fix: GPU bootstrap, refresh driver versions and list of supported GPU VM SKUs

.pre-commit-config.yaml

@@ -12,7 +12,7 @@ repos:
   hooks:
   - id: shellcheck
 - repo: https://github.com/crate-ci/typos
-  rev: v1.26.0
+  rev: v1.28.1
   hooks:
   - id: typos
     args: [--write-changes, --force-exclude, --exclude, go.mod]

go.mod

@@ -40,6 +40,7 @@ require (
 	go.uber.org/multierr v1.11.0
 	go.uber.org/zap v1.27.0
 	golang.org/x/sync v0.8.0
+	gopkg.in/yaml.v2 v2.4.0
 	k8s.io/api v0.30.3
 	k8s.io/apiextensions-apiserver v0.30.3
 	k8s.io/apimachinery v0.30.3
@@ -158,7 +159,6 @@ 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 v0.30.3 // indirect
 	k8s.io/component-base v0.30.3 // indirect

pkg/controllers/nodeclass/hash/controller.go

@@ -100,7 +100,7 @@ func (c *Controller) updateNodeClaimHash(ctx context.Context, nodeClass *v1alpha
 				v1alpha2.AnnotationAKSNodeClassHashVersion: v1alpha2.AKSNodeClassHashVersion,
 			})
 
-			// Any NodeClaim that is already drifted will remain drifted if the karpenter.k8s.aws/nodepool-hash-version doesn't match
+			// Any NodeClaim that is already drifted will remain drifted if the karpenter.azure.com/nodepool-hash-version doesn't match
 			// Since the hashing mechanism has changed we will not be able to determine if the drifted status of the NodeClaim has changed
 			if nc.StatusConditions().Get(karpv1.ConditionTypeDrifted) == nil {
 				nc.Annotations = lo.Assign(nc.Annotations, map[string]string{

pkg/providers/imagefamily/azlinux.go

@@ -88,9 +88,9 @@ func (u AzureLinux) ScriptlessCustomData(kubeletConfig *bootstrap.KubeletConfigu
 			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

@@ -204,6 +204,7 @@ 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]
@@ -467,6 +468,7 @@ 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

@@ -52,6 +52,7 @@ 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

@@ -89,6 +89,7 @@ func (u Ubuntu2204) ScriptlessCustomData(kubeletConfig *bootstrap.KubeletConfigu
 			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/suite_test.go

@@ -604,12 +604,10 @@ var _ = Describe("InstanceType Provider", func() {
 				nodes := &v1.NodeList{}
 				Expect(env.Client.List(ctx, nodes)).To(Succeed())
 				for _, node := range nodes.Items {
-					Expect(node.Labels["karpenter.k8s.azure/zone"]).ToNot(Equal(fmt.Sprintf("%s-1", fake.Region)))
+					Expect(node.Labels["karpenter.kubernetes.azure/zone"]).ToNot(Equal(fmt.Sprintf("%s-1", fake.Region)))
 					Expect(node.Labels["node.kubernetes.io/instance-type"]).To(Equal("Standard_D2_v2"))
-
 				}
 			}
-
 		})
 
 		DescribeTable("Should not return unavailable offerings", func(azEnv *test.Environment) {
@@ -655,7 +653,7 @@ var _ = Describe("InstanceType Provider", func() {
 			}}}
 			ExpectProvisioned(ctx, env.Client, cluster, cloudProvider, coreProvisioner, pod)
 			node := ExpectScheduled(ctx, env.Client, pod)
-			Expect(node.Labels["karpenter.k8s.azure/zone"]).ToNot(Equal(fmt.Sprintf("%s-1", fake.Region)))
+			Expect(node.Labels["karpenter.kubernetes.azure/zone"]).ToNot(Equal(fmt.Sprintf("%s-1", fake.Region)))
 			Expect(node.Labels["node.kubernetes.io/instance-type"]).To(Equal("Standard_D2_v2"))
 		})
 		It("should launch smaller instances than optimal if larger instance launch results in Insufficient Capacity Error", func() {
@@ -1011,20 +1009,15 @@ 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() {
@@ -1054,23 +1047,31 @@ 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

@@ -152,6 +152,7 @@ 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

@@ -29,6 +29,7 @@ type StaticParameters struct {
 	Arch                           string
 	GPUNode                        bool
 	GPUDriverVersion               string
+	GPUDriverType                  string
 	GPUImageSHA                    string
 	TenantID                       string
 	SubscriptionID                 string

pkg/utils/gpu.go

@@ -17,124 +17,81 @@ limitations under the License.
 package utils
 
 import (
+	_ "embed"
 	"strings"
+
+	"gopkg.in/yaml.v2"
 )
 
 // TODO: Get these from agentbaker
 const (
-	Nvidia470CudaDriverVersion = "cuda-470.82.01"
-	Nvidia550CudaDriverVersion = "cuda-550.54.15"
-	Nvidia535GridDriverVersion = "grid-535.161.08"
-
-	// These SHAs will change once we update aks-gpu images in aks-gpu repository. We do that fairly rarely at this time.
-	// So for now these will be kept here like this and periodically bump them
-	AKSGPUGridSHA = "sha-d1f0ca"
-	AKSGPUCudaSHA = "sha-2d4c96"
+	Nvidia470CudaDriverVersion = "470.82.01"
+
+	// https://github.com/Azure/AgentBaker/blob/ddf36a24eafd02ce0589657ff2dc799125f4ad37/parts/linux/cloud-init/artifacts/components.json#L562
+	NvidiaCudaDriverVersion = "550.90.12"
+	AKSGPUCudaVersionSuffix = "20241021235610"
+
+	NvidiaGridDriverVersion = "535.161.08"
+	AKSGPUGridVersionSuffix = "20241021235607"
 )
 
-func GetAKSGPUImageSHA(size string) string {
-	if UseGridDrivers(size) {
-		return AKSGPUGridSHA
-	}
-	return AKSGPUCudaSHA
+type NvidiaSKUConfig struct {
+	NvidiaEnabledSKUFamilies        map[string][]string `yaml:"nvidiaEnabledSKUs"`
+	MarinerNvidiaEnabledSKUFamilies map[string][]string `yaml:"marinerNvidiaEnabledSKUs"`
 }
 
 var (
-	/* If a new GPU sku becomes available, add a key to this map, but only if you have a confirmation
-	   that we have an agreement with NVIDIA for this specific gpu.
-	*/
-	NvidiaEnabledSKUs = map[string]bool{
-		// M60
-		"standard_nv6":      true,
-		"standard_nv12":     true,
-		"standard_nv12s_v3": true,
-		"standard_nv24":     true,
-		"standard_nv24s_v3": true,
-		"standard_nv24r":    true,
-		"standard_nv48s_v3": true,
-		// P40
-		"standard_nd6s":   true,
-		"standard_nd12s":  true,
-		"standard_nd24s":  true,
-		"standard_nd24rs": true,
-		// P100
-		"standard_nc6s_v2":   true,
-		"standard_nc12s_v2":  true,
-		"standard_nc24s_v2":  true,
-		"standard_nc24rs_v2": true,
-		// V100
-		"standard_nc6s_v3":   true,
-		"standard_nc12s_v3":  true,
-		"standard_nc24s_v3":  true,
-		"standard_nc24rs_v3": true,
-		"standard_nd40s_v3":  true,
-		"standard_nd40rs_v2": true,
-		// T4
-		"standard_nc4as_t4_v3":  true,
-		"standard_nc8as_t4_v3":  true,
-		"standard_nc16as_t4_v3": true,
-		"standard_nc64as_t4_v3": true,
-		// A100 40GB
-		"standard_nd96asr_v4":       true,
-		"standard_nd112asr_a100_v4": true,
-		"standard_nd120asr_a100_v4": true,
-		// A100 80GB
-		"standard_nd96amsr_a100_v4":  true,
-		"standard_nd112amsr_a100_v4": true,
-		"standard_nd120amsr_a100_v4": true,
-		// A100 PCIE 80GB
-		"standard_nc24ads_a100_v4": true,
-		"standard_nc48ads_a100_v4": true,
-		"standard_nc96ads_a100_v4": true,
-		"standard_ncads_a100_v4":   true,
-		// A10
-		"standard_nc8ads_a10_v4":  true,
-		"standard_nc16ads_a10_v4": true,
-		"standard_nc32ads_a10_v4": true,
-		// A10, GRID only
-		"standard_nv6ads_a10_v5":   true,
-		"standard_nv12ads_a10_v5":  true,
-		"standard_nv18ads_a10_v5":  true,
-		"standard_nv36ads_a10_v5":  true,
-		"standard_nv36adms_a10_v5": true,
-		"standard_nv72ads_a10_v5":  true,
-		// A100
-		"standard_nd96ams_v4":      true,
-		"standard_nd96ams_a100_v4": true,
+	nvidiaEnabledSKUs        = make(map[string]bool)
+	marinerNvidiaEnabledSKUs = make(map[string]bool)
+)
+
+//go:embed supported-gpus.yaml
+var configFile []byte
+
+func init() {
+	readNvidiaSKUConfig()
+}
+
+func readNvidiaSKUConfig() {
+	var nvidiaSKUConfig NvidiaSKUConfig
+
+	err := yaml.Unmarshal(configFile, &nvidiaSKUConfig)
+	if err != nil {
+		panic(err)
 	}
+	for _, skus := range nvidiaSKUConfig.NvidiaEnabledSKUFamilies {
+		for _, sku := range skus {
+			nvidiaEnabledSKUs[sku] = true
+		}
+	}
+	for _, skus := range nvidiaSKUConfig.MarinerNvidiaEnabledSKUFamilies {
+		for _, sku := range skus {
+			marinerNvidiaEnabledSKUs[sku] = true
+		}
+	}
+}
 
-	// List of GPU SKUs currently enabled and validated for Mariner. Will expand the support
-	// to cover other SKUs available in Azure
-	MarinerNvidiaEnabledSKUs = map[string]bool{
-		// V100
-		"standard_nc6s_v3":   true,
-		"standard_nc12s_v3":  true,
-		"standard_nc24s_v3":  true,
-		"standard_nc24rs_v3": true,
-		"standard_nd40s_v3":  true,
-		"standard_nd40rs_v2": true,
-		// T4
-		"standard_nc4as_t4_v3":  true,
-		"standard_nc8as_t4_v3":  true,
-		"standard_nc16as_t4_v3": true,
-		"standard_nc64as_t4_v3": true,
+func GetAKSGPUImageSHA(size string) string {
+	if UseGridDrivers(size) {
+		return AKSGPUGridVersionSuffix
 	}
-)
+	return AKSGPUCudaVersionSuffix
+}
 
 // IsNvidiaEnabledSKU determines if an VM SKU has nvidia driver support
 func IsNvidiaEnabledSKU(vmSize string) bool {
 	// Trim the optional _Promo suffix.
 	vmSize = strings.ToLower(vmSize)
 	vmSize = strings.TrimSuffix(vmSize, "_promo")
-	return NvidiaEnabledSKUs[vmSize]
+	return nvidiaEnabledSKUs[vmSize]
 }
 
 // IsNvidiaEnabledSKU determines if an VM SKU has nvidia driver support
 func IsMarinerEnabledGPUSKU(vmSize string) bool {
 	// Trim the optional _Promo suffix.
 	vmSize = strings.ToLower(vmSize)
 	vmSize = strings.TrimSuffix(vmSize, "_promo")
-	return MarinerNvidiaEnabledSKUs[vmSize]
+	return marinerNvidiaEnabledSKUs[vmSize]
 }
 
 // NV series GPUs target graphics workloads vs NC which targets compute.
@@ -143,12 +100,20 @@ func IsMarinerEnabledGPUSKU(vmSize string) bool {
 // NVv3 is untested on AKS, NVv4 is AMD so n/a, and NVv2 no longer seems to exist (?).
 func GetGPUDriverVersion(size string) string {
 	if UseGridDrivers(size) {
-		return Nvidia535GridDriverVersion
+		return NvidiaGridDriverVersion
 	}
 	if isStandardNCv1(size) {
 		return Nvidia470CudaDriverVersion
 	}
-	return Nvidia550CudaDriverVersion
+	return NvidiaCudaDriverVersion
+}
+
+// GetGPUDriverType returns the type of GPU driver for given VM SKU ("grid" or "cuda")
+func GetGPUDriverType(size string) string {
+	if UseGridDrivers(size) {
+		return "grid"
+	}
+	return "cuda"
 }
 
 func isStandardNCv1(size string) bool {