CUDA: faster tile FA (Pascal/AMD), headsize 256#15769

JohannesGaessler merged 1 commit intoggml-org:masterggml-org/llama.cpp:masterfrom

JohannesGaessler:cuda-fa-tile-256-7JohannesGaessler/llama.cpp:cuda-fa-tile-256-7Copy head branch name to clipboard