Understanding the Difficulty of Low-Precision Post-Training Quantization for LLMs

2024-10-18

AI Summary

This work investigates the fundamental cause of severe performance degradation in post-training quantization (PTQ) of large language models (LLMs) at ultra-low precisions (≤4 bits). We identify a structural objective mismatch: PTQ minimizes layer-wise local quantization error, which diverges from global task-loss optimization—especially under limited calibration data—leading to significantly inferior results compared to quantization-aware fine-tuning (QAT). To systematically analyze this gap, we propose a unified evaluation framework integrating inter-layer error analysis, gradient sensitivity diagnosis, and multi-granularity quantization experiments, applied across LLaMA-2 and Phi-3. Empirical results show that PTQ suffers >15 BLEU/ACC point drops at 2–4 bits, whereas QAT preserves over 95% of original performance. This study is the first to formally attribute PTQ’s intrinsic limitation at ultra-low bitwidths to objective inconsistency, thereby establishing QAT’s irreplaceable role in efficient LLM compression.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large language models of high parameter counts are computationally expensive, yet can be made much more efficient by compressing their weights to very low numerical precision. This can be achieved either through post-training quantization by minimizing local, layer-wise quantization errors, or through quantization-aware fine-tuning by minimizing the global loss function. In this study, we discovered that, under the same data constraint, the former approach nearly always fared worse than the latter, a phenomenon particularly prominent when the numerical precision is very low. We further showed that this difficulty of post-training quantization arose from stark misalignment between optimization of the local and global objective functions. Our findings explains limited utility in minimization of local quantization error and the importance of direct quantization-aware fine-tuning, in the regime of large models at very low precision.

Problem

• Analyzing challenges in low-precision post-training quantization for LLMs
• Comparing local vs global optimization in weight compression
• Explaining poor performance of layer-wise quantization at very low precision

Innovation

• Post-training quantization minimizes layer-wise errors
• Quantization-aware fine-tuning optimizes global loss
• Local and global objective functions misaligned