tutorials: Lab 20, init and sidecar container GPU resource accounting

sidebars-tutorials.js

@@ -106,6 +106,11 @@ module.exports = {
           id: "labs/hami-ascend-vnpu-slicing",
           customProps: { level: "Intermediate", duration: "about 60 minutes" },
         },
+        {
+          type: "doc",
+          id: "labs/init-sidecar-container-accounting",
+          customProps: { level: "Intermediate", duration: "about 45 minutes" },
+        },
       ],
     },
   ],

tutorials/labs/examples/20-init-sidecar-accounting/01-regular-init-container.yaml

@@ -0,0 +1,30 @@
+apiVersion: v1
+kind: Pod
+metadata:
+  name: regular-init-pod
+  labels:
+    app: regular-init-demo
+spec:
+  restartPolicy: Never
+  initContainers:
+    - name: model-prep
+      image: busybox:1.36
+      command: ["sh", "-c", "echo 'Preparing model weights on GPU...' && sleep 15 && echo 'Done'"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true" # Enables lightweight simulation in fake GPU / nvml-mock environments
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 5000 # 5000 MiB allocated during init (peak allocation)
+  containers:
+    - name: inference-app
+      image: busybox:1.36
+      command: ["sh", "-c", "echo 'Inference app running...' && sleep infinity"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 2000 # 2000 MiB allocated to the main application (steady state)

tutorials/labs/examples/20-init-sidecar-accounting/02-native-sidecar-container.yaml

@@ -0,0 +1,31 @@
+apiVersion: v1
+kind: Pod
+metadata:
+  name: native-sidecar-pod
+  labels:
+    app: native-sidecar-demo
+spec:
+  restartPolicy: Never
+  initContainers:
+    - name: proxy-sidecar
+      image: busybox:1.36
+      restartPolicy: Always # KEP-753 Native Sidecar Container: runs for the full Pod lifetime
+      command: ["sh", "-c", "echo 'Inference proxy sidecar running...' && sleep infinity"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 2000 # 2000 MiB permanently reserved for the sidecar
+  containers:
+    - name: inference-app
+      image: busybox:1.36
+      command: ["sh", "-c", "echo 'Main inference app running...' && sleep infinity"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 4000 # 4000 MiB reserved for the app (Total demand: 6000 MiB)

tutorials/labs/examples/20-init-sidecar-accounting/03-mixed-workload-ordering.yaml

@@ -0,0 +1,43 @@
+apiVersion: v1
+kind: Pod
+metadata:
+  name: mixed-workload-pod
+  labels:
+    app: mixed-workload-demo
+spec:
+  restartPolicy: Never
+  initContainers:
+    # Sidecar declared before the regular init container
+    - name: cache-sidecar
+      image: busybox:1.36
+      restartPolicy: Always # Native sidecar: stays running
+      command: ["sh", "-c", "echo 'Cache sidecar active...' && sleep infinity"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 2000 # 2000 MiB
+    # Regular init container: runs to completion
+    - name: weights-loader
+      image: busybox:1.36
+      command: ["sh", "-c", "echo 'Loading model weights...' && sleep 15 && echo 'Done'"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 5000 # 5000 MiB (Peak with preceding sidecar: 7000 MiB)
+  containers:
+    - name: serving-engine
+      image: busybox:1.36
+      command: ["sh", "-c", "echo 'Serving engine active...' && sleep infinity"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 4000 # 4000 MiB (Steady-state with sidecar: 6000 MiB)

tutorials/labs/examples/20-init-sidecar-accounting/04-oversubscription-prevention.yaml

@@ -0,0 +1,23 @@
+apiVersion: v1
+kind: Pod
+metadata:
+  name: competing-workload-pod
+  labels:
+    app: competing-workload
+  annotations:
+    # Pin this workload to the same GPU card hosting native-sidecar-pod
+    # Replace TARGET_GPU_UUID with the allocated GPU UUID from Step 3
+    nvidia.com/use-gpuuuid: "TARGET_GPU_UUID"
+spec:
+  restartPolicy: Never
+  containers:
+    - name: worker
+      image: busybox:1.36
+      command: ["sh", "-c", "echo 'Competing workload running...' && sleep infinity"]
+      env:
+        - name: CUDA_DISABLE_CONTROL
+          value: "true"
+      resources:
+        limits:
+          nvidia.com/gpu: 1
+          nvidia.com/gpumem: 36000 # Requests 36000 MiB: exceeds the 34960 MiB remaining on a 40960 MiB A100 GPU