Why does the gradient free SVGD algorithm not converge?

Question

I am trying to implement the following algorithm given in this paper:

Here is my code:

import numpy as np
import torch
i