Add custom mask-free fused backward & ring backward attention kernels (1.35x speedup on WAN 76k CP=4) - #485

Add custom mask-free fused backward & ring backward attention kernels (1.35x speedup on WAN 76k CP=4)#485

csgoogle wants to merge 1 commit into

mainAI-Hypercomputer/maxdiffusion:mainfrom

sagarchapara/custom-splash-fused-bwd-ringAI-Hypercomputer/maxdiffusion:sagarchapara/custom-splash-fused-bwd-ringCopy head branch name to clipboard