CUDA in multithreaded application

I was searching for information on OptiX multithreaded support. I read this 2013 post:

So, I am aware that CUDA is not multithread safe. I still have one question about this following line:
OptiX should therefore only be used from within a single host thread

My question is: Should OptiX API be called only from a unique thread or can be called from multiple threads as long as the API calls are serialized via some synchronization mechanism. I had assumed the latter as there doesn’t logical for OptiX to enforce that only a unique thread from the app invokes it. What is the correct way ?