disagg_planner.yaml 4.88 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-planner
8
9
  annotations:
    nvidia.com/enable-grove: "false" # temporarily disable grove because current k8s connector does not work with grove
10
spec:
11
12
13
14
15
  envs:
    - name: DYNAMO_SERVICE_CONFIG
      value: '{"Prometheus":{"global":{"scrape_interval":"5s"},"scrape_configs":[{"job_name":"prometheus","static_configs":[{"targets":["localhost:9090"]}]},{"job_name":"frontend","static_configs":[{"targets":["vllm-disagg-planner-frontend:8000"]}]}]}}'
    - name: DYNAMO_NAMESPACE
      value: "vllm-disagg-planner"
16
17
  services:
    Frontend:
18
      dynamoNamespace: vllm-disagg-planner
19
      componentType: frontend
20
      replicas: 1
21
22
      extraPodSpec:
        mainContainer:
23
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:hzhou-0814-02
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
          workingDir: /workspace/components/backends/vllm
          command:
            - /bin/sh
            - -c
          args:
            - "python3 -m dynamo.frontend --http-port 8000"
    Planner:
      dynamoNamespace: vllm-disagg-planner
      envFromSecret: hf-token-secret
      componentType: planner
      replicas: 1
      livenessProbe:
        exec:
          command:
            - /bin/sh
            - -c
            - "exit 0"
        periodSeconds: 60
        timeoutSeconds: 30
        failureThreshold: 10
      readinessProbe:
        exec:
          command:
            - /bin/sh
            - -c
            - "exit 0"
        initialDelaySeconds: 60
        periodSeconds: 60
        timeoutSeconds: 30
        failureThreshold: 10
      pvc:
        create: false
        name: profiling-pvc # Must be pre-created before deployment and SLA profiler must have been run
        mountPoint: /workspace/profiling_results
58
59
      extraPodSpec:
        mainContainer:
60
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:hzhou-0814-02
61
          workingDir: /workspace/components/planner/src/dynamo/planner
62
63
64
          ports:
            - name: metrics
              containerPort: 9085
julienmancuso's avatar
julienmancuso committed
65
66
67
          command:
            - /bin/sh
            - -c
68
          args:
julienmancuso's avatar
julienmancuso committed
69
70
71
72
73
74
            - >-
              python3 -m planner_sla
              --environment=kubernetes
              --backend=vllm
              --adjustment-interval=60
              --profile-results-dir=/workspace/profiling_results
75
              --prometheus-port=9085
76
    Prometheus: # NOTE: this is set on Prometheus to ensure a service is created for the Prometheus component. This is a workaround and should be managed differently.
77
      dynamoNamespace: vllm-disagg-planner
78
      componentType: frontend
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
      replicas: 1
      envs:
        - name: PYTHONPATH
          value: "/workspace/components/planner/src"
      livenessProbe:
        exec:
          command:
            - /bin/sh
            - -c
            - "exit 0"
        periodSeconds: 60
        timeoutSeconds: 30
        failureThreshold: 10
      readinessProbe:
        exec:
          command:
            - /bin/sh
            - -c
            - "exit 0"
        initialDelaySeconds: 30
        periodSeconds: 60
        timeoutSeconds: 30
        failureThreshold: 10
      extraPodSpec:
        mainContainer:
104
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:hzhou-0814-02
Alec's avatar
Alec committed
105
          workingDir: /workspace/components/backends/vllm
106
107
108
          command:
            - /bin/sh
            - -c
109
          args:
110
            - "python3 -m dynamo.planner.prometheus"
111
    VllmDecodeWorker:
112
      dynamoNamespace: vllm-disagg-planner
113
      envFromSecret: hf-token-secret
114
      componentType: worker
115
      replicas: 2
116
117
      resources:
        limits:
118
          gpu: "1"
119
120
      extraPodSpec:
        mainContainer:
121
122
123
124
125
126
          startupProbe:
            httpGet:
              path: /health
              port: 9090
            periodSeconds: 10
            failureThreshold: 60
127
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:hzhou-0814-02
Alec's avatar
Alec committed
128
          workingDir: /workspace/components/backends/vllm
129
130
131
          command:
            - /bin/sh
            - -c
132
          args:
133
            - "python3 -m dynamo.vllm --model Qwen/Qwen3-0.6B --migration-limit=3"
134
    VllmPrefillWorker:
135
      dynamoNamespace: vllm-disagg-planner
136
      envFromSecret: hf-token-secret
137
      componentType: worker
138
      replicas: 2
139
140
      resources:
        limits:
141
          gpu: "1"
142
143
      extraPodSpec:
        mainContainer:
144
145
146
147
148
149
          startupProbe:
            httpGet:
              path: /health
              port: 9090
            periodSeconds: 10
            failureThreshold: 60
150
          image: nvcr.io/nvidian/nim-llm-dev/vllm-runtime:hzhou-0814-02
Alec's avatar
Alec committed
151
          workingDir: /workspace/components/backends/vllm
152
153
154
          command:
            - /bin/sh
            - -c
155
          args:
156
            - python3 -m dynamo.vllm --model Qwen/Qwen3-0.6B --is-prefill-worker --migration-limit=3