Discrete GPU usage

CUDA_VISIBLE_DEVICES?
See SLURM Job Submission: Configuring node count
for an example were idle GPUs are assigned to CUDA_VISIBLE_DEVICES

and another user’s opinion on the subject: