Any one knows how this holds up on long context retrieval (needle in haystack , ruler) vs same size full attention model? efficiency gains look great but that usually where linear attention hybrids fall apart.