Configure and deploy the device config manager

Now that your control-plane workloads are protected with tolerations, you can prepare the infrastructure components that control the accelerator devices. The AMD GPU Operator manages these hardware day2 operations using a component called the device config manager (DCM).

By default, the DCM is not active. To enable it, define a ConfigMap that maps partition profiles to your hardware—for example, a single monolithic GPU (SPX) or multiple smaller slices (CPX)—so the operator knows which compute and memory layout to apply on each node.

Prerequisites

• Control plane toleration verified.
• The AMD GPU Operator installed in the openshift-amd-gpu namespace.

In this lesson, you will:

• Create a ConfigMap with GPU partition profiles
• Patch the DeviceConfig custom resource to enable and configure the DCM,  which orchestrates GPU partitioning.

Deploy DCM 

To ensure your infrastructure components are mapped correctly, create and deploy DCM profiles with ConfigMap.

1. Validate the available compute and memory partitions that your target GPU node supports with the command below. Run this on the GPU node itself (e.g., via SSH or oc debug node).

   cat /sys/module/amdgpu/drivers/pci\:amdgpu/*/{available_compute_partition,available_memory_partition}

   On a standard AMD Instinct MI300X system with an 8-GPU layout, your output will look like this:

   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   SPX, DPX, QPX, CPX
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4
   NPS1, NPS2, NPS4

2. Create DCM profiles via ConfigMap. This ConfigMap defines three partition profiles: unpartitioned (spx-profile-nps1), dual-partition (dpx-profile-nps2), and maximum-partition (cpx-profile-nps4). Run the command below to create the ConfigMap on your cluster.

   cat <<EOF | tee gpu-partition-profiles.yaml | oc apply -f -
   ---
   apiVersion: v1
   kind: ConfigMap
   metadata:
     name: config-manager-config
     namespace: openshift-amd-gpu
   data:
     config.json: |
       {
         "gpu-config-profiles": {
           "spx-profile-nps1": {
             "skippedGPUs": {
               "ids": []
             },
             "profiles": [
               {
                 "computePartition": "SPX",
                 "memoryPartition": "NPS1",
                 "numGPUsAssigned": 8
               }
             ]
           },
           "dpx-profile-nps2": {
             "skippedGPUs": {
               "ids": []
             },
             "profiles": [
               {
                 "computePartition": "DPX",
                 "memoryPartition": "NPS2",
                 "numGPUsAssigned": 8
               }
             ]
           },
           "cpx-profile-nps4": {
             "skippedGPUs": {
               "ids": []
             },
             "profiles": [
               {
                 "computePartition": "CPX",
                 "memoryPartition": "NPS4",
                 "numGPUsAssigned": 8
               }
             ]
           }
         },
         "gpuClientSystemdServices": {
           "names": ["amd-metrics-exporter", "gpuagent"]
         }
       }
   EOF

3. Patch the DeviceConfig custom resource to reference the ConfigMap and enable DCM.

   oc patch deviceconfig amdgpu-driver-install -n openshift-amd-gpu --type='merge' -p '{
     "spec": {
       "configManager": {
         "enable": true,
         "image": "docker.io/rocm/device-config-manager:v1.4.1",
         "imagePullPolicy": "IfNotPresent",
         "config": {
           "name": "config-manager-config"
         }
       }
     }
   }'

4. To determine which DCM versions are publicly available, you can use Skopeo to list the tags released by AMD.

   skopeo list-tags docker://docker.io/rocm/device-config-manager
   {
       "Repository": "docker.io/rocm/device-config-manager",
       "Tags": [
           "v1.3.0",
           "v1.3.0-beta.0",
           "v1.3.1",
           "v1.3.1-beta.0",
           "v1.4.0",
           "v1.4.0-beta.0",
           "v1.4.1"
       ]
   }

5. Wait for the DCM Pod to be ready:

   oc wait --for=condition=ready pod \
     -l app.kubernetes.io/name=device-config-manager \
     -n openshift-amd-gpu \
     --timeout=300s

You have verified your GPU's available partition modes, created the partition profiles, and enabled DCM. 

In the next lesson, you will taint and label the GPU node to trigger the partitioning process.