Found 1 document(s)
Page 1 of 1
-
Sparse attention reduces the computational cost of long-sequence video Transformers, but three-dimensional token layouts often leave selected blocks partially occupied, wasting physical tile capacity. We present KPage, a...