Cuda Error 300 when using shared libraries

Accelerated Computing CUDA CUDA Programming and Performance
,
1

I have 2 shared libraries , built with cuda toolkit 11.8. (7.5,8.0) one is called dXX and the other one is called sXX (for example). Running on T4
Another program is using library sXX. When we add library dXX we get Error 300
when the program (itself - not the libraries) calls cudaMemcpy , CudaHostMalloc…

Assuming we are not calling any function from dXX

  1. If dXX is not defined in the cuda setting libraries - OK
  2. If dXX is defined but after sXX - OK
  3. If dXX is defined before sXX - Error
  4. If dXX is defined and we call a method of dXX - Error

we have reduce dXX to a single cpp file with one method and its still happens

any idea???