ContextQMD
Back to Qwen3 Coder

Running "eval-dev-quality" on Kubernetes

qwencoder-eval/instruct/eval-dev-quality/docs/kubernetes/README.md

latest1.5 KB
Original Source

Running "eval-dev-quality" on Kubernetes

Prerequisite Checklist

  • kubectl installed and configured with authentication to the cluster.
  • Dedicated Namespace to run the jobs.
  • RWX volume to store the evaluation results (check volume.yml for inspiration).

Defining secrets

The Job automatically references to a secret called evaluation-secret which is configured to pass the values on as environment Variables. The following key is required to be created.

  • PROVIDER_TOKEN - contains the API tokens for different providers. E.g.: openrouter:abcdefgh1234,custom-provider:abcdefgh1234
bash
kubectl --namespace eval-dev-quality create secret generic evaluation-secret --from-literal='PROVIDER_TOKEN='

Running multiple evaluations with the eval-dev-quality kubernetes runtime

  • Define all the models with --model which should be run inside the containerized workload.
  • Define the parameter --runtime kubernetes to indicate that jobs should run inside a kubernetes cluster.
  • Define the parameter --parallel 20 to indicate how many jobs should run simultaneously.

Example:

bash
eval-dev-quality evaluate --runtime kubernetes --runs 5 --model symflower/symbolic-execution --model symflower/symbolic-execution --model symflower/symbolic-execution --repository golang/plain --parallel 2

This commands run 3x the symflower/symbolic-execution model with 5 runs of each model inside a containerized workload on the kubernetes cluster, it will limit the parallel execution to 2 containers at the same time.