llama.cpp

0610f855 - glm5-next : cont q_absorbed for -sm tensor with non-unified KV