Why is it okay to normalize the Q_uniform matrix by dim?

[AliceOfSNU]

in graph_lib.py's Uniform graph you have:

def rate(self, i):
      edge = torch.ones(*i.shape, self.dim, device=i.device) / self.dim
      edge = edge.scatter(-1, i[..., None], - (self.dim - 1) / self.dim)
      return edge

where you are normalizing $Q^{tok}$ by self.dim to avoid blowup.
doesn't this effect the reverse sampling probabilities given in the paper by
$\delta_{x_t^i}(x^i_{t-\Delta t }) + \Delta t Q_t^{tok}(x_t^i,x_{t-\Delta t}^i)s_\theta(x_t, t)$?