Why n_jobs=-1 in RandomizedSearchCV throws an error?

Question

I am trying to do cross validation on some dataset with Keras and Sklearn. The issue is I can not use multiprocessing for this task since when I pass any value but 1 as a parameter for n_jobs in RandomizedSearchCV it throws an error:

PicklingError: Could not pickle the task to send it to the workers.

Fun fact: everything worked just fine yesterday. My code below:

import keras
from keras import metrics
from keras import layers
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.model_selection import RandomizedSearchCV

def create_model(units=16, 
                 dropout=0.3, 
                 activation='relu', 
                 optimizer='adam', 
                 regularization=None):
    
    model = keras.Sequential(
        [
            layers.Input(shape=(x_df.shape[1], )),
            
            layers.Dense(units=units, 
                         activation=activation, 
                         kernel_regularizer=regularization),
            
            layers.Dropout(dropout),
            
            layers.Dense(units=units, 
                         activation=activation, 
                         kernel_regularizer=regularization),
            
            layers.Dropout(dropout),
            
            layers.Dense(len(names), activation='softmax')
            
        ], name='Sequential_model_vol_1'
    )

    model.compile(loss='sparse_categorical_crossentropy', 
                  optimizer=optimizer, 
                  metrics=[
                      metrics.SparseCategoricalAccuracy(name='accuracy'), 
                      metrics.SparseTopKCategoricalAccuracy(k=3, name='top-3-accuracy')
                  ]
    )
    return model

# dummy param_grid
param_grid = {
    'units': [32, 64, 128],
    'dropout': [0.2, 0.3, 0.5],
    'activation': ['relu', 'sigmoid'],
}

# instantiate clf
cv_clf = RandomizedSearchCV(
    estimator=KerasClassifier(create_model),
    param_distributions=param_grid, 
    scoring='accuracy',
    verbose=1,
    n_jobs=-1,
    cv=10
)

# fitting
cv_clf.fit(x_train, y_train, verbose=0)

As I said, when n_jobs=1 everything works fine but it takes too long to compute.