disagg.yaml 1.45 KB
Newer Older
1
2
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
Alec's avatar
Alec committed
3

4
5
6
apiVersion: nvidia.com/v1alpha1
kind: DynamoGraphDeployment
metadata:
7
  name: vllm-disagg
8
9
10
spec:
  services:
    Frontend:
11
      dynamoNamespace: vllm-disagg
12
      componentType: frontend
13
14
15
      replicas: 1
      extraPodSpec:
        mainContainer:
16
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
17
    VllmDecodeWorker:
18
      dynamoNamespace: vllm-disagg
19
      envFromSecret: hf-token-secret
20
      componentType: worker
21
22
23
      replicas: 1
      resources:
        limits:
24
          gpu: "1"
25
26
      extraPodSpec:
        mainContainer:
27
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
Alec's avatar
Alec committed
28
          workingDir: /workspace/components/backends/vllm
29
          command:
30
31
32
          - python3
          - -m
          - dynamo.vllm
33
          args:
34
35
            - --model
            - Qwen/Qwen3-0.6B
36
    VllmPrefillWorker:
37
      dynamoNamespace: vllm-disagg
38
      envFromSecret: hf-token-secret
39
      componentType: worker
40
41
42
      replicas: 1
      resources:
        limits:
43
          gpu: "1"
44
45
      extraPodSpec:
        mainContainer:
46
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:my-tag
Alec's avatar
Alec committed
47
          workingDir: /workspace/components/backends/vllm
48
          command:
49
50
51
          - python3
          - -m
          - dynamo.vllm
52
          args:
53
54
55
            - --model
            - Qwen/Qwen3-0.6B
            - --is-prefill-worker