Skip to content

Integrate Run:ai with Seldon Core

Seldon Core is software that deploys machine learning models to production over Kubernetes. The purpose of this document is to explain how to use Seldon Core together with Run:ai.

Of special importance, is the usage of Seldon together with the Run:ai fractions technology: Machine learning production tends to take less GPU Memory. As such, allocating a fraction of the GPU per job allows for better GPU Utilization.

Prerequisites

Install Seldon Core as described here. We recommend using the helm-based installation of both Seldon Core and Istio.

Create a Seldon deployment

The instructions below follow a sample machine learning model that tests the Run:ai - Seldon Core integration. Save the following in a file named <FILE-NAME>.yaml

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: seldon-model
  namespace: runai-<PROJECT-NAME>
spec:
  name: test-deployment
  predictors:
  - componentSpecs:
    - spec:
        containers:
        - name: classifier
          image: seldonio/mock_classifier:1.5.0-dev
          resources:
            limits:
              nvidia.com/gpu: <GPUs>
        schedulerName: runai-scheduler
    graph:
      children: []
      endpoint:
        type: REST
      name: classifier
      type: MODEL
    name: example
    replicas: 1

apiVersion: machinelearning.seldon.io/v1 kind: SeldonDeployment metadata: name: seldon-model namespace: runai- spec: name: test-deployment predictors: - componentSpecs: - spec: containers: - name: classifier image: seldonio/mock_classifier:1.0 resources: limits: nvidia.com/gpu: schedulerName: runai-scheduler graph: children: [] endpoint: type: REST name: classifier type: MODEL name: example replicas: 1

Replace <PROJECT-NAME> with the Run:ai projects and <GPUs> with the amount of GPUs you want to allocate (e.g. 0.5 GPUs).

kubectl apply -f <FILE-NAME>.yaml

Verification

Run: runai list jobs and verify that the job is running

Delete a deployment

Run:

kubectl delete -f <FILE-NAME>.yaml