Add base test for vLLM and its metrics

Signed-off-by: Luca Giorgi <lgiorgi@redhat.com>

ods_ci/tests/Resources/CLI/ModelServing/llm.resource

@@ -775,8 +775,9 @@ Get KServe Default Deployment Mode From DSC
     RETURN    ${mode}
 
 Start Port-forwarding
-    [Arguments]    ${namespace}    ${pod_name}    ${process_alias}=llm-query-process
-    ${process}=    Start Process    oc -n ${namespace} port-forward pod/${pod_name} 8033:8033
+    [Arguments]    ${namespace}    ${pod_name}    ${process_alias}=llm-query-process    ${local_port}=8033
+    ...    ${remote_port}=8033
+    ${process}=    Start Process    oc -n ${namespace} port-forward pod/${pod_name} ${local_port}:${remote_port}
     ...    alias=${process_alias}    stderr=STDOUT    shell=yes
     Process Should Be Running    ${process}
     sleep  7s

ods_ci/tests/Resources/Files/llm/vllm/download_model.yaml

@@ -0,0 +1,74 @@
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: vllm-gpt2
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: vlmm-gpt2-claim
+  namespace: vllm-gpt2
+spec:
+  accessModes:
+    - ReadWriteOnce
+  volumeMode: Filesystem
+  resources:
+    requests:
+      storage: 10Gi
+---
+apiVersion: v1
+kind: Pod
+metadata:
+  name: setup-gpt2-binary
+  namespace: vllm-gpt2
+  labels:
+    gpt-download-pod: 'true'
+spec:
+  volumes:
+    - name: model-volume
+      persistentVolumeClaim:
+        claimName: vlmm-gpt2-claim
+  restartPolicy: Never
+  initContainers:
+    - name: fix-volume-permissions
+      image: quay.io/quay/busybox:latest
+      imagePullPolicy: IfNotPresent
+      securityContext:
+        allowPrivilegeEscalation: true
+      resources:
+        requests:
+          memory: "64Mi"
+          cpu: "250m"
+          nvidia.com/gpu: "1"
+        limits:
+          memory: "128Mi"
+          cpu: "500m"
+          nvidia.com/gpu: "1"
+      command: ["sh"]
+      args: ["-c", "chown -R 1001:1001 /mnt/models"]
+      volumeMounts: 
+        - mountPath: "/mnt/models/"
+          name: model-volume
+  containers:
+    - name: download-model
+      image: registry.access.redhat.com/ubi9/python-311:latest
+      imagePullPolicy: IfNotPresent
+      securityContext:
+        allowPrivilegeEscalation: true
+      resources:
+        requests:
+          memory: "1Gi"
+          cpu: "1"
+          nvidia.com/gpu: "1"
+        limits:
+          memory: "1Gi"
+          cpu: "1"
+          nvidia.com/gpu: "1"
+      command: ["sh"]
+      args: [ "-c", "pip install --upgrade pip && pip install --upgrade huggingface_hub && python3 -c 'from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"gpt2\", local_dir=\"/mnt/models/gpt2\", local_dir_use_symlinks=False)'"]
+      volumeMounts:
+        - mountPath: "/mnt/models/"
+          name: model-volume
+      env:
+        - name: TRANSFORMERS_CACHE
+          value: /tmp

ods_ci/tests/Resources/Files/llm/vllm/query.json

@@ -0,0 +1,13 @@
+{
+    "model": "gpt2",
+    "messages": [
+      {
+        "role": "system",
+        "content": "You are a poetic assistant, skilled in explaining complex programming concepts with creative flair."
+      },
+      {
+        "role": "user",
+        "content": "Compose a poem that explains the concept of recursion in programming."
+      }
+    ]
+}

ods_ci/tests/Resources/Files/llm/vllm/vllm-gpt2_inferenceservice.yaml

@@ -0,0 +1,14 @@
+apiVersion: serving.kserve.io/v1beta1
+kind: InferenceService
+metadata: 
+  name: vllm-gpt2-openai
+  namespace: vllm-gpt2
+  labels:
+    modelmesh-enabled: "true"
+spec: 
+  predictor: 
+    model: 
+      runtime: kserve-vllm
+      modelFormat: 
+        name: vLLM
+      storageUri: pvc://vlmm-gpt2-claim/

ods_ci/tests/Resources/Files/llm/vllm/vllm_servingruntime.yaml

@@ -0,0 +1,79 @@
+apiVersion: serving.kserve.io/v1alpha1
+kind: ServingRuntime
+metadata: 
+  name: kserve-vllm
+  namespace: vllm-gpt2
+spec: 
+  annotations: 
+    sidecar.istio.io/inject: "true"
+    sidecar.istio.io/rewriteAppHTTPProbers: "true"
+    serving.knative.openshift.io/enablePassthrough: "true"
+    opendatahub.io/dashboard: "true"
+    openshift.io/display-name: "vLLLM Openai entry point"
+    serving.kserve.io/enable-prometheus-scraping: "false"
+    serving.kserve.io/enable-metric-aggregation: "true"
+    prometheus.io/port: '8080'
+    prometheus.io/path: "/metrics/"
+  multiModel: false
+  supportedModelFormats: 
+    - name: vLLM
+      autoSelect: true
+  containers: 
+    - name: kserve-container
+      #image: kserve/vllmserver:latest
+      image: quay.io/wxpe/tgis-vllm:release.74803b6
+      startupProbe: 
+        httpGet: 
+          port: 8080
+          path: /health
+        # Allow 12 minutes to start
+        failureThreshold: 24
+        periodSeconds: 30
+      readinessProbe: 
+        httpGet: 
+          port: 8080
+          path: /health
+        periodSeconds: 30
+        timeoutSeconds: 5
+      livenessProbe: 
+        httpGet: 
+          port: 8080
+          path: /health
+        periodSeconds: 100
+        timeoutSeconds: 8
+      terminationMessagePolicy: "FallbackToLogsOnError"
+      terminationGracePeriodSeconds: 120
+      args: 
+        - --port
+        - "8080"
+        - --model
+        - /mnt/models/gpt2
+        - --served-model-name
+        - "gpt2"
+      command: 
+        - python3
+        - -m
+        - vllm.entrypoints.openai.api_server
+      env: 
+        - name: STORAGE_URI
+          value: pvc://vlmm-gpt2-claim/
+        - name: HF_HUB_CACHE
+          value: /tmp
+        - name: TRANSFORMERS_CACHE
+          value: $(HF_HUB_CACHE)
+        - name: NUM_GPUS
+          value: "1"
+        - name: CUDA_VISIBLE_DEVICES
+          value: "0"   
+      ports: 
+        - containerPort: 8080
+          protocol: TCP
+      resources: 
+        limits: 
+          cpu: "4"
+          memory: 8Gi
+          nvidia.com/gpu: "1"
+        requests: 
+          cpu: "1"
+          memory: 4Gi
+          nvidia.com/gpu: "1"

ods_ci/tests/Tests/400__ods_dashboard/420__model_serving/LLMs/vllm/426__model_serving_vllm_metrics.robot

@@ -0,0 +1,119 @@
+*** Settings ***
+Documentation     Basic vLLM deploy test to validate metrics being correctly exposed in OpenShift
+Resource          ../../../../../Resources/Page/ODH/ODHDashboard/ODHModelServing.resource
+Resource          ../../../../../Resources/OCP.resource
+Resource          ../../../../../Resources/Page/Operators/ISVs.resource
+Resource          ../../../../../Resources/Page/ODH/ODHDashboard/ODHDashboardAPI.resource
+Resource          ../../../../../Resources/Page/ODH/ODHDashboard/ODHDataScienceProject/ModelServer.resource
+Resource          ../../../../../Resources/Page/ODH/ODHDashboard/ODHDataScienceProject/DataConnections.resource
+Resource          ../../../../../Resources/CLI/ModelServing/llm.resource
+Resource          ../../../../../Resources/Page/ODH/ODHDashboard/ODHDataScienceProject/Permissions.resource
+Library           OpenShiftLibrary
+Suite Setup       Suite Setup
+Suite Teardown    Suite Teardown
+Test Tags         KServe
+
+
+*** Variables ***
+${VLLM_RESOURCES_DIRPATH}=    ods_ci/tests/Resources/Files/llm/vllm
+${DL_POD_FILEPATH}=           ${VLLM_RESOURCES_DIRPATH}/download_model.yaml
+${SR_FILEPATH}=               ${VLLM_RESOURCES_DIRPATH}/vllm_servingruntime.yaml
+${IS_FILEPATH}=               ${VLLM_RESOURCES_DIRPATH}/vllm-gpt2_inferenceservice.yaml
+${INFERENCE_INPUT}=           @${VLLM_RESOURCES_DIRPATH}/query.json
+${INFERENCE_URL}=             http://localhost:8080/v1/chat/completions
+${METRICS_URL}=               http://localhost:8080/metrics/
+${TEST_NS}=                   vllm-gpt2
+@{SEARCH_METRICS}=            vllm:cache_config_info
+...                           vllm:num_requests_running
+...                           vllm:num_requests_swapped
+...                           vllm:num_requests_waiting
+...                           vllm:gpu_cache_usage_perc
+...                           vllm:cpu_cache_usage_perc
+...                           vllm:prompt_tokens_total
+...                           vllm:generation_tokens_total
+...                           vllm:time_to_first_token_seconds_bucket
+...                           vllm:time_to_first_token_seconds_count
+...                           vllm:time_to_first_token_seconds_sum
+...                           vllm:time_per_output_token_seconds_bucket
+...                           vllm:time_per_output_token_seconds_count
+...                           vllm:time_per_output_token_seconds_sum
+...                           vllm:e2e_request_latency_seconds_bucket
+...                           vllm:e2e_request_latency_seconds_count
+...                           vllm:e2e_request_latency_seconds_sum
+...                           vllm:avg_prompt_throughput_toks_per_s
+...                           vllm:avg_generation_throughput_toks_per_s
+
+
+*** Test Cases ***
+Verify User Can Deploy A Model With Vllm Via CLI
+    [Documentation]    Deploy a model (gpt2) using the vllm runtime and confirm that it's running
+    [Tags]    Tier1    Sanity    Resources-GPU    ODS-XXX
+    ${rc}    ${out}=    Run And Return Rc And Output    oc apply -f ${DL_POD_FILEPATH}
+    Should Be Equal As Integers    ${rc}    ${0}
+    Wait For Pods To Succeed    label_selector=gpt-download-pod=true    namespace=${TEST_NS}
+    ${rc}    ${out}=    Run And Return Rc And Output    oc apply -f ${SR_FILEPATH}
+    Should Be Equal As Integers    ${rc}    ${0}
+    ${rc}    ${out}=    Run And Return Rc And Output    oc apply -f ${IS_FILEPATH}
+    Should Be Equal As Integers    ${rc}    ${0}
+    Wait For Pods To Be Ready    label_selector=serving.kserve.io/inferenceservice=vllm-gpt2-openai
+    ...    namespace=${TEST_NS}
+    ${pod_name}=  Get Pod Name    namespace=${TEST_NS}
+    ...    label_selector=serving.kserve.io/inferenceservice=vllm-gpt2-openai
+    Start Port-forwarding    namespace=${TEST_NS}    pod_name=${pod_name}    local_port=8080   remote_port=8080
+    ${rc}    ${out}=    Run And Return Rc And Output
+    ...    curl -ks ${INFERENCE_URL} -H "Content-Type: application/json" -d ${INFERENCE_INPUT} | jq .
+    Should Be Equal As Integers    ${rc}    ${0}
+    Log    ${out}
+
+Verify Vllm Metrics Are Present
+    [Documentation]    Confirm vLLM metrics are exposed in OpenShift metrics
+    [Tags]    Tier1    Sanity    Resources-GPU    ODS-XXX
+    ${rc}    ${out}=    Run And Return Rc And Output
+    ...    curl -ks ${METRICS_URL}
+    Should Be Equal As Integers    ${rc}    ${0}
+    Log    ${out}
+    ${thanos_url}=    Get OpenShift Thanos URL
+    ${token}=    Generate Thanos Token
+    Metrics Should Exist In UserWorkloadMonitoring    ${thanos_url}    ${token}    ${SEARCH_METRICS}
+
+
+*** Keywords ***
+Suite Setup
+    Skip If Component Is Not Enabled    kserve
+    RHOSi Setup
+    Set Default Storage Class In GCP    default=ssd-csi
+    ${is_self_managed}=    Is RHODS Self-Managed
+    IF    ${is_self_managed}
+        Configure User Workload Monitoring
+        Enable User Workload Monitoring
+    END
+
+Suite Teardown
+    Set Default Storage Class In GCP    default=standard-csi
+    Terminate Process    llm-query-process    kill=true
+    ${rc}=    Run And Return Rc    oc delete inferenceservice -n ${TEST_NS} --all
+    Should Be Equal As Integers    ${rc}    ${0}
+    ${rc}=    Run And Return Rc    oc delete servingruntime -n ${TEST_NS} --all
+    Should Be Equal As Integers    ${rc}    ${0}
+    ${rc}=    Run And Return Rc    oc delete pod -n ${TEST_NS} --all
+    Should Be Equal As Integers    ${rc}    ${0}
+    ${rc}=    Run And Return Rc    oc delete namespace ${TEST_NS}
+    Should Be Equal As Integers    ${rc}    ${0}
+    RHOSi Teardown
+
+Set Default Storage Class In GCP
+    [Documentation]    If the storage class exists we can assume we are in GCP. We force ssd-csi to be the default class
+    ...    for the duration of this test suite.
+    [Arguments]    ${default}
+    ${rc}=    Run And Return Rc    oc get storageclass ${default}
+    IF    ${rc} == ${0}
+        IF    "${default}" == "ssd-csi"
+            Run    oc patch storageclass standard-csi -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"false"}}}'  #robocop: disable
+            Run    oc patch storageclass ssd-csi -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"true"}}}'  #robocop: disable
+        ELSE
+            Run    oc patch storageclass ssd-csi -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"false"}}}'  #robocop: disable
+            Run    oc patch storageclass standard-csi -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"true"}}}'  #robocop: disable
+        END
+    ELSE
+        Log    Proceeding with default storage class because we're not in GCP
+    END