Documentation
中文
Superb Performance
Improves model inference throughput several fold, significantly reducing latency and inference costs.
Simplified Deployment
Supports simplified inference deployment on a single node or multiple nodes, as well as large-scale expert parallelism.
Light Load
Uses several optimization technologies, such as efficient memory management and KV cache compression, to reduce load.
Openness and Compatibility
Supports MindIE Service serving and multiple interfaces, including vLLM, OpenAl, and Triton.

MindIE Facilitates Easy Deployment and High-Performance Inference of Ascend

Vision Generation
MindIE SD enables high-performance vision generation inference and visualization.
Text Generation
MindIE LLM enables text generation inference and accelerates the E2E inference process.
Third-Party Engine Inference
Ascend supports open source and customer-developed inference engines, such as vLLM, to accelerate inference.
Application Enablement (RAG, Agent, etc.)
MindIE
Inference Service
Inference Model Suite
Acceleration Suite
MindIE Motor for Ascend Inference
Third-party
MindIE SD for Multimodal Generation
MindIE LLM for Text Generation
vLLM/SGLang for Text Generation
MindIE Turbo Optional
AI Framework
MindSpore
PyTorch
Hardware Enablement
CANN
MindIE SD for Multimodal Generation
MindIE SD accelerates vision generation with a suite of software optimization solutions.

Key Technologies

Integrated Serving Deployment
Easy-to-use serving capabilities and proprietary and open source solutions to facilitate service integration and deployment
  • Open MindIE Serving DeploymentMindIE serving supports multiple third-party APIs, such as Triton, OpenAl, TGI, and vLLM, for session management and request scheduling.
  • Open Source vLLM EcosystemvLLM is fully compatible with Ascend and supports its core features through open source vLLM-Ascend projects.

Use Cases

China Southern Airlines
China Southern Airlines uses Ascend MindIE to deploy DeepSeek, exploring innovative applications of AI computing power in flight operations, logistics, passenger services, and business management.
Longgang, Shenzhen
MindIE is used to access and deploy DeepSeek to enhance government and citizen platforms, facilitating the handling of government services and citizen inquiries.
Southeast University
MindIE enables local deployment of DeepSeek to integrate with the university's information service portal and app, serving teachers and students with capacity for 800 concurrent QAs and 100 million+ tokens daily.