How to reduce the pre-allocated memory for a CUDA programm

Question

I am working on a MPI code written in C++. This code invokes some functions resided in a separate shared library during runtime. The code in the shared library was developed