launcher/nemo/k8s_templates/training/values.yaml

image:
  # training image
  trainingImage: cfg.container

  # image pulling policy
  pullPolicy: IfNotPresent


trainingConfig:
  # current job name
  jobName: "nil"

  # namespace to launch job
  namespace: "default"

  # script path
  scriptPath: null

  # script args
  scriptArgs: null

  # specify whether to use custom scripts
  customScript: null

  # list of custom annotations apply to jobs
  annotations: null

  # list of custom labels apply to jobs and pods
  customLabels: null

  # Kueue scheduler priority class name
  priority_class_name: null

  # device type, can be "gpu", "trainium" and "nil", "nil" means cpu
  device: "nil"

  # number of EFA devices if the instance type support EFA
  numEFADevices: 0

  # number of Neuron devices if job is for Trainium
  numNeuronDevices: null

  # number of process per node
  ntasksPerNode: 0

  # number of nodes to run
  nodes: training.trainer.num_nodes

  # restart policy
  restartPolicy: Never

  # from NeMo, not used currently
  wandbKey: "nil"

  # name of service account associated with the namespace
  serviceAccountName: null

  # relevant for Trainium chips, either 0 or 1
  compile: 0

  # persistent volume, usually used to mount FSx
  persistentVolumeClaims: null

  # temp volume, usually used to mount temp file in the host
  volumes: null

  # A github repo if user might want to use script inside
  git:
    repo_url_or_path: null
    branch: null
    commit: null
    token: null
    update_adapter: null

  # Commands to run before training
  pre_script: []
  # Commands to run after training
  post_script: []

  # select preferred and required labels for nodes
  labelSelector:
    required: null # select nodes with required labels
    preferred: null # select nodes with priority which has preferred labels
    weights: null # list of weights for the preferred labels

  # The clean up policy after the job completes or fails.
  cleanPodPolicy: null