Allocating different arrays on multiple GPUs

Yes, but please add as much relevant information as possible, because I can’t reproduce this issue.

The cu unified memory interface is experimental, I’m not sure about the API yet, is why it isn’t documented.