llama.cpp

metal : enable mat-vec kernels for bs <= 4

#10491

Merged