In my case, I had to add the feature constraint

#SBATCH --gres=gpu

in my SLURM script for the supercomputer I am using, so that the node the job gets assigned to, is guaranteed to have an NVIDIA GPU.