Редактиране

Споделяне чрез


Troubleshoot GPU extension issues for GPU VMs on Azure Stack Edge Pro GPU

APPLIES TO: Yes for Pro - GPU SKUAzure Stack Edge Pro - GPUYes for Pro 2 SKUAzure Stack Edge Pro 2Yes for Pro R SKUAzure Stack Edge Pro R  

This article gives guidance for resolving the most common issues that cause installation of the GPU extension on a GPU VM to fail on an Azure Stack Edge Pro GPU device.

For installation steps, see Install GPU extension.

In versions lower than 2205, Linux GPU extension installs old signing keys: signature and/or required key missing

Error description: The Linux GPU extension installs old signing keys, preventing download of the required GPU driver. In this case, you'll see the following error in the syslog of the Linux VM:

/var/log/syslog and /var/log/waagent.log 
May  5 06:04:53 gpuvm12 kernel: [  833.601805] nvidia:module verification failed: signature and/or required key missing- tainting kernel 

Suggested solutions: You have two options to mitigate this issue:

  • Option 1: Apply the Azure Stack Edge 2205 updates to your device.

  • Option 2: After creating a GPU virtual machine of size in NCasT4_v3-series, manually install the new signing keys before installing the extension, then set required signing keys using steps in Updating the CUDA Linux GPG Repository Key | NVIDIA Technical Blog.

    Here's an example that installs signing keys on an Ubuntu 1804 virtual machine:

    $ sudo apt-key adv --fetch-
    keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/3bf863cc.pub 
    

Failure to install GPU extension on a Windows 2016 VHD

Error description: This is a known issue in versions lower than 2205. The GPU extension requires TLS 1.2. In this case, you may see the following error message:

Failed to download https://go.microsoft.com/fwlink/?linkid=871664 after 10 attempts. Exiting!

Additional details:

  • Check the guest log for the associated error. To collect the guest logs, see Collect guest logs for VMs on an Azure Stack Edge Pro GPU device.
  • On a Linux VM, look in /var/log/waagent.log or /var/log/azure/nvidia-vmext-status.
  • On a Windows VM, find the error status in C:\Packages\Plugins\Microsoft.HpcCompute.NvidiaGpuDriverWindows\1.3.0.0\Status.
  • Review the complete execution log in C:\WindowsAzure\Logs\WaAppAgent.txt.

If the installation failed during the package download, that error indicates the VM couldn't access the public network to download the driver.

Suggested solution: Use the following steps to enable TLS 1.2 on a Windows 2016 VM, and then deploy the GPU extension.

  1. Run the following command inside the VM to enable TLS 1.2:

    sp hklm:\SOFTWARE\Microsoft\.NETFramework\v4.0.30319 SchUseStrongCrypto 1
    
  2. Deploy the template addGPUextensiontoVM.json to install the extension on an existing VM. You can install the extension manually, or you can install the extension from the Azure portal.

    Note

    The extension deployment is a long running job and takes about 10 minutes to complete.

Manually install the Nvidia driver on RHEL 7

Error description: When installing the GPU extension on an RHEL 7 VM, the installation may fail due to a certificate rotation issue and an incompatible driver version.

Suggested solution: In this case, you have two options:

  • Option 1: Resolve the certificate rotation issue and then install an Nvidia driver lower than version 510.

    1. To resolve the certificate rotation issue, run the following command:

      $ sudo yum-config-manager --add-repo  https://developer.download.nvidia.com/compute/cuda/repos/rhel7/$arch/cuda-rhel7.repo
      
    2. Install an Nvidia driver lower than version 510.

  • Option 2: Deploy the GPU extension. Use the following settings when deploying the ARM extension:

    settings": { 
    "isCustomInstall": true, 
    "InstallMethod": 0, 
    "DRIVER_URL": "  https://developer.download.nvidia.com/compute/cuda/11.4.4/local_installers/cuda-repo-rhel7-11-4-local-11.4.4_470.82.01-1.x86_64.rpm", 
    "DKMS_URL" : "  https://dl.fedoraproject.org/pub/epel/epel-release-latest-7.noarch.rpm", 
    "LIS_URL": "  https://aka.ms/lis", 
    "LIS_RHEL_ver": "3.10.0-1062.9.1.el7" 
    } 
    

VM size is not GPU VM size

Error description: A GPU VM must be either Standard_NC4as_T4_v3 or Standard_NC8as_T4_v3 size. If any other VM size is used, the GPU extension will fail to be attached.

Suggested solution: Create a VM with the Standard_NC4as_T4_v3 or Standard_NC8as_T4_v3 VM size. For more information, see Supported VM sizes for GPU VMs. For information about specifying the size, see Create GPU VMs.

Image OS is not supported

Error description: The GPU extension doesn't support the operating system that's installed on the VM image.

Suggested solution: Prepare a new VM image that has an operating system that the GPU extension supports.

Extension parameter is incorrect

Error description: Incorrect extension settings were used when deploying the GPU extension on a Linux VM.

Suggested solution: Edit the parameters file before deploying the GPU extension. For more information, see Install GPU extension.

VM extension installation failed in downloading package

Error description: Extension provisioning failed during extension installation or while in the Enable state.

  1. Check the guest log for the associated error. To collect the guest logs, see Collect guest logs for VMs on an Azure Stack Edge Pro.

    On a Linux VM:

    • Look in /var/log/waagent.log or /var/log/azure/nvidia-vmext-status.

    On a Windows VM:

    • Find out the error status in C:\Packages\Plugins\Microsoft.HpcCompute.NvidiaGpuDriverWindows\1.3.0.0\Status.
    • Review the complete execution log: C:\WindowsAzure\Logs\WaAppAgent.txt.

    If installation failed during the package download, that error indicates the VM couldn't access the public network to download the driver.

Suggested solution:

  1. Enable compute on a port that's connected to the Internet. For guidance, see Create GPU VMs.

  2. Deallocate the VM by stopping the VM in the portal. To stop the VM, go to Virtual machines > Overview, and select the VM. Then, on the VM properties page, select Stop.

  3. Create a new VM.

VM Extension failed with error dpkg is used/yum lock is used (Linux VM)

Error description: GPU extension deployment on a Linux VM failed because another process was using dpkg or another process has created a yum lock.

Suggested solution: To resolve the issue, do these steps:

  1. To find out what process is applying the lock, search the \var\log\azure\nvidia-vmext-status log for an error such as “dpkg is used by another process” or ”Another app is holding yum lock”.

  2. Either wait for the process to finish, or end the process.

  3. Install the GPU extension again.

  4. If extension deployment fails again, create a new VM and make sure the lock isn't present before you install the GPU extension.

Next steps

Collect guest logs, and create a Support package