llama.cpp

479618c9 - don't use N as workgroup size