disagg_planner.yaml 2.91 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
# 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:
  envs:
    - name: DYNAMO_NAMESPACE
      value: "dynamo"
12
13
14
  pvcs:
    - name: dynamo-pvc
      create: false # Must be pre-created before deployment and SLA profiler must have been run
15
16
17
  services:
    Frontend:
      dynamoNamespace: dynamo
18
      componentType: frontend
19
20
21
      replicas: 1
      extraPodSpec:
        mainContainer:
22
          image: my-registry/sglang-runtime:my-tag
23
24
25
26
27
    Planner:
      dynamoNamespace: dynamo
      envFromSecret: hf-token-secret
      componentType: planner
      replicas: 1
28
29
30
      volumeMounts:
        - name: dynamo-pvc
          mountPoint: /data
31
32
      extraPodSpec:
        mainContainer:
33
          image: my-registry/sglang-runtime:my-tag
34
          workingDir: /workspace/components/src/dynamo/planner
julienmancuso's avatar
julienmancuso committed
35
          command:
36
37
38
          - python3
          - -m
          - planner_sla
39
          args:
40
41
42
43
            - --environment=kubernetes
            - --backend=sglang
            - --adjustment-interval=60
            - --profile-results-dir=/data/profiling_results
44
    decode:
45
46
47
      dynamoNamespace: dynamo
      envFromSecret: hf-token-secret
      componentType: worker
48
      subComponentType: decode
49
50
51
52
53
54
      replicas: 2
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
55
          image: my-registry/sglang-runtime:my-tag
56
          workingDir: /workspace/components/backends/sglang
julienmancuso's avatar
julienmancuso committed
57
          command:
58
            - python3
59
          args:
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
            - -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
76
    prefill:
77
78
79
      dynamoNamespace: dynamo
      envFromSecret: hf-token-secret
      componentType: worker
80
      subComponentType: prefill
81
82
83
84
85
86
      replicas: 2
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
87
          image: my-registry/sglang-runtime:my-tag
88
          workingDir: /workspace/components/backends/sglang
julienmancuso's avatar
julienmancuso committed
89
          command:
90
            - python3
91
          args:
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
            - -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