llama.cpp

2c04beeb - cuda : avoid extra QxQ matrix in shared memory