llama.cpp

86221cf6 - CUDA: fix FA kernel selection logic (#21271)