"mingru/minlstm" is not a new idea (https://arxiv.org/abs/2410.01201v1), as it is close to the QUASI-RECURRENT NEURAL NETWORKS (https://openreview.net/pdf?id=H1zJ-v5xl).
What is new and very powerful is parallel instead of sequential and the log stable version!
Inspired by lucidrains pytorch version https://github.com/lucidrains/minGRU-pytorch & Bá Ngọc : https://github.com/bangoc123 (https://colab.research.google.com/drive/1hI-tk8PnZ2cvbWHbdQzWx8R8Sma423Lq?usp=sharing)