apiVersion: v1
kind: Namespace
metadata:
  name: ai-stack
  labels:
    app.kubernetes.io/part-of: spacedock-ai-stack
---
# ==============================================================================
# 1. QDRANT VECTOR DATABASE (Backed by 46.4 TiB Longhorn Flash Pool)
# ==============================================================================
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: qdrant-storage-pvc
  namespace: ai-stack
spec:
  accessModes:
    - ReadWriteOnce
  storageClassName: longhorn
  resources:
    requests:
      storage: 100Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: qdrant
  namespace: ai-stack
  labels:
    app: qdrant
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: qdrant
  template:
    metadata:
      labels:
        app: qdrant
    spec:
      containers:
        - name: qdrant
          image: qdrant/qdrant:latest
          imagePullPolicy: IfNotPresent
          ports:
            - name: http
              containerPort: 6333
            - name: grpc
              containerPort: 6334
          resources:
            requests:
              cpu: "2"
              memory: "4Gi"
            limits:
              cpu: "8"
              memory: "16Gi"
          readinessProbe:
            httpGet:
              path: /readyz
              port: 6333
            initialDelaySeconds: 5
            periodSeconds: 10
          volumeMounts:
            - name: qdrant-storage
              mountPath: /qdrant/storage
      volumes:
        - name: qdrant-storage
          persistentVolumeClaim:
            claimName: qdrant-storage-pvc
---
apiVersion: v1
kind: Service
metadata:
  name: qdrant
  namespace: ai-stack
  labels:
    app: qdrant
spec:
  type: NodePort
  selector:
    app: qdrant
  ports:
    - name: http
      port: 6333
      targetPort: 6333
      nodePort: 30633
    - name: grpc
      port: 6334
      targetPort: 6334
      nodePort: 30634
---
# ==============================================================================
# 2. vLLM HIGH-THROUGHPUT LLM INFERENCE ENGINE (Tesla T4 GPU + Longhorn Cache)
# ==============================================================================
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: vllm-models-pvc
  namespace: ai-stack
spec:
  accessModes:
    - ReadWriteOnce
  storageClassName: longhorn
  resources:
    requests:
      storage: 200Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-openai
  namespace: ai-stack
  labels:
    app: vllm-openai
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: vllm-openai
  template:
    metadata:
      labels:
        app: vllm-openai
    spec:
      nodeSelector:
        kubernetes.io/hostname: snaisvr44hrvstr04
      containers:
        - name: vllm
          image: vllm/vllm-openai:v0.7.3
          imagePullPolicy: IfNotPresent
          args:
            - "--model"
            - "Qwen/Qwen2.5-7B-Instruct-AWQ"
            - "--served-model-name"
            - "Qwen2.5-7B-Instruct"
            - "--quantization"
            - "awq"
            - "--dtype"
            - "half"
            - "--max-model-len"
            - "65536"
            - "--gpu-memory-utilization"
            - "0.90"
            - "--enable-auto-tool-choice"
            - "--tool-call-parser"
            - "hermes"
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8000"
          env:
            - name: HF_HOME
              value: "/root/.cache/huggingface"
            - name: VLLM_WORKER_MULTIPROC_METHOD
              value: "spawn"
            - name: VLLM_ALLOW_LONG_MAX_MODEL_LEN
              value: "1"
          ports:
            - name: http-openai
              containerPort: 8000
          resources:
            requests:
              cpu: "8"
              memory: "24Gi"
              nvidia.com/gpu: "1"
            limits:
              cpu: "24"
              memory: "48Gi"
              nvidia.com/gpu: "1"
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 30
            periodSeconds: 10
          volumeMounts:
            - name: vllm-models
              mountPath: /root/.cache/huggingface
            - name: dshm
              mountPath: /dev/shm
      volumes:
        - name: vllm-models
          persistentVolumeClaim:
            claimName: vllm-models-pvc
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 16Gi
---
apiVersion: v1
kind: Service
metadata:
  name: vllm-openai
  namespace: ai-stack
  labels:
    app: vllm-openai
spec:
  type: NodePort
  selector:
    app: vllm-openai
  ports:
    - name: http-openai
      port: 8000
      targetPort: 8000
      nodePort: 30800
---
# ==============================================================================
# 2B. vLLM GEMMA 4 12B INFERENCE ENGINE (snaisvr44hrvstr02 Tesla T4 GPU)
# ==============================================================================
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: vllm-gemma4-models-pvc
  namespace: ai-stack
spec:
  accessModes:
    - ReadWriteOnce
  storageClassName: longhorn
  resources:
    requests:
      storage: 200Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-gemma4
  namespace: ai-stack
  labels:
    app: vllm-gemma4
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: vllm-gemma4
  template:
    metadata:
      labels:
        app: vllm-gemma4
    spec:
      nodeSelector:
        kubernetes.io/hostname: snaisvr44hrvstr02
      containers:
        - name: vllm
          image: vllm/vllm-openai:v0.32.0-cu129
          imagePullPolicy: IfNotPresent
          command: ["/bin/bash", "-c"]
          args:
            - |
              rm -rf /usr/local/lib/python3.12/dist-packages/torchcodec*
              python3 -c '
              p = "/usr/local/lib/python3.12/dist-packages/vllm/v1/attention/ops/triton_unified_attention.py"
              s = open(p).read()
              s = s.replace("return 32", "return 16")
              s = s.replace("launch_num_stages: int | None = None", "launch_num_stages: int | None = 1")
              open(p, "w").write(s)
              '
              exec vllm serve google/gemma-4-12B-it-qat-w4a16-ct \
                --served-model-name Gemma-4-12B-it \
                --dtype half \
                --max-model-len 8192 \
                --gpu-memory-utilization 0.92 \
                --enforce-eager \
                --enable-auto-tool-choice \
                --tool-call-parser gemma4 \
                --reasoning-parser gemma4 \
                --host 0.0.0.0 \
                --port 8000
          env:
            - name: HF_HOME
              value: "/root/.cache/huggingface"
            - name: VLLM_WORKER_MULTIPROC_METHOD
              value: "spawn"
            - name: VLLM_ALLOW_LONG_MAX_MODEL_LEN
              value: "1"
          ports:
            - name: http-openai
              containerPort: 8000
          resources:
            requests:
              cpu: "8"
              memory: "24Gi"
              nvidia.com/gpu: "1"
            limits:
              cpu: "24"
              memory: "48Gi"
              nvidia.com/gpu: "1"
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 30
            periodSeconds: 10
          volumeMounts:
            - name: vllm-gemma4-models
              mountPath: /root/.cache/huggingface
            - name: dshm
              mountPath: /dev/shm
      volumes:
        - name: vllm-gemma4-models
          persistentVolumeClaim:
            claimName: vllm-gemma4-models-pvc
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 16Gi
---
apiVersion: v1
kind: Service
metadata:
  name: vllm-gemma4
  namespace: ai-stack
  labels:
    app: vllm-gemma4
spec:
  type: NodePort
  selector:
    app: vllm-gemma4
  ports:
    - name: http-openai
      port: 8000
      targetPort: 8000
      nodePort: 30804
---
# ==============================================================================
# 2C. DEEPSEEK-R1-DISTILL-QWEN-14B REASONING ENGINE (snaisvr44hrvstr01 Tesla T4 GPU)
# ==============================================================================
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: glm-superexpert-models-pvc
  namespace: ai-stack
spec:
  accessModes:
    - ReadWriteOnce
  storageClassName: longhorn
  resources:
    requests:
      storage: 200Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: glm-superexpert
  namespace: ai-stack
  labels:
    app: glm-superexpert
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: glm-superexpert
  template:
    metadata:
      labels:
        app: glm-superexpert
    spec:
      nodeSelector:
        kubernetes.io/hostname: snaisvr44hrvstr01
      containers:
        - name: vllm
          image: vllm/vllm-openai:v0.32.0-cu129
          imagePullPolicy: IfNotPresent
          command: ["/bin/bash", "-c"]
          args:
            - |
              rm -rf /usr/local/lib/python3.12/dist-packages/torchcodec*
              exec vllm serve casperhansen/deepseek-r1-distill-qwen-14b-awq \
                --served-model-name DeepSeek-R1-Distill-Qwen-14B \
                --quantization awq \
                --dtype half \
                --max-model-len 16384 \
                --gpu-memory-utilization 0.92 \
                --enforce-eager \
                --enable-auto-tool-choice \
                --tool-call-parser hermes \
                --reasoning-parser deepseek_r1 \
                --host 0.0.0.0 \
                --port 8000
          env:
            - name: HF_HOME
              value: "/root/.cache/huggingface"
          ports:
            - name: http-openai
              containerPort: 8000
          resources:
            requests:
              cpu: "8"
              memory: "20Gi"
              nvidia.com/gpu: "1"
            limits:
              cpu: "24"
              memory: "40Gi"
              nvidia.com/gpu: "1"
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 30
            periodSeconds: 10
          volumeMounts:
            - name: glm-models
              mountPath: /root/.cache/huggingface
            - name: dshm
              mountPath: /dev/shm
      volumes:
        - name: glm-models
          persistentVolumeClaim:
            claimName: glm-superexpert-models-pvc
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 16Gi
---
apiVersion: v1
kind: Service
metadata:
  name: glm-superexpert
  namespace: ai-stack
  labels:
    app: glm-superexpert
spec:
  type: NodePort
  selector:
    app: glm-superexpert
  ports:
    - name: http-openai
      port: 8000
      targetPort: 8000
      nodePort: 30805
---
# ==============================================================================
# 3. NVIDIA TRITON INFERENCE SERVER (Decision Engine & Multi-Model ML Serving)
# ==============================================================================
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: triton-models-pvc
  namespace: ai-stack
spec:
  accessModes:
    - ReadWriteOnce
  storageClassName: longhorn
  resources:
    requests:
      storage: 200Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: triton-inference-server
  namespace: ai-stack
  labels:
    app: triton-inference-server
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: triton-inference-server
  template:
    metadata:
      labels:
        app: triton-inference-server
    spec:
      initContainers:
        - name: seed-decision-engine-model
          image: busybox:1.36
          command: ["/bin/sh", "-c"]
          args:
            - |
              mkdir -p /models/repository/decision_engine/1
              cat << 'EOF' > /models/repository/decision_engine/config.pbtxt
              name: "decision_engine"
              backend: "python"
              max_batch_size: 64
              input [
                {
                  name: "FEATURES"
                  data_type: TYPE_FP32
                  dims: [ -1 ]
                }
              ]
              output [
                {
                  name: "DECISION_SCORE"
                  data_type: TYPE_FP32
                  dims: [ 1 ]
                },
                {
                  name: "ACTION_LABEL"
                  data_type: TYPE_STRING
                  dims: [ 1 ]
                }
              ]
              instance_group [
                {
                  count: 1
                  kind: KIND_GPU
                }
              ]
              EOF
              cat << 'EOF' > /models/repository/decision_engine/1/model.py
              import numpy as np
              import triton_python_backend_utils as pb_utils

              class TritonPythonModel:
                  def initialize(self, args):
                      pass

                  def execute(self, requests):
                      responses = []
                      for request in requests:
                          features = pb_utils.get_input_tensor_by_name(request, "FEATURES").as_numpy()
                          # Sigmoid risk/decision scoring over feature vector
                          mean_val = float(np.mean(features))
                          score = float(1.0 / (1.0 + np.exp(-mean_val)))
                          label = "APPROVE_EXECUTE" if score >= 0.5 else "HOLD_REVIEW"
                          out_score = pb_utils.Tensor("DECISION_SCORE", np.array([[score]], dtype=np.float32))
                          out_label = pb_utils.Tensor("ACTION_LABEL", np.array([[label.encode("utf-8")]], dtype=np.object_))
                          responses.append(pb_utils.InferenceResponse(output_tensors=[out_score, out_label]))
                      return responses

                  def finalize(self):
                      pass
              EOF
              echo "Seeded initial decision_engine model into /models/repository/decision_engine"
          volumeMounts:
            - name: triton-models
              mountPath: /models
      containers:
        - name: triton
          image: nvcr.io/nvidia/tritonserver:24.08-py3
          imagePullPolicy: IfNotPresent
          command: ["tritonserver"]
          args:
            - "--model-repository=/models/repository"
            - "--model-control-mode=poll"
            - "--repository-poll-secs=10"
            - "--strict-model-config=false"
            - "--allow-http=true"
            - "--allow-grpc=true"
            - "--allow-metrics=true"
          ports:
            - name: http
              containerPort: 8000
            - name: grpc
              containerPort: 8001
            - name: metrics
              containerPort: 8002
          resources:
            requests:
              cpu: "4"
              memory: "16Gi"
              nvidia.com/gpu: "1"
            limits:
              cpu: "16"
              memory: "32Gi"
              nvidia.com/gpu: "1"
          readinessProbe:
            httpGet:
              path: /v2/health/ready
              port: 8000
            initialDelaySeconds: 15
            periodSeconds: 10
          volumeMounts:
            - name: triton-models
              mountPath: /models
            - name: dshm
              mountPath: /dev/shm
      volumes:
        - name: triton-models
          persistentVolumeClaim:
            claimName: triton-models-pvc
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 8Gi
---
apiVersion: v1
kind: Service
metadata:
  name: triton-inference-server
  namespace: ai-stack
  labels:
    app: triton-inference-server
spec:
  type: NodePort
  selector:
    app: triton-inference-server
  ports:
    - name: http
      port: 8000
      targetPort: 8000
      nodePort: 30801
    - name: grpc
      port: 8001
      targetPort: 8001
      nodePort: 30802
    - name: metrics
      port: 8002
      targetPort: 8002
      nodePort: 30803