Open
Description
I forgot to open an issue to track an offline discussion I had with @pciolkosz...
Currently, cuda.core
diverges with cudax
in a few places, and this is one of them. cuda.core
focuses on enabling integration with existing Python projects, and having the cudart-like .set_current()
method fills in the gap. However, the two projects also share many design principles, one of which is to favor explicitness, and therefore ideally we do not want any implicit/global states.
This issue tracks if there is any design opportunity to move away from the global device state, without breaking existing libraries. If such design exists, we should raise a deprecation warning asap while working toward the new design.