use pyenv activate as switching between pyenv environments with a conda environment activated will not update the Python runtime used.

Note that it is important to make sure the auto_activate_base false command is run — which results in the following being added to your .condarc

$ grep auto_activate ~/.condarc
auto_activate_base: false

— to ensure that there won't be conflict between Conda environment Python runtimes and any other virtual environment that you have. As conda init will still place condabin onto PATH you will not need to update MLFlow's MLFLOW_CONDA_HOME shell variable.