User cluster nodes have GPU but have no an internet access. I can install julia packages using a login node but I have no possibility to download when I run code on cluster nodes with GPU.
As far as I found out in message boards - It is not so rare situation
In my case, an unsuccessful attempt to download CUDA_compat takes about 20 additional seconds of compilation time. Thus, I prefer to switch it off. I think that correct solution is just to ask the cluster admin to update CUDA driver, but it could take a while