IQ Quantization
A non-uniform quantization scheme introduced in llama.cpp. It dynamically adjusts quantization precision based on the importance of each weight to model output, preserving higher precision for critical weights and applying greater compression to weights with lesser impact.
Timeline (last 90 days)
IQ前缀（如UD-IQ1_S）表示基于重要性矩阵的量化，理论上在极低位宽下表现更优，但实际效果高度依赖校准数据质量。
IQ quantization is an importance-weighted non-uniform quantization scheme in the llama.cpp ecosystem; IQ2 averages approximately 2 bits per weight and is one of the highest-compression options among mainstream quantization approaches.
The IQ series (e.g., IQ1-XS, IQ3) is a variant of importance-aware quantization that allocates different precision levels to each weight based on its importance to the output, retaining more useful information than uniform quantization at the same compression ratio.
All Facts (3)
IQ前缀（如UD-IQ1_S）表示基于重要性矩阵的量化，理论上在极低位宽下表现更优，但实际效果高度依赖校准数据质量。
50%UnverifiedIQ quantization is an importance-weighted non-uniform quantization scheme in the llama.cpp ecosystem; IQ2 averages approximately 2 bits per weight and is one of the highest-compression options among mainstream quantization approaches.
50%UnverifiedThe IQ series (e.g., IQ1-XS, IQ3) is a variant of importance-aware quantization that allocates different precision levels to each weight based on its importance to the output, retaining more useful information than uniform quantization at the same compression ratio.
50%