[🚨 New paper] We depeloped MSPT - parallelized multi-scale attention method based on hierarchical partitioning of data. It is incredibly fast and achieves SOTA performance on multiple PDE tasks.