How to convert Keras LSTM model into tf.nn.rnn_cell.LSTMCell model

I have trained a LSTM model on pytorch, and converted it into Keras lstm model with set_weights method. The performance looks almost the same. However, I need to convert the model into lstm model with older tensorflow version (V1.xx), by using tf.nn.rnn_cell.LSTMCell method. What confused me was that I did not find any method that can load lstm weights from numpy like arrays. The only method I find is loading saved checkpoints, but not suitable for me. Any suggestion for me? Thanks.