The problem of the thickness variable being ineffective when using torch for optimization

I found that when using the** Torch backend** for optimization, when adding multiple thickness variables with problem.add_variable, usually only one is effective. And the other geometry properties remain unaffected. Additionally, directly changing lens.surfaces.surfaces[n].thickness to a tensor also cannot participate in gradient updates, while the geometry attribute does not have this issue.Is this due to my mistake or a bug?