EADST

Optimize Weight Rounding via Signed Gradient Descent for the Quantization of Large Language Models

• Paper: Optimize Weight Rounding on arXiv
• Code: Intel Neural Compressor on GitHub
• Organization: Intel

Key Feature:

• Adaptive Weight Rounding: Utilizes backward optimization to dynamically adjust the quantized integer values, either rounding them up or down, to optimize the model's performance during quantization.