tensorflow: The LSTM does not generate reproducible results, but GRU does

I can not get reproducible results by just using the LSTM, but GRU is Okay.

Let me explain what happens.

if I train an identical LSTM, 10 times in a loop, for the first 6 run it handles MSE equal to value a and for the rest 4 run it handles MSE value b. weird isn’t? Just value a OR b. it means an identical LSTM can handle two MSEs by random and just two MSE values.

and if I just replace the LSTM with GRU, all 10 runs handle identical MSEs.

I use Tensorflow 1.7 and Keras 2.1.5, I run on the CPU.