llama.cpp

69664749 - cuda : play with faster Q4_0 dequantization