In my code I have a target y which is computed using some Linear and Conv2d modules, and I can't do y.requires_grad = False before I pass it to a loss module.

Problem comes from:

Is there any particular reason why this restriction was implemented?

Also I was wondering what would happen if we have something like:
x -> var1 -> var2 -> var3 -> var_loss

and we call loss.backward() when every variable requires a gradient, apart from var2, for instance? Would it be like if var1.requires_grad is also set to False?