DL Exercises¶

Info

We put some exercises here for you, if you want to get some more hands-on.

Prepare your project folder¶

Make arrangements for the new project

Find your way into your project uppmax2025-3-5 by logging in to Rackham by ThinLinc/ssh/VSCode.
Go to private folder and make an empty folder with your name under private or nobackup/private

Answer

ssh jayan@rackham.uppmax.uu.se
ssh -X jayan@rackham.uppmax.uu.se
cd /proj/uppmax2025-3-5
mkdir
Copy contents from uppmax_workshop to your private folder.

Create environment¶

Use venv or conda to prepare your environment

Module load python and use venv to create a environment with pytorch, transformers and jupyter.
Answer
```
module load python/3.12.7
which python
python -V
python -m venv torch_env
```
If using conda, follow instructions from "Use isolated environments" section.
Activate your env and install desired packages.

Using the compute nodes¶

Submit a Slurm job

Submit sentiment_analysis.py with appropriate flags for its slurm job.

Answer

edit a file using you prefered editor, named sentiment_analysis_batch.sh, for example, with the content

    #!/bin/bash -l

    #SBATCH -A uppmax2025-3-5
    #SBATCH -p node
    #SBATCH -N 1
    #SBATCH -t 01:00:00
    #SBATCH -J sentiment_analysis
    #SBATCH -M snowy
    #SBATCH --gres=gpu:1

    source .....

    python -c "import torch; print(torch.__version__); print(torch.version.cuda); print(torch.cuda.get_device_properties(0)); print(torch.randn(1).cuda())"

    echo "running sentiment_analysis.py"

    python .....sentiment_analysis.py

make the job script executable, if not already.
```
$ chmod a+x sentiment_analysis_batch.sh
```
submit the job
```
$ sbatch sentiment_analysis_batch.sh
```

Similarly run sentiment_analysis.ipynb jupyter notebook. Add matplotblib and seaborn to your environment. Start a jupyter server on snowy compute node and then tunnel to the host from your local.
Answer
- Install matplotlib and seaborn and start an interactive snowy session:
```
    source torch_env/bin/acivate
    pip install matplotlib seaborn
    interactive -A uppmax2025-3-5 -M snowy -p node -N 1 -t 1:01:00 --gres=gpu:1
    source torch_env/bin/acivate
    jupyter notebook --ip 0.0.0.0 --no-browser
```
- Then tunnel:
ssh -L 8888:s123:8888 username@rackham.uppmax.uu.se
- Copy the localhost url and paste it in your browser

Cache management¶

By default HF will install the models and temp files to yur $HOME folde which is rather limited to 32 GB and 300k files.
To avoid that, you can set HF_HOME and HF_HUB_CACHE variables to your project folder. Follow instructions on HuggingFace's documentation.