deploy.yaml

# 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: llama3-70b-disagg
spec:
  backendFramework: vllm
  pvcs:
    - name: model-cache
      create: false
  services:
    Epp:
      envFromSecret: hf-token-secret
      componentType: epp
      replicas: 1
      extraPodSpec:
        mainContainer:
          image: nvcr.io/nvidia/ai-dynamo/epp-image:1.0.0
      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
      volumeMounts:
        - name: model-cache
          mountPoint: /opt/models
      sharedMemory:
        size: 80Gi
      frontendSidecar:
        image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.0.0
        args:
          - -m
          - dynamo.frontend
          - --router-mode
          - direct
        envFromSecret: hf-token-secret
      extraPodSpec:
        affinity:
          podAffinity:
            preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 100
              podAffinityTerm:
                labelSelector:
                  matchExpressions:
                  - key: nvidia.com/dynamo-component-type
                    operator: In
                    values:
                    - worker
                topologyKey: kubernetes.io/hostname
        mainContainer:
          env:
            - name: SERVED_MODEL_NAME
              value: "RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic"
            - name: MODEL_PATH
              value: "RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic"
            - name: HF_HOME
              value: /opt/models
          args:
          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 2 --data-parallel-size 1 --is-prefill-worker --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
          command:
          - /bin/sh
          - -c
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.0.0
          workingDir: /workspace/examples/backends/vllm
      replicas: 2
      resources:
        limits:
          gpu: "2"
        requests:
          gpu: "2"
    VllmDecodeWorker:
      componentType: worker
      subComponentType: decode
      envFromSecret: hf-token-secret
      volumeMounts:
        - name: model-cache
          mountPoint: /opt/models
      sharedMemory:
        size: 80Gi
      frontendSidecar:
        image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.0.0
        args:
          - -m
          - dynamo.frontend
          - --router-mode
          - direct
        envFromSecret: hf-token-secret
      extraPodSpec:
        affinity:
          podAffinity:
            preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 100
              podAffinityTerm:
                labelSelector:
                  matchExpressions:
                  - key: nvidia.com/dynamo-component-type
                    operator: In
                    values:
                    - worker
                topologyKey: kubernetes.io/hostname
        mainContainer:
          env:
            - name: SERVED_MODEL_NAME
              value: "RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic"
            - name: MODEL_PATH
              value: "RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic"
            - name: HF_HOME
              value: /opt/models
          args:
          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 4 --data-parallel-size 1 --gpu-memory-utilization 0.90 --no-enable-prefix-caching --block-size 128"
          command:
          - /bin/sh
          - -c
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.0.0
          workingDir: /workspace/examples/backends/vllm
      replicas: 1
      resources:
        limits:
          gpu: "4"
        requests:
          gpu: "4"