If someone wants to help, we need test cases for the cuda mpz functions to be sure that the cuda mpz library works (it's not mine). Also debugging this with gdb (the one from nvidia that can step through gpu execution) would be insightful, to see at which code line it crashes.
I can try debuggin it on linux with cuda 5.5
Btw why didnt you use cump library? it seems to be gmp (mpz) compatible