OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

DistriFusion

DistriFusion is an open-source Python implementation of distributed parallel inference for high-resolution diffusion models. It partitions image computation across multiple GPUs and uses communication between patches to accelerate image generation. The repository includes example pipelines for SDXL and Stable Diffusion 1.4/2 and an interface modeled on Diffusers.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

DistriFusion — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use DistriFusion when you want to experiment with or implement multi-GPU inference for high-resolution diffusion generation, especially with the repository's SDXL and Stable Diffusion examples.

Who it is for

  • Developers and researchers experimenting with multi-GPU diffusion inference.
  • Teams generating high-resolution images with supported Stable Diffusion model examples.
  • Researchers benchmarking parallel diffusion inference.

Documented uses

  • Run SDXL image generation across multiple GPUs using the provided example pipeline.
  • Explore distributed inference for Stable Diffusion 1.4/2.
  • Compare reported inference speedups as GPU count and output resolution increase.
Before you choose

Requirements & limits

The code is publicly available under the MIT license. The repository documents installation through PyPI, GitHub, or a local development checkout.

  • Requires Python 3, an NVIDIA GPU, CUDA 12.0 or later with corresponding cuDNN, and PyTorch 2.2.
  • The README shows SDXL and Stable Diffusion 1.4/2 examples; it does not establish support for every diffusion model.
  • Reported speedups use SDXL on A100 GPUs and should not be treated as guaranteed results on other setups.
  • The method uses synchronous communication at the first step and asynchronous activation reuse afterward to coordinate patch computation.
  • The README reports integration in TensorRT-LLM and support in ColossalAI; these are separate frameworks.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run the documented SDXL example across GPUs

Generate an SDXL image using DistriFusion's provided example script.

Before you start

  • Python 3
  • NVIDIA GPU with CUDA 12.0 or later and corresponding cuDNN
  • PyTorch 2.2
  • A pretrained SDXL model; the example names stabilityai/stable-diffusion-xl-base-1.0

Steps

  1. Install PyTorch 2.2 and the listed CUDA/cuDNN prerequisites.

  2. Install DistriFusion with pip, from PyPI or the GitHub repository.

  3. Configure DistriConfig and load the pretrained model with DistriSDXLPipeline, following scripts/sdxl_example.py.

  4. Run the example with torchrun and set --nproc_per_node to the number of GPUs to use.

Expected outcome

The example generates an image from a text prompt and saves it on rank 0.

Guide limitations

  • The guide follows the repository's example; it does not establish performance on hardware other than the benchmarked setup.
  • The listed prerequisites specify NVIDIA GPU and CUDA; no broader hardware compatibility is claimed.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
17 October 2026
Original website title
mit-han-lab/distrifuser · Official homepage title · retained in its source language
How we review sourcesSuggest a correction