"ml/vscode:/vscode.git/clone" did not exist on "73d6a82cce18f84ff5c67148783224cf25b30b32"
disagg.yaml 1.63 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/nvidian/nim-llm-dev/vllm-runtime:dep-233.17
Alec's avatar
Alec committed
17
          workingDir: /workspace/components/backends/vllm
18
19
20
          command:
            - /bin/sh
            - -c
21
          args:
22
            - "python3 -m dynamo.frontend --http-port 8000"
23
    VllmDecodeWorker:
24
      dynamoNamespace: vllm-disagg
25
      envFromSecret: hf-token-secret
26
      componentType: worker
27
28
29
      replicas: 1
      resources:
        limits:
30
          gpu: "1"
31
32
      extraPodSpec:
        mainContainer:
33
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:dep-233.17
Alec's avatar
Alec committed
34
          workingDir: /workspace/components/backends/vllm
35
36
37
          command:
            - /bin/sh
            - -c
38
          args:
39
            - "python3 -m dynamo.vllm --model Qwen/Qwen3-0.6B"
40
    VllmPrefillWorker:
41
      dynamoNamespace: vllm-disagg
42
      envFromSecret: hf-token-secret
43
      componentType: worker
44
45
46
      replicas: 1
      resources:
        limits:
47
          gpu: "1"
48
49
      extraPodSpec:
        mainContainer:
50
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:dep-233.17
Alec's avatar
Alec committed
51
          workingDir: /workspace/components/backends/vllm
52
53
54
          command:
            - /bin/sh
            - -c
55
          args:
56
            - "python3 -m dynamo.vllm --model Qwen/Qwen3-0.6B --is-prefill-worker"