LWC

Learnable Weight Clipping (LWC) is a differentiable weight-range optimization technique for neural network quantization, especially for PTQ of LLMs. It introduces channel-scaling structural transformations on top of learnable clipping to boost the accuracy of low-bit weight-activation (W/A) quantization. For details about the algorithm, see OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models.