Umar Khan

Hello, I can possibly join in on the effort to rewrite the functionality for a more permissive license. the models however will remain with their original license but at least...

@WongKinYiu would you be interested in organising this? @Youho99 An alternative I am thinking is I can create a small discord, we fork this code and go through it, separate...

I have created the Discord Server, however I haven't had the time to set it up properly, I have started looking into some core code and will start doing it...

I can pitch in on whatever you need help with, how do you plan on managing the tasks?

there are several implementation errors between this repository and the research paper. but the largest one that I can spot is: Normalizer state: The implementation is missing the normalizer state...

by introducing the normalization layer alongside a few changes in line with the paper I noticed that I get no NaN's up until 512 sequence length (that's the only length...

we should be able to construct a combination architecture using both sLSTM and mLSTM blocks, infact that is exactly how the original research paper suggest we use xLSTM. additionally I...

I mean you can create separate blocks of either of the mLSTM and sLSTM and combine them into a single architecture file, and then pass the around the data as...

> > we should be able to construct a combination architecture using both sLSTM and mLSTM blocks, infact that is exactly how the original research paper suggest we use xLSTM....

> Hi, the mlstm code change is still always nan did you try the fix in my repo? if its still there I can look this over on the weekend...