values.yaml

# -- Default values for llmstack helm chart
# -- Declare variables to be passed into your templates.

# -- Serving engine configuratoon
servingEngineSpec:
  enableEngine: true
  # -- Customized labels for the serving engine deployment
  labels:
    environment: "test"
    release: "test"
  # -- Extra service ports for models
  extraPorts: []
  # vllmApiKey: (optional) api key for securing the vLLM models. Can be either:
  #   - A string containing the token directly (will be stored in a generated secret)
  #   - An object referencing an existing secret:
  #     secretName: "my-existing-secret"
  #     secretKey: "vllm-api-key"
  #
  # modelSpec - configuring multiple serving engines deployments that runs different models
  # Each entry in the modelSpec array should contain the following fields:
  # - annotations: (Optional, map) The annotations to add to the deployment, e.g., {model: "opt125m"}
  # - serviceAccountName: (Optional, string) The name of the service account to use for the deployment, e.g., "vllm-service-account"
  # - priorityClassName: (Optional, string) The name of the priority class name for the deployment, e.g., "high-priority"
  # - runtimeClassName: (Optional, string) Runtime class for the pod, e.g., "nvidia". If not specified, falls back to servingEngineSpec.runtimeClassName
  # - podAnnotations: (Optional, map) The annotations to add to the pod, e.g., {model: "opt125m"}
  # - name: (string) The name of the model, e.g., "example-model"
  # - repository: (string) The repository of the model, e.g., "vllm/vllm-openai"
  # - tag: (string) The tag of the model, e.g., "latest"
  # - imagePullSecret: (Optional, string) Name of secret with credentials to private container repository, e.g. "secret"
  # - modelURL: (string) The URL of the model, e.g., "facebook/opt-125m"
  # - chatTemplate: (Optional, string) Chat template (Jinga2) specifying tokenizer configuration, e.g. "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ 'Question:\n' + message['content'] + '\n\n' }}{% elif message['role'] == 'system' %}\n{{ 'System:\n' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Answer:\n'  + message['content'] + '\n\n' }}{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ 'Answer:\n' }}{% endif %}{% endfor %}"
  #
  # - replicaCount: (int) The number of replicas for the model, e.g. 1
  # - requestCPU: (int) The number of CPUs requested for the model, e.g. 6
  # - requestMemory: (string) The amount of memory requested for the model, e.g., "16Gi"
  # - requestGPU: (int) The number of GPUs requested for the model, e.g., 1
  # - requestGPUType: (Optional, string) The type of GPU requested, e.g., "nvidia.com/mig-4g.71gb". If not specified, defaults to "nvidia.com/gpu"
  # - requestGPUMem: (Optional, string) Requires HAMi. The amount of GPU memory requested, e.g., 3000. Each unit equals to 1M. https://project-hami.io/docs/userguide/NVIDIA-device/specify-device-memory-usage
  # - requestGPUMemPercentage: (Optional, string) Requires HAMi. The percentage of GPU memory requested, e.g., "80". It cannot be used together with requestGPUMem.
  # - requestGPUCores: (Optional, string) Requires HAMi. The percentage of GPU cores requested, e.g., "10". Each unit equals to 1% device cores. https://project-hami.io/docs/userguide/NVIDIA-device/specify-device-core-usage
  # - limitCPU: (Optional, string) The CPU limit for the model, e.g., "8"
  # - limitMemory: (Optional, string) The memory limit for the model, e.g., "32Gi"
  # Note: If limitCPU and limitMemory are not specified, only GPU resources will have limits set equal to their requests.
  # - limitGPUMem: (Optional, string) Requires HAMi. The limit of GPU memory, e.g., 3000. Each unit equals to 1M. https://project-hami.io/docs/userguide/NVIDIA-device/specify-device-memory-usage
  # - limitGPUMemPercentage: (Optional, string) Requires HAMi. The limit of GPU memory of GPU, e.g., "80"
  # - limitGPUCores: (Optional, string) Requires HAMi. The limit of GPU cores, e.g., "10". Each unit equals to 1% device cores. https://project-hami.io/docs/userguide/NVIDIA-device/specify-device-core-usage

  # - pvcStorage: (Optional, string) The amount of storage requested for the model, e.g., "50Gi".
  # - pvcAccessMode: (Optional, list) The access mode policy for the mounted volume, e.g., ["ReadWriteOnce"]
  # - storageClass: (Optional, String) The storage class of the PVC e.g., "", default is ""
  # - pvcMatchLabels: (Optional, map) The labels to match the PVC, e.g., {model: "opt125m"}
  # - pvcLabels: (Optional, map} The labels to add to the PVC, e.g., {label_excluded_from_alerts: "true"}
  # - extraVolumes: (Optional, list) Additional volumes to add to the pod, in Kubernetes volume format. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#volume-v1-core
  #   Example for an emptyDir volume:
  #   extraVolumes:
  #   - name: tmp-volume
  #     emptyDir:
  #       medium: ""
  #       sizeLimit: 5Gi
  # - extraVolumeMounts: (Optional, list) Additional volume mounts to add to the container, in Kubernetes volumeMount format. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#volumemount-v1-core
  #   Example for mounting the tmp-volume to /tmp:
  #   extraVolumeMounts:
  #   - name: tmp-volume
  #     mountPath: /tmp
  # - initContainer: (optional, list of objects) The configuration for the init container to be run before the main container.
  #   - name: (string) The name of the init container, e.g., "init"
  #   - image: (string) The Docker image for the init container, e.g., "busybox:latest"
  #   - command: (optional, list) The command to run in the init container, e.g., ["sh", "-c"]
  #   - args: (optional, list) Additional arguments to pass to the command, e.g., ["ls"]
  #   - env: (optional, list) List of environment variables to set in the container, each being a map with:
  #   - resources: (optional, map) The resource requests and limits for the container:
  #   - mountPvcStorage: (optional, bool) Whether to mount the model's volume.
  #   - extraVolumeMounts: (Optional, list) Additional volume mounts to add to the init container, in Kubernetes volumeMount format. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#volumemount-v1-core
  #
  # - vllmConfig: (optional, map) The configuration for the VLLM model, supported options are:
  #   - v0: (optional, map) Specify to 1 to use vLLM v0, otherwise vLLM v1 is used. e.g., 0
  #   - enablePrefixCaching: (optional, bool) Enable prefix caching, e.g., false
  #   - enableChunkedPrefill: (optional, bool) Enable chunked prefill, e.g., false
  #   - maxModelLen: (optional, int) The maximum model length, e.g., 16384
  #   - dtype: (optional, string) The data type, e.g., "bfloat16"
  #   - tensorParallelSize: (optional, int) The degree of tensor parallelism, e.g., 2
  #   - maxNumSeqs: (optional, int) Maximum number of sequences to be processed in a single iteration., e.g., 32
  #   - maxLoras: (optional, int) The maximum number of LoRA models to be loaded in a single batch, e.g., 4
  #   - gpuMemoryUtilization: (optional, float) The fraction of GPU memory to be used for the model executor, which can range from 0 to 1. e.g., 0.95
  #   - extraArgs: (optional, list) Extra command line arguments to pass to vLLM, e.g., ["--disable-log-requests"]
  #
  # - lmcacheConfig: (optional, map) The configuration of the LMCache for KV offloading, supported options are:
  #   - enabled: (optional, bool) Enable LMCache, e.g., true
  #   - cpuOffloadingBufferSize: (optional, string) The CPU offloading buffer size, e.g., "30"
  #
  # - hf_token: (optional) Hugging Face token configuration. Can be either:
  #   - A string containing the token directly (will be stored in a generated secret)
  #   - An object referencing an existing secret:
  #     secretName: "my-existing-secret"
  #     secretKey: "hf-token-key"
  #
  # - envFromSecret: (optional) Reference to an existing Kubernetes Secret from which
  #   all key/value pairs will be loaded as environment variables into the container.
  #   Example:
  #     envFromSecret:
  #       name: s3-registry
  #
  #   The referenced Secret could look like:
  #     apiVersion: v1
  #     kind: Secret
  #     metadata:
  #       name: s3-registry
  #       namespace: vllm
  #     data:
  #       AWS_ACCESS_KEY_ID: <base64-encoded-value>
  #       AWS_SECRET_ACCESS_KEY: <base64-encoded-value>
  #       AWS_ENDPOINT_URL: <base64-encoded-value>
  #
  # - env: (optional, list) The environment variables to set in the container, e.g., your HF_TOKEN
  #
  # - affinity: (optional, map) Affinity configuration
  # - nodeSelectorTerms: (optional, list) The node selector terms to match the nodes. When both affinity and nodeSelectorTerms are defined, nodeSelectorTerms will be ignored.
  # - nodeName: (optional) Directly assigns a pod to a specific node (e.g., "192.168.56.5"). When both nodeName and nodeSelectorTerms are defined, the preference is given to nodeName.
  # - shmSize: (optional, string) The size of the shared memory, e.g., "20Gi"
  # - enableLoRA: (optional, bool) Whether to enable LoRA, e.g., true
  #
  # Example:
  # vllmApiKey: "vllm_xxxxxxxxxxxxx"
  # modelSpec:
  # - name: "mistral"
  #   annotations:
  #     model: "mistral"
  #   podAnnotations:
  #     model: "mistral"
  #   serviceAccountName: "vllm-service-account"
  #   repository: "lmcache/vllm-openai"
  #   tag: "latest"
  #   modelURL: "mistralai/Mistral-7B-Instruct-v0.2"
  #   replicaCount: 1
  #   imagePullPolicy: "Always" # or "IfNotPresent"
  #
  #   requestCPU: 10
  #   requestMemory: "64Gi"
  #   requestGPU: 1
  #
  #   pvcStorage: "50Gi"
  #   pvcLabels:
  #     label_excluded_from_alerts: "true"
  #   pvcAccessMode:
  #     - ReadWriteOnce
  #   pvcMatchLabels:
  #     model: "mistral"
  #   initContainer:
  #     name: my-container
  #     image: busybox
  #     command: ["sh"]
  #     env: {}
  #     args: []
  #     resources: {}
  #     mountPvcStorage: true
  #
  #   vllmConfig:
  #     enableChunkedPrefill: false
  #     enablePrefixCaching: false
  #     maxModelLen: 16384
  #     dtype: "bfloat16"
  #     maxNumSeqs: 32
  #     gpuMemoryUtilization: 0.95
  #     maxLoras: 4
  #     extraArgs: ["--disable-log-requests", "--trust-remote-code"]
  #
  #   lmcacheConfig:
  #     enabled: true
  #     cpuOffloadingBufferSize: "30"
  #
  #   hf_token: "hf_xxxxxxxxxxxxx"
  #
  #
  #   nodeSelectorTerms:
  #     - matchExpressions:
  #       - key: nvidia.com/gpu.product
  #         operator: "In"
  #         values:
  #         - "NVIDIA-RTX-A6000"

  #  extraVolumes:
  #    - name: dev-fuse
  #      hostPath:
  #        path: /dev/fuse
  #        type: CharDevice
  #    - name: cache-dir
  #      hostPath:
  #        path: /var/cache/vllm

  #    extraVolumeMounts:
  #    - name: dev-fuse
  #      mountPath: /dev/fuse
  #      readOnly: true
  #    - name: cache-dir
  #      mountPath: /cache
  #      readOnly: false

  modelSpec: []

  # -- Container port
  containerPort: 8000
  # -- Service port
  servicePort: 80

  # -- Set other environment variables from config map
  configs: {}

  # -- deployment strategy
  strategy: {}

  # -- Readiness probe configuration
  startupProbe:
    # -- Number of seconds after the container has started before startup probe is initiated
    initialDelaySeconds: 15
    # -- How often (in seconds) to perform the startup probe
    periodSeconds: 10
    # -- Number of times after which if a probe fails in a row, Kubernetes considers that the overall check has failed: the container is not ready
    failureThreshold:
      60
      # -- Configuration of the Kubelet http request on the server
    httpGet:
      # -- Path to access on the HTTP server
      path: /health
      # -- Name or number of the port to access on the container, on which the server is listening
      port: 8000

  # -- Liveness probe configuration
  livenessProbe:
    # -- Number of seconds after the container has started before liveness probe is initiated
    initialDelaySeconds: 15
    # -- Number of times after which if a probe fails in a row, Kubernetes considers that the overall check has failed: the container is not alive
    failureThreshold: 3
    # -- How often (in seconds) to perform the liveness probe
    periodSeconds: 10
    # -- Configuration of the Kubelet http request on the server
    httpGet:
      # -- Path to access on the HTTP server
      path: /health
      # -- Name or number of the port to access on the container, on which the server is listening
      port: 8000

  # -- Readiness probe configuration
  readinessProbe:
    # -- Number of seconds after the container has started before readiness probe is initiated
    initialDelaySeconds: 15
    # -- Number of times after which if a probe fails in a row, Kubernetes considers that the overall check has failed: the container is not alive
    failureThreshold: 3
    # -- How often (in seconds) to perform the readiness probe
    periodSeconds: 5
    # -- Configuration of the Kubelet http request on the server
    httpGet:
      # -- Path to access on the HTTP server
      path: /health
      # -- Name or number of the port to access on the container, on which the server is listening
      port: 8000

  # -- Disruption Budget Configuration
  maxUnavailablePodDisruptionBudget: ""

  # -- Tolerations configuration (when there are taints on nodes)
  # Example:
  # tolerations:
  #   - key: "node-role.kubernetes.io/control-plane"
  #     operator: "Exists"
  #     effect: "NoSchedule"
  tolerations: []

  # -- RuntimeClassName configuration, set to "nvidia" if the model requires GPU
  runtimeClassName: "nvidia"

  # -- SchedulerName configuration
  schedulerName: ""

  # -- Pod-level security context configuration. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#podsecuritycontext-v1-core
  securityContext: {}
    # -- Run as a non-root user ID
    # runAsUser: 1000
    # -- Run with a non-root group ID
    # runAsGroup: 1000
    # -- Run as non-root
    # runAsNonRoot: true
    # -- Set the seccomp profile
    # seccompProfile:
    #   type: RuntimeDefault
    # -- Drop all capabilities
    # capabilities:
    #   drop:
    #   - ALL
    # -- Set the file system group ID for all containers
    # fsGroup: 1000

  # -- Container-level security context configuration. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#securitycontext-v1-core
  containerSecurityContext:
    # -- Run as non-root
    runAsNonRoot: false
    # -- Don't allow privilege escalation
    # allowPrivilegeEscalation: false
    # -- Drop all capabilities
    # capabilities:
    #   drop:
    #   - ALL
    # -- Read-only root filesystem
    # readOnlyRootFilesystem: true

  # -- Sidecar configuration
  sidecar:
    image: "lmcache/lmstack-sidecar:latest"
    imagePullPolicy: "Always"

routerSpec:
  # -- The docker image of the router. The following values are defaults:
  repository: "lmcache/lmstack-router"
  tag: "latest"
  imagePullPolicy: "Always"

  # -- Pod-level security context configuration. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#podsecuritycontext-v1-core
  securityContext: {}

  # -- Container-level security context configuration. https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.32/#securitycontext-v1-core
  containerSecurityContext: {}

  # -- Whether to enable the router service
  enableRouter: true

  # -- Number of replicas
  replicaCount: 1

  # -- autoscaling configuration
  autoscaling:
    enabled: false
    minReplicas: 1
    maxReplicas: 3
    targetCPUUtilizationPercentage: 80

  # -- Priority Class
  priorityClassName: ""

  # -- Container port
  containerPort: 8000

  # -- Service type
  serviceType: ClusterIP
  # -- Service annotations if you use a LoadBalancer or NodePort service type
  serviceAnnotations: {}

  # -- Service port
  servicePort: 80

  # -- Service discovery mode, supports "k8s" or "static". Defaults to "k8s" if not set.
  serviceDiscovery: "k8s"
  # -- Service discovery mode type, supports "pod-ip" or "service-name". Defaults to "pod-ip" if not set.
  k8sServiceDiscoveryType: "pod-ip"
  # -- If serviceDiscovery is set to "static", the comma-separated values below are required. There needs to be the same number of backends and models
  staticBackends: ""
  staticModels: ""

  # -- routing logic, could be "roundrobin" or "session"
  routingLogic: "roundrobin"

  # -- session key if using "session" routing logic
  sessionKey: ""

  # -- extra router commandline arguments
  extraArgs: []
  # -- extra service ports
  extraPorts: []

  # -- Interval in seconds to scrape the serving engine metrics
  engineScrapeInterval: 15

  # -- Window size in seconds to calculate the request statistics
  requestStatsWindow: 60

  # -- deployment strategy
  strategy: {}

  # vllmApiKey: (optional) api key for securing the vLLM models. Must be an object referencing an existing secret
  #   secretName: "my-existing-secret"
  #   secretKey: "vllm-api-key"

  # -- router resource requests and limits
  resources:
    requests:
      cpu: 400m
      memory: 500Mi
    limits:
      memory: 500Mi

  # -- Customized labels for the router deployment
  labels:
    environment: "router"
    release: "router"

  # -- Customized pod annotations for the router pods
  podAnnotations: {}

  ingress:
    # -- Enable ingress controller resource
    enabled: false

    # -- IngressClass that will be used to implement the Ingress
    className: ""

    # -- Additional annotations for the Ingress resource
    annotations:
      {}
      # kubernetes.io/ingress.class: alb
      # kubernetes.io/ingress.class: nginx
      # kubernetes.io/tls-acme: "true"

    # The list of hostnames to be covered with this ingress record.
    hosts:
      - host: vllm-router.local
        paths:
          - path: /
            pathType: Prefix

    # --  The tls configuration for hostnames to be covered with this ingress record.
    tls: []
    #  - secretName: vllm-router-tls
    #    hosts:
    #      - vllm-router.local

  # -- Expose the service via gateway-api HTTPRoute
  # More routes can be added by adding a dictionary key like the 'main' route.
  # Requires Gateway API resources and suitable controller installed within the cluster
  # (see: https://gateway-api.sigs.k8s.io/guides/)
  route:
    main:
      # -- Enables or disables the route
      enabled: false

      # -- Set the route apiVersion, e.g. gateway.networking.k8s.io/v1 or gateway.networking.k8s.io/v1alpha2
      apiVersion: gateway.networking.k8s.io/v1
      # -- Set the route kind
      # Valid options are GRPCRoute, HTTPRoute, TCPRoute, TLSRoute, UDPRoute
      kind: HTTPRoute

      annotations: {}
      labels: {}

      hostnames: []
      # - my-filter.example.com
      parentRefs: []
      # - name: acme-gw

      # -- create http route for redirect (https://gateway-api.sigs.k8s.io/guides/http-redirect-rewrite/#http-to-https-redirects)
      ## Take care that you only enable this on the http listener of the gateway to avoid an infinite redirect.
      ## matches, filters and additionalRules will be ignored if this is set to true.
      httpsRedirect: false

      matches:
        - path:
            type: PathPrefix
            value: /

      ## Filters define the filters that are applied to requests that match this rule.
      filters: []

      ## Additional custom rules that can be added to the route
      additionalRules: []

  # Affinity configuration
  affinity: {}

  # The node selector terms to match the nodes. Will be ignored if affinity is configured.
  # Example:
  #   nodeSelectorTerms:
  #     - matchExpressions:
  #       - key: nvidia.com/gpu.product
  #         operator: "In"
  #         values:
  #         - "NVIDIA-RTX-A6000"
  nodeSelectorTerms: []

  # Router probes configuration
  # <probe>.httpGet.port will automatically be set to .Values.routerSpec.containerPort

  # Liveness probe configuration.
  livenessProbe:
    initialDelaySeconds: 30
    periodSeconds: 5
    failureThreshold: 3
    httpGet:
      path: /health

  # Startup probe configuration.
  startupProbe:
    initialDelaySeconds: 5
    periodSeconds: 5
    failureThreshold: 3
    httpGet:
      path: /health

  # Readiness probe configuration.
  readinessProbe:
    initialDelaySeconds: 30
    periodSeconds: 5
    failureThreshold: 3
    httpGet:
      path: /health

  # -- TODO: Readiness probe configuration
  #startupProbe:
  #  # -- Number of seconds after the container has started before startup probe is initiated
  #  initialDelaySeconds: 5
  #  # -- How often (in seconds) to perform the startup probe
  #  periodSeconds: 5
  #  # -- Number of times after which if a probe fails in a row, Kubernetes considers that the overall check has failed: the container is not ready
  #  failureThreshold: 100
  #   # -- Configuration of the Kubelet http request on the server
  #  httpGet:
  #    # -- Path to access on the HTTP server
  #

# -- LoRA Adapter Configuration
loraAdapters: []

  # -- LoRA adapter instances to deploy
  # Each instance should contain:
  # - name: (string) The name of the LoRA adapter instance
  # - baseModel: (string) The name of the base model this adapter is for
  # - vllmApiKey: (optional) API key configuration for vLLM authentication
  #   - secretRef: Reference to a secret containing the API key
  #     - secretName: (string) Name of the secret
  #     - secretKey: (string) Key in the secret containing the API key
  #   - value: (string) Direct API key value
  # - adapterSource: (object) Configuration for the adapter source
  #   - type: (string) Type of adapter source (local, s3, http, huggingface)
  #   - adapterName: (string) Name of the adapter to apply
  #   - adapterPath: (optional, string) Path to the LoRA adapter weights
  #   - repository: (optional, string) Repository to get the LoRA adapter from
  #   - pattern: (optional, string) Pattern to use for the adapter name
  #   - maxAdapters: (optional, int) Maximum number of adapters to load
  #   - credentials: (optional, object) Reference to secret with storage credentials
  #     - secretName: (string) Name of the secret
  #     - secretKey: (string) Key in the secret containing the credentials
  # - loraAdapterDeploymentConfig: (object) Configuration for adapter deployment
  #   - algorithm: (string) Placement algorithm to use (default, ordered, equalized)
  #   - replicas: (optional, int) Number of replicas that should load this adapter
  # - labels: (optional, map) Additional labels for the LoRA adapter
  #
  # Example:
  # loraAdapters:
  #   - name: "llama3-nemoguard-adapter"
  #     baseModel: "llama3-8b-instr"
  #     vllmApiKey:
  #       secretRef:
  #         secretName: "vllm-api-key"
  #         secretKey: "VLLM_API_KEY"
  #     adapterSource:
  #       type: "local"
  #       adapterName: "llama-3.1-nemoguard-8b-topic-control"
  #       adapterPath: "/data/lora-adapters/llama-3.1-nemoguard-8b-topic-control"
  #     loraAdapterDeploymentConfig:
  #       algorithm: "default"
  #       replicas: 1
  #     labels:
  #       environment: "production"
  #       model: "llama3-nemoguard"

# -- lora controller Configuration
loraController:
  enableLoraController: false
  # -- kubernetes cluster domain
  kubernetesClusterDomain: "cluster.local"

  # -- Number of lora controller replicas
  replicaCount: 1

  # -- lora controller image configuration
  image:
    repository: "lmcache/lmstack-lora-controller"
    tag: "latest"
    pullPolicy: "IfNotPresent"

  # -- Image pull secrets
  imagePullSecrets: []

  # -- Pod annotations
  podAnnotations: {}

  # -- Pod security context
  podSecurityContext:
    runAsNonRoot: true
    seccompProfile:
      type: RuntimeDefault

  # -- Container security context
  containerSecurityContext:
    allowPrivilegeEscalation: false
    capabilities:
      drop:
      - ALL

  # -- lora controller resources
  resources: {}

  # -- Node selector
  nodeSelector: {}

  # -- Affinity
  affinity: {}

  # -- Tolerations
  tolerations: []

  # -- Environment variables
  env: []

  # -- Extra arguments for the lora controller
  extraArgs: []

  # -- Metrics configuration
  metrics:
    enabled: true

  # -- Webhook configuration
  webhook:
    enabled: false