LPCD: Unified Framework from Layer-Wise to Submodule Quantization

TL;DR

LPCD is a unified quantization framework that optimizes arbitrary submodules in large language models, improving upon layer-wise methods by capturing larger submodule behaviors while maintaining efficiency.

Contribution

Introduces LPCD, a general framework for submodule quantization that extends PTQ beyond layers, unifying and improving existing methods.

Findings

LPCD improves quantization accuracy across diverse LLMs.

LPCD enhances existing layer-wise PTQ methods.

LPCD maintains efficiency comparable to traditional PTQ.

Abstract

Post-training quantization (PTQ) aims to preserve model-level behavior; however, most methods focus on individual linear layers. Even recent extensions, such as QEP and LoaQ, which mitigate error propagation or target specific submodules, still rely on layer-wise formulations and fail to capture the behavior of larger submodules. We introduce Layer-Projected Coordinate Descent (LPCD), a unified framework that extends PTQ beyond layers by optimizing relaxed objectives across arbitrary submodules and projecting the solutions with layer-wise quantizers. LPCD generalizes existing methods and provides a principled approach to quantizing complex submodules while maintaining the efficiency and compatibility of layer-wise PTQ pipelines. Across diverse LLM architectures and bit-widths, LPCD-based submodule quantization consistently enhances both layer-wise PTQ methods and existing submodule approaches.