Configure Multi-Instance GPU

This guide describes how to configure Multi-Instance GPU (MIG) on a Compute Pool. MIG partitions a supported NVIDIA GPU into as many as seven isolated instances so that several workloads can share a single physical GPU.

You can configure MIG in two ways.

• Day 1 - Set a MIG configuration on a worker pool while you create the Compute Pool. The worker nodes are partitioned as they join.
• Day 2 - Change the MIG configuration on a running Compute Pool. The GPU operator reconfigures the nodes in place.

Refer to the Multi-Instance GPU concept page for background on MIG profiles, configurations, and strategies.

Prerequisites

• A user with Project editor or admin permissions. For details, refer to Project Role Permissions.

• MIG configuration must be enabled for your scope. MIG UI visibility is gated at the System, Tenant, and Project scope, with higher scopes taking precedence. Refer to System, Tenant, and Project Governance for how the scopes interact, Manage Multi-Instance GPU in Tenant Settings for the Tenant scope, and Manage Multi-Instance GPU in Project Settings for the Project scope.

• The Compute Pool must use a Profile Bundle that deploys the NVIDIA GPU Operator. Refer to Profile Bundles for the GPU operator requirements.

• The worker pool must use a MIG-capable NVIDIA GPU model. PaletteAI offers MIG configurations only for GPUs that support MIG.

warning

Changing a MIG configuration on a running Compute Pool is disruptive. The GPU operator stops the GPU pods on each affected node, applies the new partitions, and, on some hardware, reboots the node. PaletteAI pauses autoscaling for the Compute Pool until the reconfiguration settles. Plan MIG changes during a maintenance window.

Configure MIG on Day 1

Set a MIG configuration while you create the Compute Pool so that the worker nodes are partitioned as they are provisioned.

1. Follow Create and Manage Compute Pools to start the Compute Pool creation wizard.

2. On the Node config page, select the worker pool that uses a MIG-capable NVIDIA GPU.

3. Expand GPU Resources and confirm that the worker pool specifies the NVIDIA GPU variant and count.

4. Select a MIG Configuration for the worker pool from the drop-down menu.

   The drop-down lists only the configurations that the selected GPU model supports, such as all-1g.10gb. The MIG Configuration control appears only when the System, Tenant, and Project allow MIG and the worker pool uses a MIG-capable GPU.

5. Repeat the MIG selection for each worker pool that you want to partition.

6. Complete the remaining wizard steps and submit the Compute Pool.

When the worker nodes join, the NVIDIA GPU Operator partitions their GPUs to match the selected configuration.

Configure MIG on Day 2

Change the MIG configuration on a running Compute Pool. The first Day 2 MIG change moves the pool from Day 1 management to Day 2 management. After this transition, PaletteAI manages MIG through the running pool's per-worker-pool configuration, and the transition is permanent for the life of the pool.

1. Log in to PaletteAI, and then open your Project.

2. From the left main menu, select Compute Pools.

3. Select the Compute Pool that you want to update.

4. Locate the MIG configuration control for the worker pool that you want to change.

   PaletteAI presents one MIG control per worker pool. The control appears only when the System, Tenant, and Project allow MIG and the worker pool uses a MIG-capable GPU.

5. Select a new MIG Configuration from the drop-down menu, and then save the change.

PaletteAI pauses autoscaling for the Compute Pool, and the NVIDIA GPU Operator reconfigures the affected nodes. Autoscaling resumes when the reconfiguration settles.

info

After a Compute Pool transitions to Day 2 management, you can no longer change its MIG configuration through the worker pool labels used for Day 1. Make all further MIG changes through the per-worker-pool MIG control described in this section.

Disable MIG

To remove the partitions and return the GPUs to a whole state, set the MIG configuration to all-disabled rather than clearing the control. The GPU operator only reconfigures a GPU when the configuration changes, so setting all-disabled is what reverts the partitions.

1. Open the Compute Pool and locate the MIG configuration control for the worker pool.

2. Select all-disabled from the drop-down menu, and then save the change.

The GPU operator un-partitions the nodes and restores each GPU to a single whole device. PaletteAI tracks this rollout the same way it tracks any other MIG change and resumes autoscaling when the rollout settles.

Validate

1. Open the Compute Pool and review the MIG reconfiguration progress for each worker pool.

   PaletteAI reports the node counts in the pending, rebooting, success, and failed states while the GPU operator applies the change.

2. Confirm that the targeted nodes reach the success state.

3. Confirm that the Compute Pool's hardware capacity reflects the partitioned GPUs.

   You can inspect the Compute Pool status directly with the following command.

   kubectl get computepool <pool-name> --namespace <project-namespace> --output yaml

   The status.hardwareCapacity GPU entries include a mig section that reports the applied strategy, configuration, reconfiguration state, and per-GPU device layout. Refer to MIG Hardware Capacity for the field reference.

4. Confirm that autoscaling resumes after the reconfiguration settles.

If a node reports the failed state, refer to Troubleshooting Compute Pools and review the Compute Pool events.

Next Steps

• Multi-Instance GPU to review MIG concepts, strategies, and governance.

• Create and Manage Scaling Policies to configure autoscaling for the Compute Pool.

• Configure Tenant Settings to govern which Projects may configure MIG.