I see. It could be not so easy to tell what happens. i would suggest you ask the question at iqtree3 github and there would be experts helping you. If you are working on aa (mixture) model, you can also try
phylohessian and if it doesn't work pls don't hesitate to let me know. Alt, an easier way is to use the bootstrap approach
bs_inBV which in general gives you a good approx to the hessian that is needed for mcmctree's approx lik dating method.