D = torch.sparse.sum(A, dim=1).to_dense() is no right

[scuChenscu]

because A has normalized, D would be a tensor with all 1.