The default backend uses NVIDIA NCCL. There is another backend that uses Gloo (and also doesn't work.) The backend can be set to gloo at the top of jobs/with_context.sh, by uncommenting the appropriate line (PL_TORCH_DISTRIBUTED_BACKEND).

1. Pulling with Submodules

First pull with

git clone git@github.com:JulesGM/IteratedRetrieval.git --recurse-submodules

to also download the required submodules.

2. Requirements

• Install the requirements in requirements.txt in either a regular (pip) virtual environment or a conda one.

• If you use conda, also install NCCL:

conda install -c conda-forge nccl

3. Launch the job

srun --jobid $SLURM_JOB_ID jobs/sbatch_with_context.sh [[pip] or [conda]] [output_path_name] [path_to_conda_venv_activate_file]

The first agrument after the script path can be either pip or conda according to the type of virtual environment being used.