disagg.yaml 1.84 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: trtllm-disagg
spec:
  services:
    Frontend:
      dynamoNamespace: trtllm-disagg
12
      componentType: frontend
13
14
15
      replicas: 1
      extraPodSpec:
        mainContainer:
16
          image: my-registry/tensorrtllm-runtime:my-tag
17
18
19
20
    TRTLLMPrefillWorker:
      dynamoNamespace: trtllm-disagg
      envFromSecret: hf-token-secret
      componentType: worker
21
      subComponentType: prefill
22
23
24
25
26
27
      replicas: 1
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
28
          image: my-registry/tensorrtllm-runtime:my-tag
29
          workingDir: /workspace/
30
          command:
31
32
33
          - python3
          - -m
          - dynamo.trtllm
34
          args:
35
36
37
38
39
            - --model-path
            - Qwen/Qwen3-0.6B
            - --served-model-name
            - Qwen/Qwen3-0.6B
            - --extra-engine-args
40
            - ./recipes/qwen3/trtllm/prefill.yaml
41
42
            - --disaggregation-mode
            - prefill
43
44
45
46
    TRTLLMDecodeWorker:
      dynamoNamespace: trtllm-disagg
      envFromSecret: hf-token-secret
      componentType: worker
47
      subComponentType: decode
48
49
50
51
52
53
      replicas: 1
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
54
          image: my-registry/tensorrtllm-runtime:my-tag
55
          workingDir: /workspace/
56
          command:
57
58
59
          - python3
          - -m
          - dynamo.trtllm
60
          args:
61
62
63
64
65
            - --model-path
            - Qwen/Qwen3-0.6B
            - --served-model-name
            - Qwen/Qwen3-0.6B
            - --extra-engine-args
66
            - ./recipes/qwen3/trtllm/decode.yaml
67
68
            - --disaggregation-mode
            - decode