llama.cpp

63e66fdd - opencl: use flat variants of q4_K and q6_K gemv for very large M (#24006)