it comes with its own temporary friction, torch.export is still maturing for backward/training graphs and all effort over the past two years was put on inference maturity

→ View original post on X — @soumithchintala