This is super cool. If you're interested in this, you might enjoy this proof of concept as well: https://github.com/deepai-org/omnivm
It instead embeds a bunch of runtimes onto the same OS thread.