MAGMA 1.0 RC3 has most of the components needed.
First the B matrix is factored using a Cholesky factorization and we have that for the GPUs.
Then the factors are applied to A, we have that for GPUs.
Then the resulting matrix is reduced to tridiagonal form, we have that for GPUs.
Then fined the eigenvalues of the tridiagonal matrix, we don't have that for the GPU but its not too expensive to compute on the host.
Finally, there are back transforms, for the vectors and we have these as well for GPUs.
At the moment you would have to put together these pieces from MAGMA, similarly to the way we have already done it for the standard eigen-problem.