Uploaded image for project: 'Red Hat OpenShift Data Science'
  1. Red Hat OpenShift Data Science
  2. RHODS-8796

OpenVINO Model Server runtime does not have the required flag to force GPU usage

    XMLWordPrintable

Details

    • False
    • None
    • False
    • Testable
    • No
    • 1.28.0
    • No
    • Hide
      == OpenVINO Model Server runtime did not have the required flag to force GPU usage
      OpenShift Data Science includes the OpenVINO Model Server (OVMS) model-serving runtime by default. When you configured a new model server and chose this runtime, the *Configure model server* dialog enabled you to specify a number of GPUs to use with the model server. However, when you finished configuring the model server and deployed models from it, the model server did not actually use any GPUs. This issue is now fixed and the model server uses the GPUs.
      Show
      == OpenVINO Model Server runtime did not have the required flag to force GPU usage OpenShift Data Science includes the OpenVINO Model Server (OVMS) model-serving runtime by default. When you configured a new model server and chose this runtime, the *Configure model server* dialog enabled you to specify a number of GPUs to use with the model server. However, when you finished configuring the model server and deployed models from it, the model server did not actually use any GPUs. This issue is now fixed and the model server uses the GPUs.
    • Bug Fix
    • Done
    • No
    • Pending
    • None
    • RHODS 1.28

    Description

      Description of problem:

      Following up from https://issues.redhat.com/browse/RHODS-6529, we should now be able to deploy OVMS runtimes that request (and use) GPUs.
      In the current 1.27 build we are able to request a GPU for the runtime via UI (assuming GPUs are available in the cluster), but the ServingRuntime that gets created does not have the correct flags to force the model to be served on the gpu device, i.e. (from https://github.com/opendatahub-io/modelmesh-runtime-adapter/pull/15):

      spec:
  builtInAdapter:
    env:
      - name: OVMS_FORCE_TARGET_DEVICE
        value: NVIDIA

      This is the full ServingRuntime def created when requesting 1 GPU:

       

      apiVersion: serving.kserve.io/v1alpha1
kind: ServingRuntime
metadata:
  annotations:
    enable-auth: 'false'
    enable-route: 'true'
    opendatahub.io/template-display-name: OpenVINO Model Server
    opendatahub.io/template-name: ovms
    openshift.io/display-name: ovms-test
  resourceVersion: '585132'
  name: ovms-test
  uid: 2a220f40-6988-4b24-b1f1-f39f9ba810f8
  creationTimestamp: '2023-05-18T15:27:09Z'
  generation: 1
  managedFields:
    - apiVersion: serving.kserve.io/v1alpha1
      fieldsType: FieldsV1
      fieldsV1:
        'f:metadata':
          'f:annotations':
            .: {}
            'f:enable-auth': {}
            'f:enable-route': {}
            'f:opendatahub.io/template-display-name': {}
            'f:opendatahub.io/template-name': {}
            'f:openshift.io/display-name': {}
          'f:labels':
            .: {}
            'f:name': {}
            'f:opendatahub.io/dashboard': {}
        'f:spec':
          'f:builtInAdapter':
            .: {}
            'f:memBufferBytes': {}
            'f:modelLoadingTimeoutMillis': {}
            'f:runtimeManagementPort': {}
            'f:serverType': {}
          'f:multiModel': {}
          'f:containers': {}
          'f:protocolVersions': {}
          'f:grpcEndpoint': {}
          'f:supportedModelFormats': {}
          .: {}
          'f:replicas': {}
          'f:grpcDataEndpoint': {}
      manager: unknown
      operation: Update
      time: '2023-05-18T15:27:09Z'
  namespace: test
  labels:
    name: ovms-test
    opendatahub.io/dashboard: 'true'
spec:
  builtInAdapter:
    memBufferBytes: 134217728
    modelLoadingTimeoutMillis: 90000
    runtimeManagementPort: 8888
    serverType: ovms
  containers:
    - args:
        - '--port=8001'
        - '--rest_port=8888'
        - '--config_path=/models/model_config_list.json'
        - '--file_system_poll_wait_seconds=0'
        - '--grpc_bind_address=127.0.0.1'
        - '--rest_bind_address=127.0.0.1'
      image: >-
        quay.io/opendatahub/openvino_model_server@sha256:20dbfbaf53d1afbd47c612d953984238cb0e207972ed544a5ea662c2404f276d
      name: ovms
      resources:
        limits:
          cpu: '2'
          memory: 8Gi
          nvidia.com/gpu: 1
        requests:
          cpu: '1'
          memory: 4Gi
          nvidia.com/gpu: 1
  grpcDataEndpoint: 'port:8001'
  grpcEndpoint: 'port:8085'
  multiModel: true
  protocolVersions:
    - grpc-v1
  replicas: 1
  supportedModelFormats:
    - autoSelect: true
      name: openvino_ir
      version: opset1
    - autoSelect: true
      name: onnx
      version: '1'

      Prerequisites (if any, like setup, operators/versions):

      RHODS 1.27 RC (iib:498989)

      Steps to Reproduce

      1. Provision GPU Node
      2. Install Nvidia GPU Add-On
      3. Create Data Science Project
      4. Configure Model Server (with at least 1 GPU)
      5. Deploy a model
      6. Look at the ServingRuntime def for the Model Server in the DSP namespace

      Actual results:

      No force flag for Nvidia devices

      Expected results:

      force flag for Nvidia devices in ServingRuntime

      Reproducibility (Always/Intermittent/Only Once):

      Always

      Build Details:

      Workaround:

      Add the flag manually to the ServingRuntime, let the model server redeploy

      Additional info:

      Attachments

        Activity

          People

            lferrnan@redhat.com Lucas Fernandez Aragon
            rhn-support-lgiorgi Luca Giorgi
            Luca Giorgi Luca Giorgi
            Votes:
            0 Vote for this issue
            Watchers:
            9 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved: