Published August 7, 2026 | Version v2

Artifact for "Hardware Characterization of Diffusion vs. Autoregressive Language Model Inference: Compute vs. Memory Bottlenecks"

  • 1. ROR icon University of Virginia

Description

This repository contains the profiling, measurement, and plotting code behind the IISWC 2026 paper "Hardware Characterization of Diffusion vs. Autoregressive Language Model Inference: Compute vs. Memory Bottlenecks." The paper characterizes four representative inference paradigms: an autoregressive baseline, full-sequence masked diffusion, block diffusion, and
a larger block-diffusion+MoE model, plus a tensor-parallel study of a fifth model, on NVIDIA H100 GPUs using NVIDIA Nsight Compute (kernel-level) and PyTorch allocator-level (system-level) profiling.

Files

diffusion_model_artifact.zip

Files (5.0 MB)

Name Size
md5:78e1b24a52b18da79deb658a12708e3f
5.0 MB Preview Download

Additional details

Software

Programming language
Python