FlashAttention-3

Tri Dao's Hopper-native attention kernel: 1.5–2× faster than FlashAttention-2 with FP8 support, now universally adopted across training and inference stacks. Developed with Colfax, NVIDIA, Together AI, and Meta collaborators — flag the industrial co-branding — but the FlashAttention line is Dao's Princeton-anchored efficient-kernels agenda, continued in GTA/GLA attention and the SonicMoE training kernels.