Any hint how to use this model with smaller context, like 16K or 32K. 128K is way too much for my setup. But after setting 16K it seams model breaks after set context lenght, so I must miss something. Do I need to set alpha_value or compress_pos_emb?
