Proper way to implement Spark for real-time inference with Keras model?

Question

Here is a piece of my code:

@pandas_udf(StringType())
def online_predict(values: pd.Series) -> pd.Series:
    pred = Model.from_config(bc_config.value)