Enabling the MIG mode on a GPU server

Introduction

This tutorial explains how to enable Multi-instance GPU (MIG). It follows the prerequisites in the official MIG documentation.

Prerequisites

A server with:
- Ubuntu 24.04
- A GPU that supports MIG (see official list)
- Access to the root user or a user with sudo permissions
An account with NVIDIA for the NVIDIA Display Mode Selector Tool

Step 1 - Install prerequisites

You can use the official configurator to check which commands are required on your system to download the NVIDIA CUDA Toolkit and install the local repository.

Following explains the commands required on Ubuntu 24.04 with x86_64 and deb (local).

Update the system and install Linux Kernel headers

sudo apt update
sudo apt install linux-headers-generic

Download and set up the APT pinning file and the CUDA repository

wget -O cuda.pin https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-ubuntu2404.pin
wget -O cuda.deb https://developer.download.nvidia.com/compute/cuda/13.2.1/local_installers/cuda-repo-ubuntu2404-13-2-local_13.2.1-595.58.03-1_amd64.deb

sudo mv cuda.pin /etc/apt/preferences.d/cuda-repository-pin-600
sudo cp /var/cuda-repo-ubuntu2404-13-2-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo dpkg -i cuda.deb

Install the NVIDIA CUDA Toolkit and the NVIDIA Driver open kernel module flavor:
```
sudo apt-get update
sudo apt-get -y install cuda-toolkit-13-2
sudo apt-get install -y nvidia-open
reboot
```
Note that the installation can take several minutes.

After reboot, verify that everything was installed as expected:
```
nvidia-smi --query-gpu=driver_version --format=csv,noheader
```
You should see the driver version.
If the command fails, the driver may not be loaded correctly.

Step 2 - Set up the NVIDIA Display Mode Selector Tool

Before you continue, you need to download the Display Mode Selector Tool file from NVIDIA. This requires an account. For more information, see the official NVIDIA documentation.

Once the ZIP file is downloaded (e.g. NVIDIA-Display-Mode-Selector-Tool-1.72.0-July25.zip), copy it to your server and unzip it as explained below.

If the ZIP file is on your local machine, run this command to copy it to the server, then connect to the server:
```
scp NVIDIA-Display-Mode-Selector-Tool-1.72.0-July25.zip user@<ip_address>:~/
ssh user@<ip_address>
```
You should find the ZIP file in your home directory.

On the server, move the ZIP file into a new directory and unzip it:

mkdir ~/displaymodeselector && cd ~/displaymodeselector
mv ~/NVIDIA-Display-Mode-Selector-Tool-1.72.0-July25.zip ~/displaymodeselector
unzip NVIDIA-Display-Mode-Selector-Tool-1.72.0-July25.zip

In the unzipped directory, you should find a file called displaymodeselector. For convenience, you can move the file directly into ~/displaymodeselector.

cp "NVIDIA Display Mode Selector Tool - 1.72.0-July25/linux/x64/displaymodeselector" .

Your directory should look similar to this now:

holu@example:~/displaymodeselector$ ls -al
-rw-rw-r--  displaymodeselector
drwxrwxr-x  __MACOSX
drwx------  'NVIDIA Display Mode Selector Tool - 1.72.0-July25'
-rwxrwxr-x  NVIDIA-Display-Mode-Selector-Tool-1.72.0-July25.zip

Make displaymodeselector executable and check the version to verify it works
```
chmod +x displaymodeselector
sudo ./displaymodeselector --version
```
In the output, you should get something like NVIDIA Display Mode Selector Utility (Version #.##.#).

Step 3 - Change the GPU mode

MIG requires the display mode to be set to compute.

Check current display mode

nvidia-smi -i 0 --query-gpu=pci.bus_id,mig.mode.current --format=csv

In the output, it might say [N/A].

Set the display mode to "compute" and reboot
```
sudo ./displaymodeselector --gpumode compute --auto
sudo poweroff
```
Wait 5 minutes before starting the server again.

After reboot, verify the new display mode:
```
nvidia-smi -i 0 --query-gpu=pci.bus_id,mig.mode.current --format=csv
```
In the output, it should now say Disabled.

Step 4 - Enable MIG

Now that the prerequisites are met, you can enable Multi-Instance GPU (MIG).

Check GPU status before the change:
```
nvidia-smi
```

Enable MIG:
```
nvidia-smi -i 0 -mig 1
```
The output should say Enabled MIG Mode.

Check GPU status, kernel version, and GPU vBIOS version:
```
nvidia-smi
uname -r
nvidia-smi --query-gpu=vbios_version --format=csv
```
In the output, you should see MIG devices. If you have a RTX PRO 6000 Blackwell GPU, compare your vBIOS version with the minimum vBIOS version mentioned in the NVIDIA documentation.

Conclusion

Multi-Instance GPU (MIG) should now be enabled on your system and you can start creating and managing GPU instances as explained in the official documentation:

Enabling the MIG mode on a GPU server

Author

Published

Time to read

Table of Contents