From 2:4 to 8:16 sparsity patterns in LLMs for Outliers and Weights with Variance Correction
This is an incremental improvement in compression techniques for LLMs, addressing efficiency issues for model deployment.
The paper tackles the problem of limited flexibility and sensitivity to outlier weights in structured sparsity methods for compressing large language models (LLMs), showing that 8:16 semi-structured sparsity surpasses a performance threshold with minimal storage overhead (0.875 vs. 0.75 bits/element) and that structured sparsity for outliers is competitive with unstructured approaches.
As large language models (LLMs) grow in size, efficient compression techniques like quantization and sparsification are critical. While quantization maintains performance with reduced precision, structured sparsity methods, such as N:M sparsification, often fall short due to limited flexibility, and sensitivity to outlier weights. We explore 8:16 semi-structured sparsity, demonstrating its ability to surpass the Performance Threshold-where a compressed model matches the accuracy of its uncompressed or smaller counterpart under equivalent memory constraints. Compared to 2:4 sparsity, 8:16 offers greater flexibility with minimal storage overhead (0.875 vs. 0.75 bits/element). We also apply sparse structured patterns for salient weights, showing that structured sparsity for outliers is competitive with unstructured approaches leading to equivalent or better results. Finally, we demonstrate that simple techniques such as variance correction and SmoothQuant like weight equalization improve sparse models performance.