Dynamic Bit Quantization

A model quantization technique that adaptively selects bit width per layer or according to weight importance, retaining high precision for sensitive layers while compressing redundant layers more aggressively, thereby reducing model size while preserving inference quality as much as possible.