I believe the documentation is missing regarding the C++ memory management API. I opened an issue but was told to post here instead.
From issue 85436:
I assume there should be a C++ memory management API like Python, since Issue 14618 to do this was marked as completed. Does this API exist or no?
If not do I need to get the CUDA caching allocator and read from the DeviceStats
struct, e.g. here? I’m assuming this isn’t exactly future proof so I would rather not do this unless I have to.