It would be nice if we could get GPU by UUIDs or GPUs by indices.
if you use an UUID (e.g. export CUDA_VISIBLE_DEVICES="GPU-08041152-9141-4970-a445-5e8b634025b5" the file 'source/train/Local.py ' will exit with an error because the code try to cast the uuid to an integer value.
Could you please support also GPU device UUID ? Newest HPC schedulers use gpu uids instead of integer indexes.