Different results when use the PESCalculator and Potential model

[SmallBearC]

I am verifying the accuracy of my model. But I found that the results obtained using PESCalculator and Potential are different: some of the code is as follows:
I get the Potential by:
model = matgl.load_model("test.out")
I get the PESCalculator by:
calc = PESCalculator(model)
When calculate the force directly：
forces_calc = calc.get_forces(atom) (or by atom.set_calcualtor(calc); forces_calc = atom.get_forces() )
When calculate the force by Potential:
g, lat, state_attr = ase_converter.get_graph(atom)
e, f, s, h = model(g=g, lat=lat, state_attr=state_attr)
forces_ml_pot = f

I found that the forces_ml_pot is the true predicted value because it is very similar to the data in the training set and the obtained error is consistent with the training data. However, forces_calc is very very far away from the training data.
Usually, we use ASE for various simulations, and we calculate energy or force by the function: atom. set_calculator(). If we get a wrong result by this function, all the simulation results would be also wrong. So, What is the reason for this?