llama.cpp

cuda : replace remaining shfl_xor with calls to warp_reduce functions

#5744

Merged