disagg_planner.yaml 3.07 KB
Newer Older
1
2
3
4
5
6
7
8
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

apiVersion: nvidia.com/v1alpha1
kind: DynamoGraphDeployment
metadata:
  name: sglang-disagg-planner
spec:
9
10
11
  pvcs:
    - name: dynamo-pvc
      create: false # Must be pre-created before deployment and SLA profiler must have been run
12
13
14
  services:
    Frontend:
      dynamoNamespace: dynamo
15
      componentType: frontend
16
17
18
      replicas: 1
      extraPodSpec:
        mainContainer:
19
          image: my-registry/sglang-runtime:my-tag
20
21
22
23
24
    Planner:
      dynamoNamespace: dynamo
      envFromSecret: hf-token-secret
      componentType: planner
      replicas: 1
25
26
27
      volumeMounts:
        - name: dynamo-pvc
          mountPoint: /data
28
29
      extraPodSpec:
        mainContainer:
30
          image: my-registry/sglang-runtime:my-tag
31
          workingDir: /workspace/components/src/dynamo/planner
julienmancuso's avatar
julienmancuso committed
32
          command:
33
34
35
          - python3
          - -m
          - planner_sla
36
          args:
37
38
39
40
            - --environment=kubernetes
            - --backend=sglang
            - --adjustment-interval=60
            - --profile-results-dir=/data/profiling_results
41
    decode:
42
43
44
      dynamoNamespace: dynamo
      envFromSecret: hf-token-secret
      componentType: worker
45
      subComponentType: decode
46
47
48
49
50
51
      replicas: 2
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
52
          image: my-registry/sglang-runtime:my-tag
53
          workingDir: /workspace/components/backends/sglang
julienmancuso's avatar
julienmancuso committed
54
          command:
55
            - python3
56
          args:
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
            - -m
            - dynamo.sglang
            - --model-path
            - Qwen/Qwen3-0.6B
            - --served-model-name
            - Qwen/Qwen3-0.6B
            - --page-size
            - "16"
            - --tp
            - "1"
            - --trust-remote-code
            - --skip-tokenizer-init
            - --disaggregation-mode
            - decode
            - --disaggregation-transfer-backend
            - nixl
73
74
75
76
            - --disaggregation-bootstrap-port
            - "12345"
            - --host
            - "0.0.0.0"
77
    prefill:
78
79
80
      dynamoNamespace: dynamo
      envFromSecret: hf-token-secret
      componentType: worker
81
      subComponentType: prefill
82
83
84
85
86
87
      replicas: 2
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
88
          image: my-registry/sglang-runtime:my-tag
89
          workingDir: /workspace/components/backends/sglang
julienmancuso's avatar
julienmancuso committed
90
          command:
91
            - python3
92
          args:
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
            - -m
            - dynamo.sglang
            - --model-path
            - Qwen/Qwen3-0.6B
            - --served-model-name
            - Qwen/Qwen3-0.6B
            - --page-size
            - "16"
            - --tp
            - "1"
            - --trust-remote-code
            - --skip-tokenizer-init
            - --disaggregation-mode
            - prefill
            - --disaggregation-transfer-backend
            - nixl
109
110
111
112
            - --disaggregation-bootstrap-port
            - "12345"
            - --host
            - "0.0.0.0"