The Super Weight in Large Language Models

2024-11-11

arXiv.org

AI Summary

LLMs contain a sparse set of “super weights” whose perturbation—by altering a single parameter—triggers catastrophic performance collapse (e.g., sharp perplexity increase or near-zero zero-shot accuracy). This work introduces the first data-free, single-forward-pass method to identify super weights, grounded in forward sensitivity analysis and joint detection of weight/activation anomalies; it formally defines both super weights and their associated “super activations.” We further propose a non-uniform quantization strategy: preserving critical parameters at high precision while aggressively pruning and round-to-nearest quantizing the rest. This scheme is extended to block-wise quantization for larger weight groups. Experiments demonstrate that our approach matches or exceeds state-of-the-art quantization methods in accuracy, significantly improves robustness and inference efficiency, and achieves up to 4× faster quantization. We publicly release super-weight coordinate indices for major open-source LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Recent works have shown a surprising result: a small fraction of Large Language Model (LLM) parameter outliers are disproportionately important to the quality of the model. LLMs contain billions of parameters, so these small fractions, such as 0.01%, translate to hundreds of thousands of parameters. In this work, we present an even more surprising finding: Pruning as few as a single parameter can destroy an LLM's ability to generate text -- increasing perplexity by 3 orders of magnitude and reducing zero-shot accuracy to guessing. We propose a data-free method for identifying such parameters, termed super weights, using a single forward pass through the model. We additionally find that these super weights induce correspondingly rare and large activation outliers, termed super activations. When preserved with high precision, super activations can improve simple round-to-nearest quantization to become competitive with state-of-the-art methods. For weight quantization, we similarly find that by preserving the super weight and clipping other weight outliers, round-to-nearest quantization can scale to much larger block sizes than previously considered. To facilitate further research into super weights, we provide an index of super weight coordinates for common, openly available LLMs.

Problem

• Identify critical parameters (super weights) in LLMs
• Preserve super activations to improve quantization methods
• Enable efficient pruning without destroying model performance

Innovation

• Data-free method identifies super weights
• Preserve super activations for better quantization
• Clip weight outliers to enable larger blocks