homelab/cluster/applications/litellm/values.yaml

# ============================================================================
# LiteLLM Helm values — overrides for oci://ghcr.io/berriai/litellm-helm 1.106.1
# Cluster: fastpass (Talos). Host: skyliner.wdw-labs.cloud. Wave 8.
# Upstream defaults for reference: values.upstream.yaml (same dir; not applied).
# ============================================================================

replicaCount: 1

image:
  repository: ghcr.io/berriai/litellm
  # Empty tag => chart appVersion (1.106.1). Version is pinned via chart version
  # in application.yaml; bump both together.
  tag: ""
  pullPolicy: IfNotPresent

# Exposure is via Gateway API HTTPRoute (httproute.yaml), not the chart ingress.
ingress:
  enabled: false

service:
  type: ClusterIP
  port: 4000

resources:
  requests:
    cpu: 250m
    memory: 512Mi
  limits:
    memory: 2Gi

# --- Secrets -----------------------------------------------------------------
# Master key from ExternalSecret litellm-credentials (key: master-key) -> PROXY_MASTER_KEY
masterkeySecretName: litellm-credentials
masterkeySecretKey: master-key

# All keys of litellm-env exported as env vars (LITELLM_SALT_KEY, ILLUMINATIONS_/FANTASMIC_/WISHES_API_KEY)
# (includes DATABASE_URL: the Deployment template ignores db.useExisting and would
#  otherwise use the chart's placeholder DB password; envFrom supplies the real URL.
#  Not extraEnvVars: the chart also injects those into the migration Job -> duplicate env.)
environmentSecrets:
  - litellm-env

# --- Database: bundled Bitnami PostgreSQL ------------------------------------
db:
  # Bundled Postgres is still deployed (deployStandalone), but the app and the
  # migration job are pointed at it as an "existing" DB so credentials come from
  # litellm-credentials (ExternalSecret) instead of the chart's placeholder secret.
  deployStandalone: true
  useExisting: true
  endpoint: litellm-postgresql
  database: litellm
  secret:
    name: litellm-credentials
    usernameKey: username
    passwordKey: password

postgresql:
  architecture: standalone
  auth:
    username: litellm
    database: litellm
    existingSecret: litellm-credentials
    secretKeys:
      adminPasswordKey: postgres-password
      userPasswordKey: password
  primary:
    persistence:
      enabled: true
      storageClass: px-fa-direct-access   # Portworx direct access to FlashArray (as openviking)
      size: 10Gi
    resources:
      requests:
        cpu: 100m
        memory: 256Mi
      limits:
        memory: 1Gi

redis:
  enabled: false

# Schema migrations. The chart's default is an ArgoCD PreSync hook, which runs
# BEFORE Postgres/ExternalSecrets exist and would block first install. Instead run
# it as a Sync-phase hook ordered by wave: ExternalSecrets (-1) -> Postgres (0)
# -> migration Job (1) -> Deployment (2).
migrationJob:
  enabled: true
  hooks:
    argocd:
      enabled: false
    helm:
      enabled: false
  jobAnnotations:
    argocd.argoproj.io/hook: Sync
    argocd.argoproj.io/hook-delete-policy: BeforeHookCreation
    argocd.argoproj.io/sync-wave: "1"

deploymentAnnotations:
  argocd.argoproj.io/sync-wave: "2"

logLevel: INFO

# --- Prometheus ----------------------------------------------------------------
serviceMonitor:
  enabled: true
  labels:
    release: monitoring        # kube-prometheus-stack selector (matches Harbor/external-dns)
  interval: 30s
  scrapeTimeout: 10s
  namespaceSelector:
    matchNames:
      - litellm

# --- LiteLLM proxy config (rendered to ConfigMap) ------------------------------
proxy_config:
  model_list:
    # --- illuminations (RTX 3090, vLLM) --------------------------------------
    - model_name: Gemma-4-26B-A4B-it-AWQ
      litellm_params:
        model: openai/Gemma-4-26B-A4B-it-AWQ
        api_base: http://illuminations.wdw-labs.cloud:8000/v1
        api_key: os.environ/ILLUMINATIONS_API_KEY

    # --- Mac Studio (fantasmic, oMLX) — Tier 1 Hermes ------------------------
    - model_name: Gemma-4-26B-A4B-Hermes-VLM-MLX
      litellm_params:
        model: openai/Gemma-4-26B-A4B-Hermes-VLM-MLX
        api_base: http://fantasmic.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/FANTASMIC_API_KEY
    - model_name: Qwen3.8-27B-4bit
      litellm_params:
        model: openai/mlx-community--Qwen3.8-27B-4bit
        api_base: http://fantasmic.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/FANTASMIC_API_KEY
    # Also served, not exposed: incoai--Qwen3.8-27B-DFlash2, ukisai-Swift-Qwen3.8-27b-MTPLX

    # --- Mac Mini M4 (wishes, oMLX) — Tier 3 aux tasks -----------------------
    - model_name: Qwen2.5-7B-Instruct-4bit
      litellm_params:
        model: openai/Qwen2.5-7B-Instruct-4bit
        api_base: http://wishes.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/WISHES_API_KEY
    - model_name: Phi-3.5-mini-instruct-4bit
      litellm_params:
        model: openai/Phi-3.5-mini-instruct-4bit
        api_base: http://wishes.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/WISHES_API_KEY
    - model_name: Phi-3.5-mini-instruct-bf16
      litellm_params:
        model: openai/Phi-3.5-mini-instruct-bf16
        api_base: http://wishes.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/WISHES_API_KEY
    - model_name: Phi-3.5-vision-instruct-4bit
      litellm_params:
        model: openai/Phi-3.5-vision-instruct-4bit
        api_base: http://wishes.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/WISHES_API_KEY
    - model_name: gemma-2-2b-it-4bit
      litellm_params:
        model: openai/gemma-2-2b-it-4bit
        api_base: http://wishes.litellm.svc.cluster.local:8000/v1
        api_key: os.environ/WISHES_API_KEY

  litellm_settings:
    drop_params: true
    callbacks: ["prometheus"]     # exposes /metrics for the ServiceMonitor

  general_settings:
    master_key: os.environ/PROXY_MASTER_KEY
    # Allow adding models / virtual keys from the Admin UI (/ui), stored in Postgres
    store_model_in_db: true