llama.cpp

ggml-quants : avoid invalid rounding in qkx3 scale search

#29817

Merged