After doing more testing it seems that the amount of memory allocated after calling torch.cuda.init() depends on the numbers of GPUs visible. (With 1xP100 about 10GB is allocated. With 2xP100 about 18GB is allocated…)
I’m afraid not
The fact that it uses a lot of virtual memory is expected because it use that for GPU memory management from what I remember. And you should have plenty of virtual memory to spare anyway
The fact that it uses 2GB of RAM is a bit but I’m not sure what is the root reason for it.