Use a single warp per element instead of a single block per element if the K-dimension is small#13

gaugarg-nv wants to merge 2 commits intoJohannesGaessler:ggml-meta-backend-8JohannesGaessler/llama.cpp:ggml-meta-backend-8from

gaugarg-nv:small_k_optgaugarg-nv/llama.cpp:small_k_optCopy head branch name to clipboard