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.
- [julia-users] CUDA package and multiple GPUs oliver . mirus
- [julia-users] Re: CUDA package and multiple GPUs Oliver Mirus
- Re: [julia-users] Re: CUDA package and multiple ... Stefan Karpinski
