Its been a while I last used CUDA , but I doubt it. I think e.g. OpenCL supports something like this.

I have not tried that so far. But, as per this page, there is an option called --device-c (-dc).

http://docs.nvidia.com/cuda/nvrtc/index.html#axzz4GwvjLyCt

It is not really clear to me , but if the function can be part of an object (class /structure) then yes you can override  it. Additionally, if the difference is data type,  you may define your function using template.  

The function is a normal __device__ function. It is not part of a class nor structure. The difference is not the data type, so I cannot rely on templates. I actually must change the calculations performed in the function (i.e., propensity calculations) according to the model that I am simulating.

I am adding this information in the original post.

for override part:

I had conducted basic code  test and I found it does support override you can use the following to test it as well:

define in your cuh (header file )

__device__ int test_Max_override(int, int );

__device__ int test_Max_override(int, int,int );

in your implantation file 

__device__ int test_Max_override(int x, int y){

if(x>y)return x;

return y;

}

__device__ int test_Max_override(int x, int y, int z){

if(x>y && x>z) return x;

if(y>x && y>z) return y;

if(z>x && z>y) return z;

}

while it is defined as __device__ than it should be called from global or other __device__ function i.e. you can not call it from host function. 

I assume this override with number of arguments; it will also work with different arguments  data type by modifying one of above with different datatype . 

Related to global variables performance you may explore using __shared__ which will be within the scope of the grid-block.

I am not for "real-time compilation (e.g., using NVRTC) to dynamically create a "__device__" function"  point if I understand it correctly, I assume you can use system call system(" YOUR EXEC COMMAND")from your code to execute NVRTC with your cuda files as work around. 

I hope this answer your question.