Quantization Meets dLLMs: A Systematic Study of Post-training Quantization for Diffusion LLMs

TL;DR

This paper systematically investigates post-training quantization techniques for diffusion large language models (dLLMs), addressing challenges posed by activation outliers and evaluating various configurations to facilitate efficient deployment on edge devices.

Contribution

It is the first comprehensive study on quantizing diffusion-based language models, analyzing the impact of activation outliers and evaluating multiple PTQ methods across different tasks and model variants.

Findings

Activation outliers significantly affect low-bit quantization accuracy.

State-of-the-art PTQ methods can be adapted for dLLMs with careful handling of outliers.

Quantization performance varies across task types and model configurations.

Abstract

Recent advances in diffusion large language models (dLLMs) have introduced a promising alternative to autoregressive (AR) LLMs for natural language generation tasks, leveraging full attention and denoising-based decoding strategies. However, the deployment of these models on edge devices remains challenging due to their massive parameter scale and high resource demands. While post-training quantization (PTQ) has emerged as a widely adopted technique for compressing AR LLMs, its applicability to dLLMs remains largely unexplored. In this work, we present the first systematic study on quantizing diffusion-based language models. We begin by identifying the presence of activation outliers, characterized by abnormally large activation values that dominate the dynamic range. These outliers pose a key challenge to low-bit quantization, as they make it difficult to preserve precision for the majority of values. More importantly, we implement state-of-the-art PTQ methods and conduct a comprehensive evaluation across multiple task types and model variants. Our analysis is structured along four key dimensions: bit-width, quantization method, task category, and model type. Through this multi-perspective evaluation, we offer practical insights into the quantization behavior of dLLMs under different configurations. We hope our findings provide a foundation for future research in efficient dLLM deployment. Our code is publicly available at https://github.com/FelixMessi/QDLM.