disagg.yaml 2.12 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
spec:
  services:
    Frontend:
      dynamoNamespace: sglang-disagg
12
      componentType: frontend
13
14
15
      replicas: 1
      extraPodSpec:
        mainContainer:
julienmancuso's avatar
julienmancuso committed
16
          image: my-registry/sglang-runtime:my-tag
17
    decode:
18
19
20
      envFromSecret: hf-token-secret
      dynamoNamespace: sglang-disagg
      componentType: worker
21
      subComponentType: decode
22
23
24
25
26
27
      replicas: 1
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
julienmancuso's avatar
julienmancuso committed
28
          image: my-registry/sglang-runtime:my-tag
29
          workingDir: /workspace/components/backends/sglang
30
          command:
31
32
33
          - python3
          - -m
          - dynamo.sglang
34
          args:
35
36
37
38
39
40
41
42
43
44
45
46
47
48
            - --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
julienmancuso's avatar
julienmancuso committed
49

50
    prefill:
51
52
53
      envFromSecret: hf-token-secret
      dynamoNamespace: sglang-disagg
      componentType: worker
54
      subComponentType: prefill
55
56
57
58
59
60
      replicas: 1
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
julienmancuso's avatar
julienmancuso committed
61
          image: my-registry/sglang-runtime:my-tag
62
          workingDir: /workspace/components/backends/sglang
63
          command:
64
          - python3E
65
66
          - -m
          - dynamo.sglang
67
          args:
68
69
70
71
72
73
74
75
76
77
78
79
80
81
            - --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