Set CUDA_VISIBLE_DEVICES if GPU count mismatch
If the count of GPUs configured in slurm.conf and gres.conf differ and FastSchedule>=1 then the bitmap identifying the GPU allocation sent from slurmctld to slurmd will differ. Previously this resulted in CUDA_VISIBLE_DEVICES being set to NULL. Now it will be set correctly. bug 6725
Please register or sign in to comment