Change nvidia device plugin to tegra strategy

Author: nwthomas

helm/nvidia-device-plugin/templates/daemonset.yaml

@@ -37,6 +37,9 @@ spec:
           value: "true"
         - name: FAIL_ON_INIT_ERROR
           value: "false"
+        {{- if .Values.env }}
+        {{- toYaml .Values.env | nindent 8 }}
+        {{- else }}
         - name: DEVICE_LIST_STRATEGY
           value: envvar
         - name: DEVICE_ID_STRATEGY
@@ -45,6 +48,7 @@ spec:
           value: all
         - name: NVIDIA_DRIVER_CAPABILITIES
           value: "compute,utility"
+        {{- end }}
       volumes:
         {{- toYaml .Values.volumes | nindent 8 }}
       hostNetwork: true

helm/nvidia-device-plugin/values.yaml

@@ -20,6 +20,26 @@ resources:
 securityContext:
   privileged: true
 
+# Tolerations for GPU nodes
+tolerations:
+  - key: gpu
+    operator: Equal
+    value: "true"
+    effect: NoExecute
+
+# Environment variables for nvidia-device-plugin
+env:
+  - name: DEVICE_DISCOVERY_STRATEGY
+    value: "tegra"  # Use tegra discovery strategy
+  - name: DEVICE_LIST_STRATEGY
+    value: "envvar"  # Use envvar for device list strategy
+  - name: DEVICE_ID_STRATEGY
+    value: "uuid"
+  - name: NVIDIA_VISIBLE_DEVICES
+    value: "all"
+  - name: NVIDIA_DRIVER_CAPABILITIES
+    value: "compute,utility"
+
 # Mount the NVIDIA libraries and device files
 volumeMounts:
   - name: device-plugin

helm/vllm/templates/deployment.yaml

@@ -15,6 +15,9 @@ spec:
       labels:
         {{- include "vllm.selectorLabels" . | nindent 8 }}
     spec:
+      {{- if .Values.runtimeClassName }}
+      runtimeClassName: {{ .Values.runtimeClassName }}
+      {{- end }}
       {{- with .Values.nodeSelector }}
       nodeSelector:
         {{- toYaml . | nindent 8 }}
@@ -46,6 +49,9 @@ spec:
         - "{{ .Values.service.targetPort }}"
         - "--tensor-parallel-size"
         - "{{ .Values.model.tensorParallelSize }}"
+        - "--device"
+        - "cuda"
+        - "--trust-remote-code"
         {{- if .Values.persistence.enabled }}
         - "--download-dir"
         - "{{ .Values.model.downloadDir }}"
@@ -54,11 +60,14 @@ spec:
         {{- toYaml .Values.env | nindent 8 }}
         resources:
           {{- toYaml .Values.resources | nindent 10 }}
-        {{- if .Values.persistence.enabled }}
         volumeMounts:
+        {{- if .Values.persistence.enabled }}
         - name: model-storage
           mountPath: {{ .Values.model.downloadDir }}
         {{- end }}
+        {{- if .Values.volumeMounts }}
+        {{- toYaml .Values.volumeMounts | nindent 8 }}
+        {{- end }}
         readinessProbe:
           httpGet:
             path: /health
@@ -75,9 +84,12 @@ spec:
           periodSeconds: 30
           timeoutSeconds: 10
           failureThreshold: 3
-      {{- if .Values.persistence.enabled }}
       volumes:
+      {{- if .Values.persistence.enabled }}
       - name: model-storage
         persistentVolumeClaim:
           claimName: {{ include "vllm.fullname" . }}-models
+      {{- end }}
+      {{- if .Values.volumes }}
+      {{- toYaml .Values.volumes | nindent 6 }}
       {{- end }}

helm/vllm/values.yaml

@@ -1,6 +1,6 @@
 image:
   repository: vllm/vllm-openai
-  tag: latest
+  tag: v0.6.3
   pullPolicy: Always
 
 replicaCount: 1
@@ -57,11 +57,40 @@ persistence:
 # Environment variables
 env:
   - name: NVIDIA_VISIBLE_DEVICES
-    value: "all"
+    value: "0"
+  - name: CUDA_VISIBLE_DEVICES
+    value: "0"
   - name: NVIDIA_DRIVER_CAPABILITIES
     value: "compute,utility"
+  - name: VLLM_LOGGING_LEVEL
+    value: "DEBUG"
+  - name: CUDA_HOME
+    value: "/usr/local/cuda"
+  - name: LD_LIBRARY_PATH
+    value: "/usr/lib/x86_64-linux-gnu:/usr/local/cuda/lib64"
 
 # Security context
 securityContext:
   runAsNonRoot: false
-  runAsUser: 0
+  runAsUser: 0
+
+# Runtime class for NVIDIA GPU support (commented out for k3s compatibility)
+# runtimeClassName: nvidia
+
+# Volume mounts for NVIDIA libraries
+volumeMounts:
+  - name: nvidia-libs
+    mountPath: /usr/lib/x86_64-linux-gnu
+    readOnly: true
+  - name: nvidia-driver
+    mountPath: /usr/local/nvidia
+    readOnly: true
+
+# Volumes for NVIDIA libraries
+volumes:
+  - name: nvidia-libs
+    hostPath:
+      path: /usr/lib/x86_64-linux-gnu
+  - name: nvidia-driver
+    hostPath:
+      path: /usr/local/nvidia