Rare find

ffpa-attn. ๐Ÿค–FFPA: Extends FlashAttention-2 via Split-D for large headdims, 1.5x~3ร—โ†‘๐ŸŽ‰ vs SDPA, up to 430T๐ŸŽ‰ on H200.

github.com/xlite-dev/ffpa-attn

Problem

Audience

Updates

No founder updates yet. Check back soon.