MindIE SD

English | 简体中文

Latest News

• 08/03/2026: Completed high-performance MiniMax-H3 inference adaptation and optimization on Ascend
• 06/2026: Enhanced low-precision inference, multimodal MoE, sparse attention, and the deployment ecosystem
• 05/28/2026: Cache-DiT completed its MindIE-SD integration
• 05/2026: Completed multimodal MoE inference capability development and multi-model adaptation validation
• 04/2026: Enhanced graph compilation, operator fusion, and sparse computation capabilities
• 03/2026: Extended W4A4 and MXFP4 quantization capabilities
• 02/2026: Added sparse attention operators and Dynamic EPLB
• 01/2026: Added memory optimization capabilities and serving examples
• 12/31/2025: MindIE SD provides sparse Attention computation capability
• 12/25/2025: vLLM Omni x MindIE SD achieves native high-performance Ascend inference for Qwen-Image-Edit-2511 / Qwen-Image-Layered
• 11/30/2025: MindIE SD officially open-sourced and available to the public! Meeting Calendar

Introduction

MindIE SD (Mind Inference Engine Stable Diffusion) is the visual generation inference model suite of MindIE. Its goal is to provide an end-to-end inference solution for Stable Diffusion (SD) series large models on Ascend hardware and its software stack. The software system internally integrates various functional modules and provides a unified programming interface externally.

Below are two AI Agents for the MindIE-SD code repositories. Simply click the "Ask AI" badge to enter their dedicated pages, effectively alleviating the difficulty of source code reading and starting an intelligent code learning and Q&A experience! They will help you understand the operating principles of MindIE-SD more deeply and assist in resolving issues and errors encountered during use.

Installation Guide

Quick Start:

pip install mindiesd

For other installation methods and environment dependencies, see Installation Guide.

Architecture Overview

For MindIE SD architecture and key features, see Architecture Overview.

Quick Start

Quick Start: Using Wan2.1 as an example, introduces the overall acceleration effect with MindIE SD.

MiniMax-H3 Inference and Performance Optimization: Introduces MiniMax-H3 inference deployment, adaptation, and performance optimization on Ascend.

Developer Documentation

Contribution Guide: Explains how to submit Issues, Pull Requests, and coding standards. AI Agent Support: Introduces the default skills used in the repository and the agent-assisted development workflow. Test Verification: Introduces the unit test example execution workflow.

Ecosystem Support

The following projects have integrated MindIE SD as a multimodal inference backend:

• vLLM Omni:
• Cache-DiT:

Paper Citations

@misc{RainFusion2.0@2025,
    title = {RainFusion2.0: Temporal-Spatial Awareness and Hardware-Efficient Block-wise Sparse Attention},
    url = {https://gitcode.com/Ascend/MindIE-SD.git},
    note = {Open-source software available at https://gitcode.com/Ascend/MindIE-SD.git},
    author = {Aiyue Chen and others},
    year = {2025}
    }