Scikit-learn: StackingRegressor with GridSearchCV

I want a stacked regression where the final_estimator is regfinal and the estimators are reg1 and reg2. For example:

reg1 = RandomForestRegressor()
reg2 = KNeighborsRegressor()
regfinal = ridge_regression()

I am unsure about the parametrization of the estimators so I want to perform a grid search first. I see two ways.

The first way gives model_A, which is StackingRegressor with two GridSearchCV (on reg1 and reg2) as estimators:

params_reg1 = {some grid}
params_reg2 = {some grid}

grid1 = GridSearchCV(estimator=reg1, param_grid=params_reg1)
grid2 = GridSearchCV(estimator=reg2, param_grid=params_reg2)

estimators = [
    ('rfr', grid1),
    ('knr', grid2)
]

model_A = StackingRegressor(
    estimators=estimators,
    final_estimator=regfinal)

The second way gives model_B which is a GridSearchCV on a StackingRegressor model with reg1 and reg2 as estimators:

estimators = [
    ('rfr', reg1),
    ('knr', reg2)
]

params_reg_all = {some grid for reg1 and reg2}

stacked = StackingRegressor(
    estimators=estimators,
    final_estimator=regfinal)

model_B = GridSearchCV(estimator=stacked, param_grid=params_reg_all)

What are the differences between model_A and model_B?