llama.cpp

c24a2a6e - cuda : replace remaining shfl_xor with calls to warp_reduce functions (#5744)