Hey @elisio,
There are 2 errors in your implementation of GRULM model. The first error lies in ShiftRight. Although, we do have to specify the mode, but the mode is passed in as a function parameter, and we have to use that, so that as per the scenario in-case, we can change the argument for this parameter to use the model in different modes.
The second error lies in the way you have implemented the stack of GRU layers. Check this highly relevant thread to your discussion, and you will get to know how to fix this issue.
Cheers,
Elemento