Skip to main content

TopK Attention

Experiments using topk operator within self attention block
Created on March 25|Last edited on March 25

Normal (red) vs. TopK HARD (orange) vs. TopK Soft (TBD)


2k4k6k8kStep405060708090
group: KXtakAjjB3Q58vacVKpX8A
group: SJvwbBjUVfknhhoiCQUmmu
Run set
12



Run set
12