Installing the MindIO ACP SDK on Compute Nodes

The MindIO ACP SDK is used to adapt to Torch and MindSpore to accelerate checkpoint saving and loading.

Procedure

  1. Log in to the installation node as the {MindIO-install-user} user.

    The password set by the installation user must meet the password complexity requirements. For details, see Password Complexity Requirements. The password validity period is 90 days. You can change the validity period in the /etc/login.defs file or run the chage command to set the validity period. For details, see Setting the Validity Period of a User Account.

  2. Upload the memory cache system package to a path on the device on which the installation user has the read and write permissions.
    • Use the actual memory cache system package.
    • If the Python environment is a shared directory, upload the installation package to any compute node. Otherwise, upload the installation package to all compute nodes.
  3. Go to the package upload path and decompress the memory cache system package.
    unzip Ascend-mindxdl-mindio_{version}_linux-{arch}.zip
    Table 1 Extracted files

    File

    Description

    mindio_acp-{mindio_acp_version}-py3-none-linux_{arch}.whl

    MindIO ACP installation package

    mindio_ttp-{mindio_ttp_version}-py3-none-linux_{arch}.whl

    MindIO TFT installation package

  4. Go to the upload path and install the MindIO ACP SDK.
    pip3 install mindio_acp-{mindio_acp_version}-py3-none-linux_{arch}.whl --force-reinstall
    • If the following information is displayed when the MindIO ACP SDK is installed for the first time, the installation is successful.
      1
      2
      3
      Processing ./mindio_acp-{mindio_acp_version}-py3-none-linux_{arch}.whl
      Installing collected packages: mindio_acp
      Successfully installed mindio_acp-{version}
      
    • If the following information is displayed when the MindIO ACP SDK is not installed for the first time, the installation is successful.
      1
      2
      3
      4
      5
      6
      7
      Processing ./mindio_acp-{mindio_acp_version}-py3-none-linux_{arch}.whl
       Installing collected packages: mindio_acp
         Attempting uninstall: mindio_acp
           Found existing installation: mindio_acp{version}
           Uninstalling mindio_acp{version}:
             Successfully uninstalled mindio_acp{version}
       Successfully installed mindio_acp{version}
      
  5. Change the permission on executable files and code scripts in the software installation directory to 550 to prevent unauthorized tampering.
    chmod -R 550 {MindIO ACPSDK installation directory}