Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster

Zyphra’s ZYFA FP8 attention kernels for AMD MI355X run 1.16-1.76x faster than AITER on captured real model activations.

MarkTechPost · Michal Sutter · https://www.facebook.com/MarkTechPost/