Dell-Omnia-Clusters-HPC-AI/examples/TensorRT-InferenceServer/README.md

Run Nvidia's TensorRT Inference Server

Clone the repo

git clone https://github.com/NVIDIA/tensorrt-inference-server.git

Download models

cd tensorrt-inference-server/docs/examples/ ./fetch_models.sh

Copy models to shared NFS location

cp -rp model_repository ensemble_model_repository /home/k8sSHARE

Deploy Prometheus and Grafana

Prometheus collects metrics for viewing in Grafana. Install the prometheus-operator for these components. The serviceMonitorSelectorNilUsesHelmValues flag is needed so that Prometheus can find the inference server metrics in the example release deployed below:

helm install --name example-metrics --set prometheus.prometheusSpec.serviceMonitorSelectorNilUsesHelmValues=false stable/prometheus-operator

Setup port-forward to the Grafana service for local access

kubectl port-forward service/example-metrics-grafana 8080:80

Navigate in your browser to localhost:8080 for the Grafana login page. username=admin password=prom-operator

Deploy TensorRT Inference Server

Change to helm chart directory cd ~/tensorrt-inference-server/deploy/single_server/'

Modify values.yaml changing modelRepositoryPath

image:
  imageName: nvcr.io/nvidia/tensorrtserver:20.01-py3
  pullPolicy: IfNotPresent
  #modelRepositoryPath: gs://tensorrt-inference-server-repository/model_repository
  modelRepositoryPath: /data/model_repository
  numGpus: 1
 

Modify templates/deployment.yaml in bold to add the local NFS mount:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: {{ template "tensorrt-inference-server.fullname" . }}
  namespace: {{ .Release.Namespace }}
  labels:


app: {{ template "tensorrt-inference-server.name" . }}
chart: {{ template "tensorrt-inference-server.chart" . }}
release: {{ .Release.Name }}
heritage: {{ .Release.Service }}


spec:
  replicas: {{ .Values.replicaCount }}
  selector:

matchLabels:
  app: {{ template "tensorrt-inference-server.name" . }}
  release: {{ .Release.Name }}


template:

metadata:
  labels:
    app: {{ template "tensorrt-inference-server.name" . }}
    release: {{ .Release.Name }}

spec:
  containers:
    - name: {{ .Chart.Name }}
      image: "{{ .Values.image.imageName }}"
      imagePullPolicy: {{ .Values.image.pullPolicy }}
     <b style='background-color:yellow'> volumeMounts:
        - mountPath: /data/
          name: work-volume</b>
      resources:
        limits:
          nvidia.com/gpu: {{ .Values.image.numGpus }}

      args: ["trtserver", "--model-store={{ .Values.image.modelRepositoryPath }}"]

      ports:
        - containerPort: 8000
          name: http
        - containerPort: 8001
          name: grpc
        - containerPort: 8002
          name: metrics
      livenessProbe:
        httpGet:
          path: /api/health/live
          port: http
      readinessProbe:
        initialDelaySeconds: 5
        periodSeconds: 5
        httpGet:
          path: /api/health/ready
          port: http

      securityContext:
        runAsUser: 1000
        fsGroup: 1000


   volumes:

  - name: work-volume
    hostPath:
      # directory locally mounted on host
      path: /home/k8sSHARE
      type: Directory



   


Deploy the inference server using the default configuration with:

cd ~/tensorrt-inference-server/deploy/single_server/
$ helm install --name example .