# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
apiVersion: nvidia.com/v1alpha1
kind: DynamoGraphDeployment
metadata:
  name: qwen
spec:
  backendFramework: vllm
  services:
    Epp:
      envFromSecret: hf-token-secret
      componentType: epp
      replicas: 1
      extraPodSpec:
        mainContainer:
          image: nvcr.io/nvidia/ai-dynamo/epp-image:my-tag
          imagePullPolicy: IfNotPresent
          env:
            - name: DYN_KV_CACHE_BLOCK_SIZE
              value: "16"
            - name: DYN_MODEL_NAME
              value: "Qwen/Qwen3-0.6B"
            - name: DYN_ENFORCE_DISAGG
              value: "true"
      eppConfig:
        config:
          plugins:
            - type: disagg-profile-handler
            - name: prefill-filter
              type: label-filter
              parameters:
                label: "nvidia.com/dynamo-sub-component-type"
                validValues:
                  - "prefill"
                allowsNoLabel: false
            - name: decode-filter
              type: label-filter
              parameters:
                label: "nvidia.com/dynamo-sub-component-type"
                validValues:
                  - "decode"
                allowsNoLabel: false
            - name: picker
              type: max-score-picker
            - name: dyn-prefill
              type: dyn-prefill-scorer
            - name: dyn-decode
              type: dyn-decode-scorer
          schedulingProfiles:
          - name: prefill
            plugins:
            - pluginRef: prefill-filter
              weight: 1
            - pluginRef: dyn-prefill
              weight: 1
            - pluginRef: picker
              weight: 1
          - name: decode
            plugins:
            - pluginRef: decode-filter
              weight: 1
            - pluginRef: dyn-decode
              weight: 1
            - pluginRef: picker
              weight: 1
    VllmPrefillWorker:
      componentType: worker
      subComponentType: prefill
      envFromSecret: hf-token-secret
      sharedMemory:
        size: 2Gi
      frontendSidecar:
        image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
        args:
          - -m
          - dynamo.frontend
          - --router-mode
          - direct
        envFromSecret: hf-token-secret
      extraPodSpec:
        tolerations:
          - key: "nvidia.com/gpu"
            operator: "Exists"
            effect: "NoSchedule"
        affinity:
          podAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchExpressions:
                - key: nvidia.com/dynamo-component-type
                  operator: In
                  values:
                  - worker
              topologyKey: kubernetes.io/hostname
        mainContainer:
          env:
            - name: SERVED_MODEL_NAME
              value: "Qwen/Qwen3-0.6B"
            - name: MODEL_PATH
              value: "Qwen/Qwen3-0.6B"
            - name: UCX_TLS
              value: "tcp,self"
          args:
          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 1 --data-parallel-size 1 --disaggregation-mode prefill --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\"}' --gpu-memory-utilization 0.90 --enable-prefix-caching --block-size 16 --kv-events-config '{\"enable_kv_cache_events\":true}'"
          command:
          - /bin/sh
          - -c
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
          imagePullPolicy: IfNotPresent
          workingDir: /workspace/examples/backends/vllm
      replicas: 1
      resources:
        limits:
          gpu: "1"
        requests:
          gpu: "1"
    VllmDecodeWorker:
      componentType: worker
      subComponentType: decode
      envFromSecret: hf-token-secret
      sharedMemory:
        size: 2Gi
      frontendSidecar:
        image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
        args:
          - -m
          - dynamo.frontend
          - --router-mode
          - direct
        envFromSecret: hf-token-secret
      extraPodSpec:
        tolerations:
          - key: "nvidia.com/gpu"
            operator: "Exists"
            effect: "NoSchedule"
        affinity:
          podAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchExpressions:
                - key: nvidia.com/dynamo-component-type
                  operator: In
                  values:
                  - worker
              topologyKey: kubernetes.io/hostname
        mainContainer:
          env:
            - name: SERVED_MODEL_NAME
              value: "Qwen/Qwen3-0.6B"
            - name: MODEL_PATH
              value: "Qwen/Qwen3-0.6B"
            - name: UCX_TLS
              value: "tcp,self"
          args:
          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 1 --data-parallel-size 1 --disaggregation-mode decode --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\"}' --gpu-memory-utilization 0.90 --block-size 16"
          command:
          - /bin/sh
          - -c
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
          imagePullPolicy: IfNotPresent
          workingDir: /workspace/examples/backends/vllm
      replicas: 1
      resources:
        limits:
          gpu: "1"
        requests:
          gpu: "1"