SFMP: Fine-Grained, Hardware-Friendly and Search-Free Mixed-Precision Quantization for Large Language Models

TL;DR

SFMP addresses memory-constrained quantization of large language models by replacing costly discrete optimization with a fractional bit-width and block-wise, salience-guided scheme. It adds row-column weight reordering to align salient weights with hardware-friendly blocks and a unified one-bit LUT-based GEMM kernel to run mixed-precision without explicit dequantization. The method comprises a parameter-free pipeline that computes Fisher-based salience, assigns two nearby bit-widths using a fractional budget, and executes with a block-major layout. Empirically, SFMP outperforms state-of-the-art layer-wise mixed-precision methods under the same memory budgets and delivers improved throughput, enabling practical edge deployment of LLMs.

Abstract

Mixed-precision quantization is a promising approach for compressing large language models under tight memory budgets. However, existing mixed-precision methods typically suffer from one of two limitations: they either rely on expensive discrete optimization to determine precision allocation, or introduce hardware inefficiencies due to irregular memory layouts. We propose SFMP, a search-free and hardware-friendly mixed-precision quantization framework for large language models. The framework is built upon four novel ideas: Fractional bit-width, which extends integer bit-width for weight matrix to fractional value and transforms discrete precision allocation as a continuous problem; 2)Block-wise mixed-precision, enabling fine-grained precision within weight matrices while remaining hardware-friendly; 3)Row-column weight reordering, which aggregates salient weights via row and column reordering, incurring only a small activation reordering overhead during inference; 4)Unified GEMM kernel, which supports mixed-precision GEMM at arbitrary average bit-width. Extensive experiments demonstrate that SFMP outperforms state-of-the-art layer-wise mixed-precision methods under the same memory constraints, while significantly reducing quantization cost and improving inference efficiency. Code is available at https://github.com/Nkniexin/SFMP

Figures (17)

• Figure 1: Trade-off between memory usage and average zero-shot accuracy on ARC-Easy, ARC-Challenge, PIQA, HellaSwag, WinoGrande, and BoolQ.
• Figure 2: Comparison of two GEMM computation paradigms: Left) dequant-based GEMM; Right) one-bit LUT-based GEMM.
• Figure 3: Motivation of SFMP.
• Figure 4: Pipeline of SFMP.
• Figure 5: An example of unified mixed-precision GEMM.