For cusolverDnSgesvdaStridedBatched_bufferSize, the lwork should be long long instead of int.
I think there are cases where there will be enough memory on the GPU card to do the computation, but the computation of lwork will overflow an int.
Keep your eye out for the next release.