I have found the solution. CUDA_VISIBLE_DEVICES is only to be interpreted 
by the CUDA library, which in turn maps the device IDs assigned by slurm to 
a continuous device ID range always beginning with 0.

Reply via email to