README

Setup (In SCITAS)

Install libraries

In the root of the project execute those commands to install the environments with the libraries. If needed, all the libraries are listed in the pyproject.toml file

conda create -n saga304 python=3.10 -y
conda activate saga304
pip install --upgrade pip
pip install -e . --upgrade
pip install git+https://github.com/NVIDIA/Cosmos-Tokenizer.git --no-dependencies
python -m ipykernel install --user --name saga304 --display-name "nano4M kernel (saga304)"
conda install -c conda-forge ffmpeg --yes

Train model

Make sure to change the wandb attributes in the .yaml files

wandb_project: ...
wandb_entity: ...
wandb_run_name: ...

And in the root of the project run this command to run any of the configuration files in the ./cfgs/ folder.

sbatch submit_job_multi_node_scitas.sh <config_file> <your_wandb_key>

Run the notebooks

srun -t 120 -A com-304 --qos=com-304 --gres=gpu:1 --mem=16G --pty bash

jupyter lab --no-browser --port=8888 --ip=$(hostname -i)

On your local machine, run:

ssh -L 8888:<IP+Port> -l <SCITAS_username> izar.epfl.ch -f -N

Useful resources

Foundation Models repository

Overleaf Project