How to make a process more efficient with Python multiprocessing

Question

I have a function func_app which the argument is a huge Dataframe (df1) with 17.000.000 rows. In order to make the process more efficient I tried to use the multiprocessing library but it still takes a lot of time and memory because I need to do a groupBy before executing the function. This GROUP BY is the reason why I cannot split the Dataframe in smaller portions before executing the function which would make things a lot easier and faster but i don't see the way of getting rid of it. Is there a way I am not thinking of to fix this?

Here below is a portion of the code in python:

def func_app(g):
    g = g[1].reset_index(drop=True)
    niv_app = g.reindex(columns=['A1','B1','B2','B3'])
    niv_app = niv_app.sort_values(by = ['A1','B1'],ascending = [True,False]).groupby('A1').head(1)
    niv_app = niv_app.reset_index(drop=True)
    g['B1_BIS'] = niv_app.loc[0,'B1']
    g['B2_BIS'] = niv_app.loc[0,'B2']
    g['B3_BIS'] = niv_app.loc[0,'B3']
    return g

with parallel_backend('loky', n_jobs=dict_download['num_cores']):
    var = Parallel(verbose=5)(delayed(func_app)(g) for g in df1.groupby(['A1','A2']))
print('\nManaging results...')
df_output = pd.concat(var, sort=False)

THANK YOU!