AMQ: Enabling AutoML for Mixed-precision Weight-Only Quantization of Large Language Models

2025-09-15

AI Summary

Efficient deployment of large language models (LLMs) under stringent memory constraints demands automated search for high-performance mixed-precision weight quantization configurations—a key challenge due to the astronomical configuration space (~10¹⁰⁰). This paper proposes a Pareto-optimal automated quantization search framework. First, it prunes the search space using domain-informed heuristics. Second, it introduces a lightweight layer-wise quality predictor and a proxy evaluation mechanism that avoids full-format conversion, drastically reducing overhead. Third, it performs iterative optimization for fine-grained, layer-adaptive bit-width allocation. The method significantly improves search efficiency and stability, achieving state-of-the-art accuracy–memory trade-offs across multiple LLMs. Compared to baseline approaches, it accelerates search by over an order of magnitude while reducing accuracy degradation by more than 50%.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• 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

Abstract

To enable broader deployment of Large Language Models (LLMs), it is essential to identify the best-performing model under strict memory constraints. We present AMQ, Automated Mixed-Precision Weight-Only Quantization, a framework that assigns layer-wise quantization bit-widths to optimally balance model quality and memory usage. However, the combinatorial search space, with over 10^{100} possible configurations, makes conventional black-box optimization infeasible. AMQ overcomes this challenge through four key innovations:(1) search space pruning using prior knowledge to exclude unpromising configurations, (2) quantization proxy to bypass costly format conversions during search, (3) quality predictor to minimize evaluation overhead, and (4) iterative search-and-update strategy for fast and stable convergence. By integrating these components, AMQ efficiently explores the quality-efficiency landscape, reaching the Pareto frontier and yielding LLMs that are both compact and high-performing. Our code is available at https://github.com/dlwns147/amq.

Problem

• Automating mixed-precision quantization for large language models
• Reducing memory usage while maintaining model quality
• Solving combinatorial search space exceeding 10^100 configurations

Innovation

• Search space pruning with prior knowledge
• Quantization proxy to avoid conversions
• Quality predictor minimizing evaluation overhead