Experimental ultra-low precision quantization formats for AMD hardware, enabling faster LLM inference with 2-8 bit weights
