llama.cpp

cuda: optimize SOLVE_TRI using registers and FMAF

#17703

Merged