# OpenShift deployment for the Shield GUARDRAIL DATA PLANE with in-cluster vLLM (GPU).
#
# This is a thin wrapper around the existing image — the image's entrypoint
# (scripts/start_vllm.sh) already launches vLLM on :8000, waits for it, then runs
# the guard API (handler.py -> uvicorn). This manifest only schedules that image
# on a GPU, gives it the volumes/Route it needs, and sets its env. No vLLM logic
# is reimplemented here.
#
# Model runs IN the cluster — nothing leaves for inference (full data residency).
# For the lighter "app calls a remote model" option, use the llm-shield-cloud
# image with SKIP_VLLM=true + LLM_BACKEND_URL instead (no GPU); see
# docs/partners/rafay/rafay-integration.md Appendix A.
#
# PREREQUISITES
#   - NVIDIA GPU Operator installed (provides nvidia.com/gpu, drivers, device plugin).
#   - A GPU node pool. Default below targets an fp8-capable card (L40S / L4 / H100).
#     On a GPU WITHOUT fp8 (A100, V100, T4), set VLLM_QUANTIZATION=none and
#     VLLM_KV_CACHE_DTYPE=none below (bf16 uses more VRAM).
#   - Redis for the tenant store: oc apply -f openshift/redis.yaml
#   - If votal-ai/vai35-4B-v2 is a private HF repo, set HUGGING_FACE_HUB_TOKEN
#     in the Secret below.
#
# APPLY
#   oc new-project shield 2>/dev/null || oc project shield
#   oc apply -f openshift/redis.yaml
#   # put real values in the Secret first (do NOT commit real keys), then:
#   oc apply -f openshift/shield-guardrail-vllm.yaml
#   oc get route shield-guardrail -o jsonpath='{.spec.host}'   # base URL for callers
#
# SCC NOTE: the vllm/vllm-openai base may expect to run as root. If the pod
# CrashLoops under the default restricted-v2 SCC, grant anyuid to its SA:
#   oc adm policy add-scc-to-user anyuid -z shield-guardrail -n shield
---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: shield-guardrail
  namespace: shield
---
apiVersion: v1
kind: Secret
metadata:
  name: shield-guardrail
  namespace: shield
stringData:
  # Tenant store / policy cache. Point at openshift/redis.yaml's Service.
  REDIS_URL: "redis://redis:6379/0"
  # Authorizes the tenant-provisioning endpoints (see integration guide §4a).
  SHIELD_ADMIN_KEY: "REPLACE_ME"
  # Only if the model repo is private/gated on Hugging Face.
  HUGGING_FACE_HUB_TOKEN: ""
---
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: shield-model-cache
  namespace: shield
spec:
  accessModes: ["ReadWriteOnce"]
  resources:
    requests:
      storage: 50Gi          # HF model cache, so the 4B model isn't re-pulled on restart
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: shield-guardrail
  namespace: shield
  labels: { app: shield-guardrail }
spec:
  replicas: 1                # bounded by available GPUs; scale with more GPU nodes
  selector:
    matchLabels: { app: shield-guardrail }
  template:
    metadata:
      labels: { app: shield-guardrail }
    spec:
      serviceAccountName: shield-guardrail
      # Schedule onto a GPU node and tolerate the GPU Operator's taint.
      tolerations:
        - key: nvidia.com/gpu
          operator: Exists
          effect: NoSchedule
      containers:
        - name: shield
          image: docker.io/sundi133/llm-shield:latest   # the "vllm" image (app + vLLM); pin a tag for prod
          imagePullPolicy: IfNotPresent
          ports:
            - { name: http, containerPort: 8080 }        # guard API (8000 vLLM is pod-internal)
          env:
            # --- guard API ---
            - { name: PORT, value: "8080" }               # OpenShift restricted SCC can't bind :80
            - { name: SHIELD_GUARD_REQUIRE_KEY, value: "enforce" }  # REQUIRED for tenant isolation
            - { name: REDIS_URL,        valueFrom: { secretKeyRef: { name: shield-guardrail, key: REDIS_URL } } }
            - { name: SHIELD_ADMIN_KEY, valueFrom: { secretKeyRef: { name: shield-guardrail, key: SHIELD_ADMIN_KEY } } }
            # --- vLLM (read by scripts/start_vllm.sh) ---
            - { name: MODEL_NAME,   value: "votal-ai/vai35-4B-v2" }   # or nvidia/Nemotron-3.5-Content-Safety (pin VLLM_BASE_IMAGE at build)
            - { name: VLLM_PORT,    value: "8000" }
            - { name: MAX_MODEL_LEN, value: "8196" }
            - { name: GPU_MEM_UTIL, value: "0.85" }
            - { name: VLLM_QUANTIZATION,   value: "fp8" }   # set "none" on GPUs without fp8 (A100/V100/T4)
            - { name: VLLM_KV_CACHE_DTYPE, value: "fp8" }   # set "none" to match
            - { name: HUGGING_FACE_HUB_TOKEN, valueFrom: { secretKeyRef: { name: shield-guardrail, key: HUGGING_FACE_HUB_TOKEN, optional: true } } }
            - { name: HF_HOME, value: "/model-cache" }       # persist the model download on the PVC
          volumeMounts:
            - { name: dshm,        mountPath: /dev/shm }     # vLLM/NCCL need a large /dev/shm
            - { name: model-cache, mountPath: /model-cache }
          resources:
            requests: { cpu: "2", memory: "16Gi", nvidia.com/gpu: "1" }
            limits:   { cpu: "8", memory: "32Gi", nvidia.com/gpu: "1" }
          # Model load takes minutes — a startupProbe keeps it from being killed
          # while vLLM downloads/loads, then readiness/liveness take over.
          startupProbe:
            httpGet: { path: /health, port: 8080 }
            periodSeconds: 10
            failureThreshold: 60          # ~10 min to become ready
          readinessProbe:
            httpGet: { path: /health, port: 8080 }
            periodSeconds: 10
          livenessProbe:
            httpGet: { path: /health, port: 8080 }
            periodSeconds: 20
            failureThreshold: 6
      volumes:
        - name: dshm
          emptyDir: { medium: Memory, sizeLimit: 16Gi }
        - name: model-cache
          persistentVolumeClaim: { claimName: shield-model-cache }
---
apiVersion: v1
kind: Service
metadata:
  name: shield-guardrail
  namespace: shield
spec:
  selector: { app: shield-guardrail }
  ports:
    - { name: http, port: 80, targetPort: 8080 }
---
apiVersion: route.openshift.io/v1
kind: Route
metadata:
  name: shield-guardrail
  namespace: shield
  annotations:
    # Guardrail calls (esp. Tier-2/agentic) can exceed the 30s default.
    haproxy.router.openshift.io/timeout: 120s
spec:
  to: { kind: Service, name: shield-guardrail }
  port: { targetPort: http }
  tls: { termination: edge, insecureEdgeTerminationPolicy: Redirect }