Artifact for "Hardware Characterization of Diffusion vs. Autoregressive Language Model Inference: Compute vs. Memory Bottlenecks"
Authors/Creators
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
Additional details
Software
- Programming language
- Python