74 lines
1.3 KiB
YAML
74 lines
1.3 KiB
YAML
fullnameOverride: ollama
|
|
|
|
replicaCount: 1
|
|
updateStrategy:
|
|
type: Recreate
|
|
terminationGracePeriodSeconds: 60
|
|
|
|
ollama:
|
|
port: 11434
|
|
|
|
extraEnv:
|
|
- name: OLLAMA_KEEP_ALIVE
|
|
value: "-1"
|
|
- name: OLLAMA_LOAD_TIMEOUT
|
|
value: 20m
|
|
- name: OLLAMA_CONTEXT_LENGTH
|
|
value: "8192"
|
|
- name: OLLAMA_MAX_LOADED_MODELS
|
|
value: "2"
|
|
- name: OLLAMA_NUM_PARALLEL
|
|
value: "1"
|
|
|
|
resources:
|
|
requests:
|
|
cpu: "2"
|
|
memory: 8Gi
|
|
limits:
|
|
memory: 96Gi
|
|
|
|
securityContext:
|
|
allowPrivilegeEscalation: false
|
|
capabilities:
|
|
drop:
|
|
- ALL
|
|
seccompProfile:
|
|
type: RuntimeDefault
|
|
|
|
livenessProbe:
|
|
path: /api/tags
|
|
initialDelaySeconds: 0
|
|
periodSeconds: 30
|
|
timeoutSeconds: 5
|
|
failureThreshold: 3
|
|
|
|
readinessProbe:
|
|
path: /api/tags
|
|
initialDelaySeconds: 0
|
|
periodSeconds: 10
|
|
timeoutSeconds: 3
|
|
failureThreshold: 3
|
|
|
|
persistentVolume:
|
|
enabled: true
|
|
existingClaim: ollama-models
|
|
|
|
nodeSelector:
|
|
kubernetes.io/arch: amd64
|
|
|
|
tolerations:
|
|
- key: node-role.kubernetes.io/control-plane
|
|
operator: Exists
|
|
effect: NoSchedule
|
|
|
|
affinity:
|
|
nodeAffinity:
|
|
preferredDuringSchedulingIgnoredDuringExecution:
|
|
- weight: 100
|
|
preference:
|
|
matchExpressions:
|
|
- key: kubernetes.io/hostname
|
|
operator: In
|
|
values:
|
|
- arschrock
|