Author of metal-flash-attention
GPU performance engineer · current
Philip ported FlashAttention to Metal and beat Apple's own MPSGraph by up to an order of magnitude, then documented Apple GPU behavior in benchmarks the whole community cites. Draw Things and other on-device apps are faster today because of his kernels.
We built this from your public work because we think it deserves celebrating. You did not ask us to, so the only fair thing is that you decide what happens to it. Claim it and it is yours to edit. Ask us to change something and we will. Ask us to take it down and it is gone within 72 hours — free, no account needed, and nobody will try to talk you out of it.