fix: change the annotation key for enabling LB type service

api/v1alpha1/workspace_labels.go

@@ -7,15 +7,12 @@ const (
 	// KAITOPrefix Kubernetes Data Mining prefix.
 	KAITOPrefix = "kaito.sh/"
 
-	// AnnotationServiceType determines whether kaito creates ClusterIP or LoadBalancer type service.
-	AnnotationServiceType = KAITOPrefix + "service-type"
+	// AnnotationEnableLB determines whether kaito creates LoadBalancer type service.for testing.
+	AnnotationEnableLB = KAITOPrefix + "enablelb"
 
 	// LabelWorkspaceName is the label for workspace name.
 	LabelWorkspaceName = KAITOPrefix + "workspace"
 
 	// LabelWorkspaceName is the label for workspace namespace.
 	LabelWorkspaceNamespace = KAITOPrefix + "workspacenamespace"
-
-	ServiceTypeClusterIP    = "cluster-ip"
-	ServiceTypeLoadBalancer = "load-balancer"
 )

examples/README.md

@@ -0,0 +1,2 @@
+# Notes:
+For **testing** purposes, users can add the `kaito.sh/enablelb: "True"` annotation to the workspace custom resource. As a result, a `loadbalancer` type service will be created for the inference service with a public IP being assigned. However, this is **NOT** recommended for production use. An [ingress controller](https://learn.microsoft.com/en-us/azure/aks/ingress-basic?tabs=azure-cli) is recommended to expose the service to public.

examples/kaito_workspace_falcon_7b.yaml

@@ -2,6 +2,8 @@ apiVersion: kaito.sh/v1alpha1
 kind: Workspace
 metadata:
   name: workspace-falcon-7b
+  annotations:
+    kaito.sh/enablelb: "True"
 resource:
   instanceType: "Standard_NC12s_v3"
   labelSelector:

examples/kaito_workspace_llama2_13b-chat.yaml

@@ -6,7 +6,7 @@ resource:
   instanceType: "Standard_NC12s_v3"
   labelSelector:
     matchLabels:
-      apps: llama-2-13b
+      apps: llama-2-13b-chat
 inference:
   preset:
     name: "llama-2-13b-chat"

examples/kaito_workspace_llama2_13b.yaml

@@ -0,0 +1,15 @@
+apiVersion: kaito.sh/v1alpha1
+kind: Workspace
+metadata:
+  name: workspace-llama-2-13b
+resource:
+  instanceType: "Standard_NC12s_v3"
+  labelSelector:
+    matchLabels:
+      apps: llama-2-13b
+inference:
+  preset:
+    name: "llama-2-13b"
+    accessMode: private
+    presetOptions:
+      image: <YOUR IMAGE URL>

examples/kaito_workspace_llama2_70b-chat.yaml

@@ -7,7 +7,7 @@ resource:
   count: 2
   labelSelector:
     matchLabels:
-      apps: llama-2-70b
+      apps: llama-2-70b-chat
 inference:
   preset:
     name: "llama-2-70b-chat"

examples/kaito_workspace_llama2_70b.yaml

@@ -0,0 +1,16 @@
+apiVersion: kaito.sh/v1alpha1
+kind: Workspace
+metadata:
+  name: workspace-llama-2-70b
+resource:
+  instanceType: "Standard_NC96ads_A100_v4"
+  count: 2
+  labelSelector:
+    matchLabels:
+      apps: llama-2-70b
+inference:
+  preset:
+    name: "llama-2-70b"
+    accessMode: private
+    presetOptions:
+      image: <YOUR IMAGE URL>

examples/kaito_workspace_llama2_7b-chat.yaml

@@ -6,7 +6,7 @@ resource:
   instanceType: "Standard_NC12s_v3"
   labelSelector:
     matchLabels:
-      apps: llama-2-7b
+      apps: llama-2-7b-chat
 inference:
   preset:
     name: "llama-2-7b-chat"

examples/kaito_workspace_llama2_7b.yaml

@@ -0,0 +1,15 @@
+apiVersion: kaito.sh/v1alpha1
+kind: Workspace
+metadata:
+  name: workspace-llama-2-7b
+resource:
+  instanceType: "Standard_NC12s_v3"
+  labelSelector:
+    matchLabels:
+      apps: llama-2-7b
+inference:
+  preset:
+    name: "llama-2-7b"
+    accessMode: private
+    presetOptions:
+      image: <YOUR IMAGE URL>

pkg/controllers/workspace_controller.go

@@ -384,8 +384,8 @@ func (c *WorkspaceReconciler) ensureService(ctx context.Context, wObj *kaitov1al
 	wAnnotation := wObj.GetAnnotations()
 
 	if len(wAnnotation) != 0 {
-		_, found := lo.FindKey(wAnnotation, kaitov1alpha1.ServiceTypeLoadBalancer)
-		if found {
+		val, found := wAnnotation[kaitov1alpha1.AnnotationEnableLB]
+		if found && val == "True" {
 			serviceType = corev1.ServiceTypeLoadBalancer
 		}
 	}