Anyone implemented bi-lstm wrapper that takes customized cell and able to handle packed_sequence

I’m trying to reproduce the result of a paper which uses layernorm lstm cells. The author uses tensorflow to implement it so it’s fairly simple. I searched for bi-lstm with layer norm online but I only found those works that does not support different lengths. I implemented a bi-lstm that handles different lengths before but it’s slow and it doesn’t take packed sequence as the official one does.
I’m not familiar with Jitter also.
I wonder if a good wrapper exists.