apiVersion: apps.nvidia.com/v1alpha1
kind: NIMCache
metadata:
  name: qwen3-32b-instruct
  namespace: nim-service
spec:
  source:
    ngc:
      modelPuller: nvcr.io/nim/qwen/qwen3-32b-dgx-spark:1.1.0-variant
      pullSecret: ngc-secret
      authSecret: ngc-api-secret
      model:
        engine: "vllm"
        tensorParallelism: "1"
        profiles:
          - c4f105d92c72ab56200884dfacde9d2128b139755c06b9c883eeb3e287b7408a
  storage:
    pvc:
      create: true
      size: "100Gi"
      volumeAccessMode: ReadWriteOnce
---
apiVersion: apps.nvidia.com/v1alpha1
kind: NIMCache
metadata:
  name: qwen36-27b-fp8
  namespace: nim-service
spec:
  source:
    hf:
      endpoint: "https://huggingface.co"
      namespace: "Qwen"
      authSecret: hf-api-secret
      modelPuller: nvcr.io/nim/nvidia/llm-nim:1.12
      pullSecret: ngc-secret
      modelName: "Qwen3.6-27B-FP8"
  storage:
    pvc:
      create: true
      size: "50Gi"
      volumeAccessMode: ReadWriteOnce
---
apiVersion: apps.nvidia.com/v1alpha1
kind: NIMService
metadata:
  name: qwen36-27b-fp8
  namespace: nim-service
spec:
  userID: 0
  groupID: 0
  image:
    repository: vllm/vllm-openai
    tag: v0.19.1-cu130
    pullPolicy: IfNotPresent
  env:
    - name: VLLM_CACHE_ROOT
      value: /model-store/vllm-cache
  command:
    - python3
  args:
    - -m
    - vllm.entrypoints.openai.api_server
    - --model
    - /model-store
    - --host
    - 0.0.0.0
    - --served-model-name
    - qwen36
    - --gpu-memory-utilization
    - "0.85"
    - --max-model-len
    - "256000"
    - --language-model-only
    - --reasoning-parser
    - qwen3
    - --enable-auto-tool-choice
    - --tool-call-parser
    - qwen3_coder
    - --enable-chunked-prefill
    - --max-num-batched-tokens
    - "32768"
    - --max-num-seqs
    - "10"
    - --enable-prefix-caching
    - --speculative-config
    - '{"method":"mtp","num_speculative_tokens":2}'

  authSecret: hf-api-secret
  storage:
    nimCache:
      name: qwen36-27b-fp8
  replicas: 1
  resources:
    limits:
      nvidia.com/gpu: 1
  livenessProbe:
    httpGet:
      path: /health
      port: 8000
    initialDelaySeconds: 120
    periodSeconds: 30
    timeoutSeconds: 10
    failureThreshold: 10

  readinessProbe:
    httpGet:
      path: /health
      port: 8000
    initialDelaySeconds: 30
    periodSeconds: 15
    timeoutSeconds: 10
    failureThreshold: 20

  startupProbe:
    httpGet:
      path: /health
      port: 8000
    initialDelaySeconds: 10
    periodSeconds: 20
    timeoutSeconds: 10
    failureThreshold: 60
  expose:
    service:
      type: ClusterIP
      port: 8000
---
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
  name: llm-route
  namespace: nim-service
spec:
  parentRefs:
    - name: envoy-gateway
      namespace: default
  hostnames:
    - "mcp.corredorconect.com"
  rules:
    - matches:
        - path:
            type: PathPrefix
            value: /
      backendRefs:
        - name: qwen36-27b-fp8
          port: 8000
      timeouts:
        request: 600s
        backendRequest: 600s
---
apiVersion: gateway.envoyproxy.io/v1alpha1
kind: SecurityPolicy
metadata:
  name: llm-bearer-auth
  namespace: nim-service
spec:
  targetRefs:
    - group: gateway.networking.k8s.io
      kind: HTTPRoute
      name: llm-route

  apiKeyAuth:
    credentialRefs:
    - group: ""
      kind: Secret
      name: mcp-bearer-token
    extractFrom:
    - headers:
      - authorization