llama.cpp

ggml : fix quants nans when all the group weights are very close to zero

#7313

Merged