llama.cpp

Use Q4_K for attn_v for Q2_K_S when n_gqa >= 4

#4996

Merged