jalving
commented
2 weeks ago This is largely addressed by the complete re-write (#105). There are certainly still improvements we could make with how we store node and edge data, but in general I am seeing comparable model build speeds from before with ~100x less memory for large optigraphs.
This issue comes up if a user wants to model individual variables as optinodes to do things like partitioning. The problem with this is that memory use explodes because each optinode creates an empty JuMP model. A possible solution is for us to inspect the hypergraph of an optinode directly and provide the mechanisms to perform and apply partitions.