DOCSEARCHbyAlgolia
Submit Docs

Megrez Computing Platform

website·Indexed 31 days ago

https://docs.megrez.xsheng-ai.com

A GPU cluster management platform for deep learning, featuring containerized instance management, node orchestration, and image control. The platform supports Ubuntu-based environments with MiniConda, providing tools for performance monitoring, resource allocation (CPU, RAM, GPU), and integration with IDEs like VSCode, Jupyter Lab, and SSH.

Tokens
11.1K
Snippets
46
Records
109
Agent score
50%

What's inside Megrez

  1. Overview of MEGREZ High-Performance Computing Platform

    MEGREZ (Chinese name: 天权) is a modern, elegant, open-source high-performance computing platform designed to make machine learning more accessible and efficient for developers. Its primary goal is to make extendable GPU resource management easy.

  3. Understand Code-Server features and persistence

    Code-Server provides a full IDE experience similar to local VSCode, supporting the installation of various extensions. A key advantage of the web version is that all operations, including terminal commands, are saved in real-time. This ensures that running processes are not interrupted by network fluctuations or disconnections.

  4. Understand Megrez platform default system environment

    The platform provides images based on Ubuntu, primarily Ubuntu 22.04 (with some Ubuntu 20.04). These images come pre-packaged with deep learning frameworks, Python, and CUDA environments as specified in the image descriptions. If you require different versions of Python, CUDA, or other frameworks, you must install them manually.

  6. View server node resource status

    To monitor the status of server nodes, navigate to the node list. The system provides a comprehensive overview of each node's resources, including:

    • Node Name and IP address
    • GPU Type
    • GPU Quantity and current usage
    • Data disk total size and current usage
    • CPU allocation per GPU card
    • Memory allocation per GPU card

  7. Use pre-installed MiniConda for environment management

    All built-in images on the platform come with MiniConda pre-installed at /root/miniconda3/. This allows users to easily build environments with specific versions of CUDA or cuDNN. For a clean environment without pre-installed deep learning frameworks, it is recommended to select the Ubuntu 22.04 image when creating an instance.

  9. Add a new server node to the Megrez panel

    To add a new node, navigate to the management panel and click the 'Add Node' button. You will be required to provide connectivity details and hardware specifications for the controlled end. Note that automatic detection for CPU, memory, disks, GPU types, and driver versions is currently under development.

  10. Access and log into Jupyter Lab

    To access Jupyter Lab, click the Jupyter Lab button in the interface. When prompted, enter your SSH password to complete the login process. Once the Jupyter Lab interface appears, you can use it as you would a local installation.

    Important Note: The default working directory for Jupyter Lab is the /root directory, not the system root directory (/).

  11. Access and log in to Code-Server (Web VSCode)

    To access the web-based VSCode environment, click the 'VSCode Web' button located to the right of the corresponding instance. Upon clicking, a prompt will appear requesting the SSH password; enter the password to log in and access the Code-Server interface.

  12. Install CUDA and cuDNN using Conda

    Conda provides a simple installation method, but it is generally limited to CUDA versions prior to 11.8 and typically does not include header files. If you need to compile code, use the apt installation method instead.
    # Search for available versions
conda search cudatoolkit

# Install specific versions
conda install cudatoolkit=12.1
conda install cudnn=8.9.0