I forgot to close this thread, I obtain the answer with another thread. The issue was resolved using the reduce_sum_static function instead of the reduce_sum function, which enables to put a seed in the within-chain parallelization:
Thank you to all !