llama.cpp

e5aa067d - llama : rotate activations for better quantization