Help Center> ModelArts> Troubleshooting> Training Jobs> Permission Issues> Error Message "Permission denied" Displayed in Logs
Updated on 2024-04-30 GMT+08:00

Error Message "Permission denied" Displayed in Logs

Symptom

When a training job accesses the attached EFS disks or executes the .sh boot script, an error occurs.

  • [Errno 13]Permission denied: '/xxx/xxxx'
    Figure 1 Error log
  • bash: /bin/ln: Permission denied
  • bash:/home/ma-user/.pip/pip.conf: Permission Denied (in a custom image)
  • tee: /xxx/xxxx: Permission denied cp: cannot stat '' No such file or directory (in a custom image)

Possible Causes

The possible causes are as follows:

  • [Errno 13]Permission denied: '/xxx/xxxx'
    • When data is uploaded, the ownership and permissions to the file are not changed. As a result, the work user group does not have the permission to access the training job.
    • After the .sh file in the code directory is copied to the container, the execution permission is not granted for the file.
  • bash: /bin/ln: Permission denied

    For security purposes, the ln command is not supported.

  • bash:/home/ma-user/.pip/pip.conf: Permission Denied

    After the version of training jobs is switched from V1 to V2, the UID of the ma-user user is still 1102.

  • tee: /xxx/xxxx: Permission denied cp: cannot stat '': No such file or directory

    The used startup script is run_train.sh of an earlier version. Some environment variables in the script are unavailable in the training jobs of the new version.

  • The APIs using the Python file concurrently read and write the same file.

Solution

  1. Add permissions to access the attached EFS disks so that the permissions are the same as those of user group (1000) used in the training container. For example, if the /nas disk is attached, run the following command:
    chown -R 1000: 1000 /nas
    Or
    chmod 777 -R /nas
  2. If the execution permission has not been granted for the .sh file used by the custom image, run chmod +x xxx.sh to grant the permission before starting the script.
  3. On the ModelArts console, if a training job is created using a custom image, a V2 container image is started using UID 1000 by default. In this case, change the UID of the ma-user user from 1102 to 1000. To obtain the sudo permission, comment out the sudoers line.

  4. Migrate environment variables from V1 training jobs to V2 training jobs.
    • Use V2 MA_NUM_HOSTS (the number of selected training nodes) to replace V1 DLS_TASK_NUMBER.
    • Use V2 VC_TASK_INDEX (or MA_TASK_INDEX that will be available later) to replace V1 DLS_TASK_INDEX. Obtain the environment variable using the method provided in the demo script for compatibility.
    • Use V2 ${MA_VJ_NAME}-${MA_TASK_NAME}-0.${MA_VJ_NAME}:6666 to replace V1 BATCH_CUSTOM0_HOSTS.
    • Use V2 ${MA_VJ_NAME}-${MA_TASK_NAME}-{N}.${MA_VJ_NAME}:6666 to replace V1 BATCH_CUSTOM{N}_HOSTS generally.
  5. Check whether there are settings that allow concurrent reading and writing of the same file in the code. If so, modify the settings to forbid this operation.

    If a job uses multiple cards, the same code for reading and writing data may be available on each card. In this case, do as follows to modify the code:

    import moxing as mox
    from mindspore.communication import init, get_rank, get_group_size
    init()
    rank_id = get_rank()
    # Enable only card 0 to download data.
    if rank_id % 8 == 0:
        mox.file.copy_parallel('obs://bucket-name/dir1/dir2/', '/cache')

Summary and Suggestions

Before creating a training job, use the ModelArts development environment to debug the training code to maximally eliminate errors in code migration.