N_PER_RNG values

I have a question regarding the RNG on CUDA. If for example one has to generate lets say 1000 pseudo random numbers. Assuming, MT_RNG_COUNT=4096, we have N_PER_RNG=ceil(1000/4096)=1. My question is, for this small N_PER_RNG, does one has to use local seeds for each thread. If so, how? Thanks!