MindStudio Training Tools
✨ What's New
🔹 [2026.03.28]: The Accuracy Debugging module (debug directory) has been officially sunset. For details, see the Announcement
🔹 [2026.02.25]: The Tinker parallel strategy automatic tuning system is officially open-sourced. For details, see the Tinker Project.
🔹 [2026.01.12]: The license for this repository has been updated. For details, see the Announcement.
🔹 [2025.12.31]: The MindStudio training development toolchain is fully open source.
ℹ️ Introduction
MindStudio Training Tools (msTT) is a training development toolchain focused on solving key challenges in model training. It provides three core tools for analysis & migration, accuracy debugging, and performance tuning, to efficiently address issues such as migration failures, loss anomalies, and performance bottlenecks, delivering a minimalist development experience with optimized accuracy and performance.
⚙️ Features
msTT provides the following series of tools:
| Category | Tool | Function |
|---|---|---|
| Migration | msTransplant | 【Analysis and migration】 One-click migration of PyTorch training scripts to the Ascend NPU, supporting migration with minimal or zero code changes. |
| Accuracy | msProbe | 【Accuracy debugging】 Ascend full-scenario accuracy tool for training accuracy debugging and issue locating. |
| Accuracy | TensorBoard | 【Hierarchical visualization】 Hierarchically visualizes model structure and accuracy, supporting debugging and comparison with benchmark models to locate accuracy issues. |
| Performance | msProf | 【Model tuning】 Full-scenario performance tuning base, profiling CANN and NPU data to improve device tuning efficiency. |
| Performance | msprof-analyze | 【Performance analysis】 Performs performance analysis based on collected data to quickly identify performance bottlenecks. |
| Performance | msMemScope | 【Memory tuning】 Dedicated memory tuning tool for multi-dimensional memory collection at the entire network level, supporting automatic diagnosis and optimization analysis. |
| Performance | msInsight | 【Visualized tuning】 Visualized performance analysis covering system, operator, and service-oriented scenarios, assisting in performance diagnosis. |
| Performance | Tinker | 【Parallelism optimization】 Automatic parallelism strategy optimization for foundation modele. It evaluates single-node NPU based on training scripts and recommends high-performance parallel solutions. |
| Performance | bind_core | 【One-click core binding】 CPU core binding tool that binds cores according to CPU affinity policies without intrusive project modifications. |
| Performance | msPTI | 【Performance profiling】 Ascend-oriented profiling API, enabling the development of NPU application performance analysis tools. |
| Monitoring | msMonitor | 【Online monitoring】 One-stop monitoring, supporting drive persistence and online sampling, for cluster monitoring and issue locating. |
🚀 Quick Start
For PyTorch and MindSpore scenarios, executable samples are provided to connect migration analysis, accuracy debugging, and performance tuning, helping you quickly get started with end-to-end training optimization.
| Training Framework | Quick Start Guide |
|---|---|
| PyTorch | msTT Quick Start in PyTorch Scenarios |
| MindSpore | msTT Quick Start in MindSpore Scenarios |
📘 User Guide
For detailed usage instructions of each tool, refer to the README file in its source code repository, or directly jump via the links in the function introduction table above.
🛠️ Contribution Guide
Contributions to the project are welcome. For details, see Contribution Guide.
⚖️ Related Notes
🔹 Release Notes
🔹 License Notice
🔹 Security Statement
🔹 Disclaimer
🤝 Suggestions and Communication
Everyone is welcome to contribute to the community. If you have any questions or suggestions, please submit an Issue, and we will respond as soon as possible. Thank you for your support.
| 📱 Follow the MindStudio Official Account | 💬 More Communication and Support |
|---|---|
 Scan the code to follow for the latest updates |
💡 Join the WeChat Group: Follow the official account and reply "communication group" to get the QR code for joining the group. 🛠️ Other Channels: 👉 Ascend Assistant:  👉 Ascend Forum:  |
🙏 Acknowledgments
msTT is jointly contributed by the following departments of Huawei:
🔹 Ascend Computing MindStudio Development Dept
🔹 Ascend Computing Ecosystem Enablement Dept
🔹 Ascend AI Cloud Service
🔹 2012 Distributed Parallel Computing Lab
🔹 2012 Network Technology Lab
Thanks for every PR from the community, and contributions to msTT are welcome.