add amd gpu test

Author: r2k1

e2e/config/azure.go

@@ -291,25 +291,27 @@ func (a *AzureClient) UploadAndGetSignedLink(ctx context.Context, blobName strin
 }
 
 func (a *AzureClient) CreateVMManagedIdentity(ctx context.Context) (string, error) {
-	identity, err := a.UserAssignedIdentities.CreateOrUpdate(ctx, ResourceGroupName, VMIdentityName, armmsi.Identity{
-		Location: to.Ptr(Config.Location),
-	}, nil)
-	if err != nil {
-		return "", fmt.Errorf("create managed identity: %w", err)
-	}
-	err = a.createBlobStorageAccount(ctx)
-	if err != nil {
-		return "", err
-	}
-	err = a.createBlobStorageContainer(ctx)
-	if err != nil {
-		return "", err
-	}
-
-	if err := a.assignRolesToVMIdentity(ctx, identity.Properties.PrincipalID); err != nil {
-		return "", err
-	}
-	return *identity.Properties.ClientID, nil
+	// HACK: temporary disable to allow running test in different subscription, without enough permissions
+	return "", nil
+	// identity, err := a.UserAssignedIdentities.CreateOrUpdate(ctx, ResourceGroupName, VMIdentityName, armmsi.Identity{
+	// 	Location: to.Ptr(Config.Location),
+	// }, nil)
+	// if err != nil {
+	// 	return "", fmt.Errorf("create managed identity: %w", err)
+	// }
+	// err = a.createBlobStorageAccount(ctx)
+	// if err != nil {
+	// 	return "", err
+	// }
+	// err = a.createBlobStorageContainer(ctx)
+	// if err != nil {
+	// 	return "", err
+	// }
+
+	// if err := a.assignRolesToVMIdentity(ctx, identity.Properties.PrincipalID); err != nil {
+	// 	return "", err
+	// }
+	// return *identity.Properties.ClientID, nil
 }
 
 func (a *AzureClient) createBlobStorageAccount(ctx context.Context) error {

e2e/kube.go

@@ -637,3 +637,52 @@ func nvidiaDevicePluginDaemonSet() *appsv1.DaemonSet {
 		},
 	}
 }
+
+func podEnableAMDGPUResource(s *Scenario) *corev1.Pod {
+	return &corev1.Pod{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      fmt.Sprintf("%s-amdgpu-device-plugin", s.Runtime.KubeNodeName),
+			Namespace: defaultNamespace,
+		},
+		Spec: corev1.PodSpec{
+			PriorityClassName: "system-node-critical",
+			NodeSelector: map[string]string{
+				"kubernetes.io/hostname": s.Runtime.KubeNodeName,
+			},
+			Containers: []corev1.Container{
+				{
+					Name:  "amdgpu-device-plugin-container",
+					Image: "rocm/k8s-device-plugin",
+					VolumeMounts: []corev1.VolumeMount{
+						{
+							Name:      "device-plugin",
+							MountPath: "/var/lib/kubelet/device-plugins",
+						},
+						{
+							Name:      "sys",
+							MountPath: "/sys",
+						},
+					},
+				},
+			},
+			Volumes: []corev1.Volume{
+				{
+					Name: "device-plugin",
+					VolumeSource: corev1.VolumeSource{
+						HostPath: &corev1.HostPathVolumeSource{
+							Path: "/var/lib/kubelet/device-plugins",
+						},
+					},
+				},
+				{
+					Name: "sys",
+					VolumeSource: corev1.VolumeSource{
+						HostPath: &corev1.HostPathVolumeSource{
+							Path: "/sys",
+						},
+					},
+				},
+			},
+		},
+	}
+}

e2e/scenario_test.go

@@ -1664,3 +1664,64 @@ func Test_Ubuntu2404ARM(t *testing.T) {
 		},
 	})
 }
+
+func Test_Ubuntu2204Gen2Containerd_AMDGPU_MI300(t *testing.T) {
+	//t.Skip("Provisioning of Standard_ND96isr_MI300X_v5 isn't reliable yet")
+	RunScenario(t, &Scenario{
+		Description: "Tests that a GPU-enabled node using a MarinerV2 VHD can be properly bootstrapped",
+		Tags: Tags{
+			GPU: true,
+		},
+		Config: Config{
+			Cluster: ClusterKubenet,
+			VHD:     config.VHDUbuntu2204Gen2Containerd, //TODO: add support for older
+			BootstrapConfigMutator: func(nbc *datamodel.NodeBootstrappingConfiguration) {
+				nbc.ContainerService.Properties.AgentPoolProfiles[0].VMSize = "Standard_ND96isr_MI300X_v5"
+				nbc.ContainerService.Properties.AgentPoolProfiles[0].Distro = "aks-cblmariner-v2-gen2"
+				nbc.AgentPoolProfile.VMSize = "Standard_ND96isr_MI300X_v5"
+				nbc.AgentPoolProfile.Distro = "aks-cblmariner-v2-gen2"
+				nbc.EnableAMDGPU = true
+				nbc.ConfigGPUDriverIfNeeded = true
+			},
+			VMConfigMutator: func(vmss *armcompute.VirtualMachineScaleSet) {
+				vmss.SKU.Name = to.Ptr("Standard_ND96isr_MI300X_v5")
+				vmss.Properties.VirtualMachineProfile.StorageProfile.OSDisk.DiskSizeGB = to.Ptr[int32](128) // drivers and gpu images are huge, give us some headroom
+			},
+			Validator: func(ctx context.Context, s *Scenario) {
+				ValidateAMDGPU(ctx, s)
+			},
+		},
+	})
+}
+
+func Test_Ubuntu2204Gen2Containerd_AMDGPU_V710(t *testing.T) {
+	// the SKU isn't available in subscriptrion/region we run tests
+	//t.Skip("Provisioning of NV4ads_V710_v5 isn't reliable yet")
+	// LOCATION=southcentralus
+	RunScenario(t, &Scenario{
+		Description: "Tests that a GPU-enabled node using a MarinerV2 VHD can be properly bootstrapped",
+		Tags: Tags{
+			GPU: true,
+		},
+		Config: Config{
+			Cluster: ClusterKubenet,
+			VHD:     config.VHDUbuntu2204Gen2Containerd,
+			BootstrapConfigMutator: func(nbc *datamodel.NodeBootstrappingConfiguration) {
+				nbc.ContainerService.Properties.AgentPoolProfiles[0].VMSize = "Standard_NV4ads_V710_v5"
+				nbc.ContainerService.Properties.AgentPoolProfiles[0].Distro = "aks-cblmariner-v2-gen2"
+				nbc.AgentPoolProfile.VMSize = "Standard_NV4ads_V710_v5"
+				nbc.AgentPoolProfile.Distro = "aks-cblmariner-v2-gen2"
+				nbc.EnableAMDGPU = true
+				nbc.ConfigGPUDriverIfNeeded = true
+
+			},
+			VMConfigMutator: func(vmss *armcompute.VirtualMachineScaleSet) {
+				vmss.SKU.Name = to.Ptr("Standard_NV4ads_V710_v5")
+				vmss.Properties.VirtualMachineProfile.StorageProfile.OSDisk.DiskSizeGB = to.Ptr[int32](128) // drivers and gpu images are huge, give us some headroom
+			},
+			Validator: func(ctx context.Context, s *Scenario) {
+				ValidateAMDGPU(ctx, s)
+			},
+		},
+	})
+}

e2e/validators.go

@@ -427,3 +427,15 @@ func GetFieldFromJsonObjectOnNode(ctx context.Context, s *Scenario, fileName str
 
 	return podExecResult.stdout.String()
 }
+
+func ValidateAMDGPU(ctx context.Context, s *Scenario) {
+	s.T.Logf("validating pod using AMD GPU")
+
+	execResult := execScriptOnVMForScenario(ctx, s, "lspci -k")
+	require.Equal(s.T, "0", execResult.exitCode, "expected to find lspci command, but did not")
+	assert.Contains(s.T, execResult.stdout.String(), "amdgpu", "expected to see amdgpu kernel module managing a PCI device, but did not")
+
+	ensurePod(ctx, s, podEnableAMDGPUResource(s))
+	waitUntilResourceAvailable(ctx, s, "amd.com/gpu")
+	//ensureJob(ctx, s, jobAMDGPUWorkload(s))
+}