Given that your model features a few roots very close to a unit root, computing GIRFs based on simulations usually performs poorly. That is the reason why people often use GIRFs at stochastic steady state/ergodic mean in the absence of shocks. See Third order perturbation