Why I get different ouputs with different construction way of LSTM using tensorflow?

Question

I am exploring the LSTM of tensorflow2.0. I found the function of LSTM can be implemented by LSTMcell besides the normal API of LSTM layer. I tried three kinds of implementation