Running Modern LLMs on a 2GB GPU

I've got boxes of mismatched old parts rather than one built machine — any rule of thumb for what's worth assembling vs. what's too old to bother with for this kind of use case?