OOD Jupyter Notebook Python with virtualenv and GPU

[gtdang]

How do I configure OOD Jupyter Notebook Python with virtualenv and GPU?

[gtdang]

OOD Jupyter Notebook Python with virtualenv and GPU

Tags: GPU, Jupyter Notebook, OOD, Python, virtualenv

Set up for an OOD Jupyter Notebook with python virtual environment w/ tensorflow installation and GPU set-up.

Several students from Deep Learning for Scientists and Engineers (APMA-2070 / ENGN 2912) came in for help on 3/10/23 for help with this set-up.

1. GPU setup on Oscar

   1. GPU setup is a one-time set-up and should be available to all of your future sessions to load.
   2. In the OSCAR Shell Access enter the following in order:

   interact -q gpu -g 1 -f ampere -m 20g -n 4  # Request an interactive/batch session on an ampere GPU node
   module load python/3.11.0  # alternatively, 3.9.0 
   module load cuda/11.3.1
   module load cudnn/8.6.0
   module load openssl/3.0.0

   b. Modified from the documentation at the following page: https://docs.ccv.brown.edu/oscar/gpu-computing/ampere-architecture-gpus

2. Create a python virtual environment using virtualenv and install tensorflow

   cd ~
   virtualenv -p python3 <myenv>
   source <myenv>/bin/activate 
   pip install tensorflow

   • Install any other packages you may need with pip while your environment is activated.
   • You can start from this point if you need to create new environments after the first time you set up the GPU

3. Install jupyter notebook and create the jupyter kernel

   pip install notebook
   pip install ipykernel
   python -m ipykernel install --user --name=<myenv> 

   • Where <myenv> is the name of your environment.

   • https://docs.ccv.brown.edu/oscar/connecting-to-oscar/open-ondemand/using-python-or-conda-environments-in-the-jupyter-app

4. Back at the OOD dashboard start a request for a Jupyter Notebook for Python Environments GUI and fill in the following fields.

   1. Python Module on Oscar: Select the python version you loaded in step 1 (python/3.7.4 for this example)

   2. Python Virtual Environment: Use the full path to <myenv> set up in step 2. It should look something like /users/<USERNAME>/<myenv>

   3. Modules: List all the modules loaded in step 1 separated by a space.

      python/3.11.0 cuda/11.3.1 cudnn/8.6.0 openssl/3.0.0

   4. Partition: gpu

Links

1. https://docs.ccv.brown.edu/oscar/gpu-computing/ampere-architecture-gpus
2. https://docs.ccv.brown.edu/oscar/connecting-to-oscar/open-ondemand/using-python-or-conda-environments-in-the-jupyter-app
3. https://docs.ccv.brown.edu/oscar/software/python-instal

[singhsaluja]

Thanks! This looks good, very nice writeup.

Step 1 should be, to request an interactive/batch session on an ampere GPU node. The environment built on Ampere is backwards compatible across older GPUs (mostly except for TensorRT functions)
interact -q gpu -g 1 -f ampere -m 20g -n 4

[gtdang]

Great! I'll add that. Thank you!

[ntolley]

@gtdang apologies to be digging up such an old question 😄

I'm attempting to use jupyter notebooks with the code server but get this message:

You ever run into a similar issue? It's odd as I can definitely select the kernel corresponding to my custom conda environment, but fails when I try to run a cell

[ntolley]

Nevermind! Turns out you need to specify loading the anaconda module when you request the code server (not in the terminal afterwards)