What is a good way to use grpc c++ with each thread having its own resource objects?

Question

I am a newbie to grpc c++, using it to deploy deep learning models. Deep learning models need an inference engine to be initialized and then to process incoming requests. The time to initialize can be very long, so it is reasonable to initialize only once and reuse it for every request. As there is only one Service object and the inference engine is not thread-safe, I need every thread to own an engine object. How to achieve that with sync or async API?