MindStudio
This Huawei-designed full-pipeline development toolchain equips Ascend AI developers with an end-to-end solution for building Ascend AI applications, accelerating the processes of training, inference, and operator development.
- Release Notes
CANN version matching MindStudio, and feature changes and issue fixed of the current version.
- Quick Start
Quick start for using the tool in the training, inference, and operator development scenarios.
Visualization Tools
- MindStudio Insight
This is a visualized performance tuning tool, which assists in locating model and operator performance issues.
- Layered Model Visualization Tool (TensorBoard)
This is a visualized accuracy debugging tool, which displays the collected MindSpore and PyTorch training accuracy data and model structure.
Operator Development Tools
- Operator Design (msKPP)
msKPP provides functions such as performance modeling and analysis, calling of msOpGen operator projects, and automatic tuning based on the Ascend C template library.
- Operator Project Creation (msOpGen)
msOpGen provides functions such as template project creation, simplified operator project setup, and operator test and verification.
- Operator Test (msOpST)
msOpST provides the function of testing the input and output of an operator in a real hardware environment to verify the operator function.
- Anomaly Detection (msSanitizer)
msSanitizer provides the memory, contention, and uninitialization detection functions, and accurately locates memory issues in multi-core programs.
- Operator Debugging (msDebug)
Based on the native environment debugging capability of the Ascend AI Processor, msDebug implements flexible variable display and supports operator function debugging and single-step debugging (whether on board or through simulations).
- Operator Tuning (msProf)
msProf collects on-board and simulation profile data, and displays the data in a visual format through MindStudio Insight, allowing users to quickly locate operator performance bottlenecks.
Training Development Tools
- Performance Tuning Tools
These tools collect and analyze profile data in MindSpore, PyTorch, and TensorFlow training scenarios.
- Memory Leak Detection Tool (msLeaks)
This tool detects memory issues during model training.
- Performance Analysis Tool (msprof_analyze)
This tool compares and analyzes the collected profile data in PyTorch and MindSpore training scenarios, and outputs the analysis result.
Inference Development Tools
Environment Deployment
- Prechecker Tool for Foundation Model Environment Deployment
This tool provides precheck before inference, data flushing during inference, and comparison after inference.
- Large Language Model Debug Tool
This tool provides automatic bad case analysis, dump, and automatic comparison of LLM inference data.
- Traditional Model Inference Accuracy Tool (msit debug)
This is a one-stop debugging and tuning tool for traditional model inference. It provides data dump, automatic comparison, accuracy precheck, and ONNX optimization and graph modification functions.
- Accuracy Debugging Tool
This tool compares accuracy data in TensorFlow, ONNX, and Caffe inference scenarios.
- Model Tuning Tool (msprof)
This tool collects and parses profile data in command line mode.
- Serving Tuning Tool (msServiceProfiler)
This tool supports serving tuning of the MindIE Service and vLLM frameworks. It provides the following functions: serving profile data comparison, vLLM serving profile data collection, automatic serving parameter optimization, and serving expert suggestions.
- Memory Leak Detection Tool (msLeaks)
This tool detects memory issues during model inference.
- Expert Load Balancing Tools
The tools are used to perform affinity-based expert determination for load balancing in both static and dynamic scenarios within DeepSeek models.
Cases, APIs, and References
Cases
- General Performance Troubleshooting Guide
- Case Study on Locating Foundation Model Training Performance Bottlenecks
- Case Study on Locating Foundation Model Training Accuracy Issues
- Analysis Cases of Foundation Model Inference Accuracy Issues
- Debugging and Tuning Guide for Foundation Model Inference Quantization
- Migration, Debugging, and Tuning Guide for Traditional Model Inference
- Memory Issue Analysis Cases