Via ssh ssh firstname.lastname@chimera.umb.edu
Every CS faculty has a cs_* account like cs_daniel.haehn. You can request the H200 resources like this:
salloc -A cs_daniel.haehn -p pomplun --mem=32G -w chimera21 --gres=gpu -t 60
This will grab the next available GPU with at least 18GB GPU memory for 60 minutes. In total, 26 users can request their own GPU device with at least 18GB GPU memory.
If you need specific GPU memory, you can specify the following:
--gres=gpu:1g.18gb (18GB GPU memory, 12 available)
--gres=gpu:2g.35gb (35GB GPU memory, 6 available)
--gres=gpu:3g.71gb (71GB GPU memory, 6 available)
--gres=gpu:h200 (full H200 GPU with 131GB GPU memory, 2 available)
Here I requested 2x 35GB GPU memory and 128GB RAM with 2 CPU cores and a timelimit of 3 hours:
salloc -A cs_daniel.haehn -p pomplun --mem=128G -w chimera21 -c2 --gres=gpu:2g.35gb:2 -t 180
We can verify with nvidia-smi:
For Tensorflow, the following conda environment works.
- Download and install miniconda locally in your homefolder.
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod +x Miniconda*
./Miniconda*
- After re-login, create the following environment.
conda create --name H200 python=3.10
conda activate H200
pip install "tensorflow[and-cuda]"
pip install jupyterlab ipykernel
-
Activate the environment.
conda activate H200 -
And configure some paths.
This has to happen only once!
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
nano $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
And then in the editor, paste the following:
export LD_LIBRARY_PATH=$(python - <<'PY'
import site, os, glob
paths=[]
for p in site.getsitepackages():
paths += glob.glob(os.path.join(p,"nvidia","*","lib"))
print(":".join(paths))
PY
):$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
Save and exit.
And then reload the environment.
conda deactivate
conda activate H200
python -c "import tensorflow as tf; print(tf.config.list_logical_devices('GPU'))"
This should list the GPUs that are available!
...is available here: https://www.umb.edu/rc/hpc/
Have fun :)