HKUSTDial/flash-sparse-attention
Trainable fast and memory-efficient sparse attention
A star delta has been computed for 62,678 of 63,382 tracked repositories in this period, and only those can be ranked in it.
A repository missing here has not been measured far enough back yet. That does not mean it did not move.
Star counts measured between 09:42 and 17:21 (BKK).
Trainable fast and memory-efficient sparse attention
Veda Sparse Attention for ComfyUI
This repo is a alternative solution for standard self-attention. It uses tree routing with learnable functions for cheaper self-attention while not compromising accuracy.
That's all 3 repositories in this period.