Can you clarify what you mean by failing to update rho? Do you mean it has failed to actually solve the problem when adaptive rho is enabled, or that there is a different error returned?

Moving from doubles to floats will change how the solver converges and behaves, especially if the convergence criteria is left the same, since ADMM can be more susceptible to the reduced precision effects and so it can be more difficult to reach the higher tolerance.

Any further debugging would probably require code/problem data so that we can look at what the solver is doing during the solve.