Signed Symmetric Quantization Reduces Error in Few-Bit Integers

A new quantization method addresses clipping errors in signed integer alphabets by utilizing signed symmetric quantization. In llama.cpp on AMD EPYC hardware, this approach maintains symmetric runtime profiles while avoiding the performance penalties of asymmetric quantization.