Small and Fast LLMs on Commodity Hardware: Post-Training Quantization in llama. cpp