The equation to predict a test instance $\mathbf{z}$ is as follows:
$$f(\mathbf{z}) = \sum_{i=1}^{n_{SV}} \alpha_i y_i \kappa(\mathbf{x}_i, \mathbf{z}) + \rho,$$
where $\alpha$ is the vector of dual weights, $\mathbf{y}$ is the vector of labels of support vectors, $\kappa(\cdot,\cdot)$ is the kernel function, $\mathbf{x}_i$ is the i'th support vector and $\rho$ is the model's bias.
This equation will yield the (signed) distance of $\mathbf{z}$ to the separating hyperplane, which you can then cutoff at a suitable value $\tau$ to obtain a binary label (by default $\tau=0$).