copy-on-write (COW) fork() & performance

[SpitchAG]

Hello,

First, bravo for the library!

Question: is it good practice to share runtime session between copy-on-write (COW) fork() processes (i.e the Session is created/loaded in the main process, then children run the session (copied from main proc memory) ?

Asking because for some models (heavily using MLasgemmfloatkernelfma3),
if we set the PARALLEL session options when loading session on main process, children processes freeze when runing the session.
If we set SEQUENTIAL, children can run the session, but perf is worth.