llama.cpp

CUDA: implement __hmax and __hmax2 for CUDA < 11.7

#7019

Merged