Merge pull request #1311 from porter-dev/ym/hf_llm_health

health checks for hf-llm addons and other minor modifications
porter-dev · Jun 18, 2024 · 40f7740 · 40f7740
2 parents ab65310 + 1fa88b2
commit 40f7740
Show file tree

Hide file tree

Showing 4 changed files with 22 additions and 12 deletions.
diff --git a/addons/hf-llm-models/templates/deployment.yaml b/addons/hf-llm-models/templates/deployment.yaml
@@ -2,21 +2,21 @@ apiVersion: apps/v1
 kind: Deployment
 metadata:
   labels:
-    llm-model: {{ .Release.Name }}
+    llm-model: {{ .Release.Name }}-hf-llm
   annotations:
     porter.run/hf-llm-model-version: "{{ .Chart.Version }}"
-  name: {{ .Release.Name }}-workload
+  name: {{ .Release.Name }}-hf-llm
 spec:
   replicas: 1
   strategy:
     type: Recreate
   selector:
     matchLabels:
-      llm-model: {{ .Release.Name }}
+      llm-model: {{ .Release.Name }}-hf-llm
   template:
     metadata:
       labels:
-        llm-model: {{ .Release.Name }}
+        llm-model: {{ .Release.Name }}-hf-llm
     spec:
       tolerations:
         - key: "removable"
@@ -49,14 +49,24 @@ spec:
         - --max-model-len={{ .Values.maxModelLen }}
       {{- end }}
         image: {{ .Values.vllmImage }}
-        imagePullPolicy: Always
+        imagePullPolicy: IfNotPresent
         env:
         - name: HF_TOKEN
           value: {{ .Values.huggingFaceToken }}
         ports:
         - containerPort: 8000
           protocol: TCP
           name: https
+        readinessProbe:
+          failureThreshold: 3
+          httpGet:
+            path: /health
+            port: 8000
+            scheme: HTTP
+          initialDelaySeconds: 15
+          periodSeconds: 15
+          successThreshold: 1
+          timeoutSeconds: 2
         resources:
           requests:
               {{- if .Values.resources.requests.cpu }}
@@ -88,4 +98,4 @@ spec:
       volumes:
       - name: model-volume
         persistentVolumeClaim:
-          claimName: "{{ .Release.Name }}-model-pvc"
+          claimName: "{{ .Release.Name }}-hf-llm"
diff --git a/addons/hf-llm-models/templates/pvc.yaml b/addons/hf-llm-models/templates/pvc.yaml
@@ -1,11 +1,11 @@
 apiVersion: v1
 kind: PersistentVolumeClaim
 metadata:
-  name: {{ .Release.Name }}-model-pvc
+  name: {{ .Release.Name }}-hf-llm
 spec:
   accessModes:
     - ReadWriteMany
-  storageClassName: efs-{{ .Release.Name }}
+  storageClassName: efs-{{ .Release.Name }}-hf-llm
   resources:
     requests:
       storage: 20Gi
diff --git a/addons/hf-llm-models/templates/service.yaml b/addons/hf-llm-models/templates/service.yaml
@@ -2,12 +2,12 @@ apiVersion: v1
 kind: Service
 metadata:
   labels:
-    llm-model: {{ .Release.Name }}
-  name: {{ .Release.Name }}
+    llm-model: {{ .Release.Name }}-hf-llm
+  name: {{ .Release.Name }}-hf-llm
 spec:
   ports:
     - name: https
       port: 8000
       targetPort: https
   selector:
-    llm-model: {{ .Release.Name }}
+    llm-model: {{ .Release.Name }}-hf-llm
diff --git a/addons/hf-llm-models/templates/storageclass.yaml b/addons/hf-llm-models/templates/storageclass.yaml
@@ -1,7 +1,7 @@
 apiVersion: storage.k8s.io/v1
 kind: StorageClass
 metadata:
-  name: efs-{{ .Release.Name }}
+  name: efs-{{ .Release.Name }}-hf-llm
 provisioner: efs.csi.aws.com
 parameters:
   provisioningMode: efs-ap