Weight-only Quantization

Compresses stored weights only; requires dequantization during compute, saving memory but not necessarily FLOPs