disagg.yaml 1.79 KB
Newer Older
1
# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
4
5
6
7
8
9
10
# SPDX-License-Identifier: Apache-2.0

apiVersion: nvidia.com/v1alpha1
kind: DynamoGraphDeployment
metadata:
  name: trtllm-disagg
spec:
  services:
    Frontend:
11
      componentType: frontend
12
13
14
      replicas: 1
      extraPodSpec:
        mainContainer:
15
          image: my-registry/tensorrtllm-runtime:my-tag
16
17
18
    TRTLLMPrefillWorker:
      envFromSecret: hf-token-secret
      componentType: worker
19
      subComponentType: prefill
20
21
22
23
24
25
      replicas: 1
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
26
          image: my-registry/tensorrtllm-runtime:my-tag
27
          workingDir: /workspace/
28
          command:
29
30
31
          - python3
          - -m
          - dynamo.trtllm
32
          args:
33
34
35
36
37
            - --model-path
            - Qwen/Qwen3-0.6B
            - --served-model-name
            - Qwen/Qwen3-0.6B
            - --extra-engine-args
38
            - ./examples/backends/trtllm/engine_configs/qwen3/prefill.yaml
39
40
            - --disaggregation-mode
            - prefill
41
42
43
    TRTLLMDecodeWorker:
      envFromSecret: hf-token-secret
      componentType: worker
44
      subComponentType: decode
45
46
47
48
49
50
      replicas: 1
      resources:
        limits:
          gpu: "1"
      extraPodSpec:
        mainContainer:
51
          image: my-registry/tensorrtllm-runtime:my-tag
52
          workingDir: /workspace/
53
          command:
54
55
56
          - python3
          - -m
          - dynamo.trtllm
57
          args:
58
59
60
61
62
            - --model-path
            - Qwen/Qwen3-0.6B
            - --served-model-name
            - Qwen/Qwen3-0.6B
            - --extra-engine-args
63
            - ./examples/backends/trtllm/engine_configs/qwen3/decode.yaml
64
65
            - --disaggregation-mode
            - decode