Merge branch 'release-v0.6.x' into tallaxes/release-v0.6.x/zones

.pre-commit-config.yaml

@@ -12,7 +12,7 @@ repos:
   hooks:
   - id: shellcheck
 - repo: https://github.com/crate-ci/typos
-  rev: v1.17.2
+  rev: v1.28.1
   hooks:
   - id: typos
 - repo: https://github.com/pre-commit/pre-commit-hooks

go.mod

@@ -32,6 +32,7 @@ require (
 	github.com/stretchr/testify v1.9.0
 	go.uber.org/multierr v1.11.0
 	go.uber.org/zap v1.27.0
+	gopkg.in/yaml.v2 v2.4.0
 	k8s.io/api v0.29.3
 	k8s.io/apiextensions-apiserver v0.29.3
 	k8s.io/apimachinery v0.29.3
@@ -146,7 +147,6 @@ require (
 	google.golang.org/protobuf v1.34.2 // indirect
 	gopkg.in/dnaeon/go-vcr.v3 v3.2.0 // indirect
 	gopkg.in/inf.v0 v0.9.1 // indirect
-	gopkg.in/yaml.v2 v2.4.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 	k8s.io/cloud-provider v0.29.3 // indirect
 	k8s.io/component-base v0.29.3 // indirect

hack/codegen.sh

@@ -32,7 +32,7 @@ skugen() {
   NO_UPDATE=" pkg/fake/zz_generated.sku.$location.go | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)"
   SUBJECT="SKUGEN"
 
-  go run hack/code/instancetype_testdata_gen/main.go -- "${GENERATED_FILE}" "$location" "Standard_B1s,Standard_A0,Standard_D2_v2,Standard_D2_v3,Standard_DS2_v2,Standard_D2s_v3,Standard_D2_v5,Standard_D16plds_v5,Standard_F16s_v2,Standard_NC6s,Standard_NC6s_v3,Standard_NC16as_T4_v3,Standard_NC24ads_A100_v4,Standard_M8-2ms,Standard_D4s_v3,Standard_D64s_v3,Standard_DC8s_v3"
+  go run hack/code/instancetype_testdata_gen/main.go -- "${GENERATED_FILE}" "$location" "Standard_B1s,Standard_A0,Standard_D2_v2,Standard_D2_v3,Standard_DS2_v2,Standard_D2s_v3,Standard_D2_v5,Standard_D16plds_v5,Standard_F16s_v2,Standard_NC6s,Standard_NC6s_v3,Standard_NC16as_T4_v3,Standard_NC24ads_A100_v4,Standard_M8-2ms,Standard_D4s_v3,Standard_D64s_v3,Standard_DC8s_v3,Standard_D2as_v6"
   go fmt "${GENERATED_FILE}"
 
   GIT_DIFF=$(git diff --stat "${GENERATED_FILE}")

pkg/providers/imagefamily/azlinux.go

@@ -84,9 +84,9 @@ func (u AzureLinux) UserData(kubeletConfig *corev1beta1.KubeletConfiguration, ta
 			CABundle:         caBundle,
 			GPUNode:          u.Options.GPUNode,
 			GPUDriverVersion: u.Options.GPUDriverVersion,
-			// GPUImageSHA: u.Options.GPUImageSHA - GPU image SHA only applies to Ubuntu
-			// See: https://github.com/Azure/AgentBaker/blob/f393d6e4d689d9204d6000c85623ad9b764e2a29/vhdbuilder/packer/install-dependencies.sh#L201
-			SubnetID: u.Options.SubnetID,
+			GPUDriverType:    u.Options.GPUDriverType,
+			GPUImageSHA:      u.Options.GPUImageSHA,
+			SubnetID:         u.Options.SubnetID,
 		},
 		Arch:                           u.Options.Arch,
 		TenantID:                       u.Options.TenantID,

pkg/providers/imagefamily/bootstrap/aksbootstrap.go

@@ -205,6 +205,7 @@ type NodeBootstrapVariables struct {
 	SwapFileSizeMB                    int      // t   user input
 	GPUImageSHA                       string   // s	  static sha rarely updated
 	GPUDriverVersion                  string   // k   determine by OS + GPU hardware requirements; can be determined automatically, but hard. suggest using GPU operator.
+	GPUDriverType                     string   // k
 	GPUInstanceProfile                string   // t   user-specified
 	CustomSearchDomainName            string   // c   user-specified [presumably cluster-level]
 	CustomSearchRealmUser             string   // c   user-specified [presumably cluster-level]
@@ -468,6 +469,7 @@ func (a AKS) applyOptions(nbv *NodeBootstrapVariables) {
 		nbv.GPUNode = true
 		nbv.ConfigGPUDriverIfNeeded = true
 		nbv.GPUDriverVersion = a.GPUDriverVersion
+		nbv.GPUDriverType = a.GPUDriverType
 		nbv.GPUImageSHA = a.GPUImageSHA
 	}
 

pkg/providers/imagefamily/bootstrap/bootstrap.go

@@ -31,6 +31,7 @@ type Options struct {
 	CABundle         *string
 	GPUNode          bool
 	GPUDriverVersion string
+	GPUDriverType    string
 	GPUImageSHA      string
 	SubnetID         string
 }

pkg/providers/imagefamily/bootstrap/cse_cmd.sh.gtpl

@@ -128,6 +128,7 @@ KUBELET_CONFIG_FILE_CONTENT="{{.KubeletConfigFileContent}}"
 SWAP_FILE_SIZE_MB="{{.SwapFileSizeMB}}"
 GPU_IMAGE_SHA="{{.GPUImageSHA}}"
 GPU_DRIVER_VERSION="{{.GPUDriverVersion}}"
+GPU_DRIVER_TYPE="{{.GPUDriverType}}"
 GPU_INSTANCE_PROFILE="{{.GPUInstanceProfile}}"
 CUSTOM_SEARCH_DOMAIN_NAME="{{.CustomSearchDomainName}}"
 CUSTOM_SEARCH_REALM_USER="{{.CustomSearchRealmUser}}"

pkg/providers/imagefamily/ubuntu_2204.go

@@ -85,6 +85,7 @@ func (u Ubuntu2204) UserData(kubeletConfig *corev1beta1.KubeletConfiguration, ta
 			GPUNode:          u.Options.GPUNode,
 			GPUDriverVersion: u.Options.GPUDriverVersion,
 			GPUImageSHA:      u.Options.GPUImageSHA,
+			GPUDriverType:    u.Options.GPUDriverType,
 			SubnetID:         u.Options.SubnetID,
 		},
 		Arch:                           u.Options.Arch,

pkg/providers/instancetype/instancetypes.go

@@ -29,6 +29,7 @@ import (
 	"github.com/mitchellh/hashstructure/v2"
 	"github.com/samber/lo"
 
+	"github.com/Azure/azure-sdk-for-go/profiles/latest/compute/mgmt/compute"
 	"github.com/Azure/go-autorest/autorest/to"
 	"github.com/Azure/karpenter-provider-azure/pkg/apis/v1alpha2"
 	kcache "github.com/Azure/karpenter-provider-azure/pkg/cache"
@@ -212,7 +213,8 @@ func (p *Provider) getInstanceTypes(ctx context.Context) (map[string]*skewer.SKU
 			continue
 		}
 
-		if !skus[i].HasLocationRestriction(p.region) && p.isSupported(&skus[i], vmsize) {
+		useSIG := false // replace with options.FromContext(ctx).UseSIG when available
+		if !skus[i].HasLocationRestriction(p.region) && p.isSupported(&skus[i], vmsize, useSIG) {
 			instanceTypes[skus[i].GetName()] = &skus[i]
 		}
 	}
@@ -229,13 +231,14 @@ func (p *Provider) getInstanceTypes(ctx context.Context) (map[string]*skewer.SKU
 }
 
 // isSupported indicates SKU is supported by AKS, based on SKU properties
-func (p *Provider) isSupported(sku *skewer.SKU, vmsize *skewer.VMSizeType) bool {
+func (p *Provider) isSupported(sku *skewer.SKU, vmsize *skewer.VMSizeType, useSIG bool) bool {
 	return p.hasMinimumCPU(sku) &&
 		p.hasMinimumMemory(sku) &&
 		!p.isUnsupportedByAKS(sku) &&
 		!p.isUnsupportedGPU(sku) &&
 		!p.hasConstrainedCPUs(vmsize) &&
-		!p.isConfidential(sku)
+		!p.isConfidential(sku) &&
+		isCompatibleImageAvailable(sku, useSIG)
 }
 
 // at least 2 cpus
@@ -344,3 +347,15 @@ var (
 func hasZonalSupport(region string) bool {
 	return zonalRegions.Has(region)
 }
+
+func isCompatibleImageAvailable(sku *skewer.SKU, useSIG bool) bool {
+	hasSCSISupport := func(sku *skewer.SKU) bool { // TODO: move capability determination to skewer
+		const diskControllerTypeCapability = "DiskControllerTypes"
+		declaresSCSI := sku.HasCapabilityWithSeparator(diskControllerTypeCapability, string(compute.SCSI))
+		declaresNVMe := sku.HasCapabilityWithSeparator(diskControllerTypeCapability, string(compute.NVMe))
+		declaresNothing := !(declaresSCSI || declaresNVMe)
+		return declaresSCSI || declaresNothing // if nothing is declared, assume SCSI is supported
+	}
+
+	return useSIG || hasSCSISupport(sku) // CIG images are not currently tagged for NVMe
+}

pkg/providers/instancetype/suite_test.go

@@ -350,6 +350,9 @@ var _ = Describe("InstanceType Provider", func() {
 		It("should not include confidential SKUs", func() {
 			Expect(instanceTypes).ShouldNot(ContainElement(WithTransform(getName, Equal("Standard_DC8s_v3"))))
 		})
+		It("should not include SKUs without compatible image", func() {
+			Expect(instanceTypes).ShouldNot(ContainElement(WithTransform(getName, Equal("Standard_D2as_v6"))))
+		})
 	})
 	Context("Filtering GPU SKUs ProviderList(AzureLinux)", func() {
 		var instanceTypes corecloudprovider.InstanceTypes
@@ -1074,20 +1077,15 @@ var _ = Describe("InstanceType Provider", func() {
 			ExpectApplied(ctx, env.Client, nodePool, nodeClass)
 			pod := coretest.UnschedulablePod(coretest.PodOptions{})
 			ExpectProvisioned(ctx, env.Client, cluster, cloudProvider, coreProvisioner, pod)
-			ExpectScheduled(ctx, env.Client, pod)
+			node := ExpectScheduled(ctx, env.Client, pod)
 
 			Expect(azureEnv.VirtualMachinesAPI.VirtualMachineCreateOrUpdateBehavior.CalledWithInput.Len()).To(Equal(1))
 			vm := azureEnv.VirtualMachinesAPI.VirtualMachineCreateOrUpdateBehavior.CalledWithInput.Pop().VM
 			Expect(vm.Properties).ToNot(BeNil())
 			Expect(vm.Properties.HardwareProfile).ToNot(BeNil())
 			Expect(utils.IsNvidiaEnabledSKU(string(*vm.Properties.HardwareProfile.VMSize))).To(BeFalse())
 
-			clusterNodes := cluster.Nodes()
-			node := clusterNodes[0]
-			if node.Name() == pod.Spec.NodeName {
-				nodeLabels := node.Labels()
-				Expect(nodeLabels).To(HaveKeyWithValue("karpenter.k8s.azure/sku-gpu-count", "0"))
-			}
+			Expect(node.Labels).To(HaveKeyWithValue("karpenter.azure.com/sku-gpu-count", "0"))
 		})
 
 		It("should schedule GPU pod on GPU capable node", func() {
@@ -1117,23 +1115,31 @@ var _ = Describe("InstanceType Provider", func() {
 			})
 
 			ExpectProvisioned(ctx, env.Client, cluster, cloudProvider, coreProvisioner, pod)
-			ExpectScheduled(ctx, env.Client, pod)
-
-			// Verify that the node has the GPU label set that the pod was scheduled on
-			clusterNodes := cluster.Nodes()
-			Expect(clusterNodes).ToNot(BeEmpty())
-			Expect(len(clusterNodes)).To(Equal(1))
-			node := clusterNodes[0]
-			Expect(node.Node.Status.Allocatable).To(HaveKeyWithValue(v1.ResourceName("nvidia.com/gpu"), resource.MustParse("1")))
+			node := ExpectScheduled(ctx, env.Client, pod)
 
-			if node.Name() == pod.Spec.NodeName {
-				nodeLabels := node.Labels()
+			// the following checks assume Standard_NC16as_T4_v3 (surprisingly the cheapest GPU in the test set), so test the assumption
+			Expect(node.Labels).To(HaveKeyWithValue("node.kubernetes.io/instance-type", "Standard_NC16as_T4_v3"))
 
-				Expect(nodeLabels).To(HaveKeyWithValue("karpenter.k8s.azure/sku-gpu-name", "A100"))
-				Expect(nodeLabels).To(HaveKeyWithValue("karpenter.k8s.azure/sku-gpu-manufacturer", v1alpha2.ManufacturerNvidia))
-				Expect(nodeLabels).To(HaveKeyWithValue("karpenter.k8s.azure/sku-gpu-count", "1"))
+			// Verify GPU related settings in bootstrap (assuming one Standard_NC16as_T4_v3)
+			customData := ExpectDecodedCustomData(azureEnv)
+			Expect(customData).To(SatisfyAll(
+				ContainSubstring("GPU_NODE=true"),
+				ContainSubstring("SGX_NODE=false"),
+				ContainSubstring("MIG_NODE=false"),
+				ContainSubstring("CONFIG_GPU_DRIVER_IF_NEEDED=true"),
+				ContainSubstring("ENABLE_GPU_DEVICE_PLUGIN_IF_NEEDED=false"),
+				ContainSubstring("GPU_DRIVER_TYPE=\"cuda\""),
+				ContainSubstring(fmt.Sprintf("GPU_DRIVER_VERSION=\"%s\"", utils.NvidiaCudaDriverVersion)),
+				ContainSubstring(fmt.Sprintf("GPU_IMAGE_SHA=\"%s\"", utils.AKSGPUCudaVersionSuffix)),
+				ContainSubstring("GPU_NEEDS_FABRIC_MANAGER=\"false\""),
+				ContainSubstring("GPU_INSTANCE_PROFILE=\"\""),
+			))
 
-			}
+			// Verify that the node the pod was scheduled on has GPU resource and labels set
+			Expect(node.Status.Allocatable).To(HaveKeyWithValue(v1.ResourceName("nvidia.com/gpu"), resource.MustParse("1")))
+			Expect(node.Labels).To(HaveKeyWithValue("karpenter.azure.com/sku-gpu-name", "T4"))
+			Expect(node.Labels).To(HaveKeyWithValue("karpenter.azure.com/sku-gpu-manufacturer", v1alpha2.ManufacturerNvidia))
+			Expect(node.Labels).To(HaveKeyWithValue("karpenter.azure.com/sku-gpu-count", "1"))
 		})
 	})
 

pkg/providers/launchtemplate/launchtemplate.go

@@ -144,6 +144,7 @@ func (p *Provider) getStaticParameters(ctx context.Context, instanceType *cloudp
 		Arch:                           arch,
 		GPUNode:                        utils.IsNvidiaEnabledSKU(instanceType.Name),
 		GPUDriverVersion:               utils.GetGPUDriverVersion(instanceType.Name),
+		GPUDriverType:                  utils.GetGPUDriverType(instanceType.Name),
 		GPUImageSHA:                    utils.GetAKSGPUImageSHA(instanceType.Name),
 		TenantID:                       p.tenantID,
 		SubscriptionID:                 p.subscriptionID,

pkg/providers/launchtemplate/parameters/types.go

@@ -28,6 +28,7 @@ type StaticParameters struct {
 	Arch                           string
 	GPUNode                        bool
 	GPUDriverVersion               string
+	GPUDriverType                  string
 	GPUImageSHA                    string
 	TenantID                       string
 	SubscriptionID                 string