llama.cpp

cuda : fix dmmv cols requirement to 2*GGML_CUDA_DMMV_X

#8800

Merged