dynamo/recipes/deepseek-v4/deepseek-v4-pro/sglang/agg/deploy.yaml at main · nvyutwu/dynamo · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# DeepSeek-V4-Pro SGLang DynamoGraphDeployment
# Aggregated mode, B200x8 TP8, MXFP4 MoE
# EAGLE disabled: crashes on long context (>18k tokens) and tool calling on B200 TP8
# due to insufficient GPU memory headroom (4.23 GB/GPU with EAGLE vs 7.75 GB without)

apiVersion: nvidia.com/v1alpha1
kind: DynamoGraphDeployment
metadata:
  name: sglang-dsv4-pro
spec:
  services:
    Frontend:
      componentType: frontend
      replicas: 1
      volumeMounts:
        - name: shared-model-cache
          mountPoint: /models
      extraPodSpec:
        mainContainer:
          image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.2.0-deepseek-v4-cuda12-dev.3
          imagePullPolicy: Always
          env:
            - name: HF_HOME
              value: /models
            - name: HF_HUB_OFFLINE
              value: "1"
          startupProbe:
            httpGet:
              path: /health
              port: 8000
            periodSeconds: 10
            timeoutSeconds: 10
            failureThreshold: 360

    decode:
      componentType: worker
      subComponentType: decode
      replicas: 1
      resources:
        limits:
          gpu: "8"
      volumeMounts:
        - name: shared-model-cache
          mountPoint: /models
      sharedMemory:
        size: 200Gi
      extraPodSpec:
        nodeSelector:
          nvidia.com/gpu.product: NVIDIA-B200
        tolerations:
          - key: nvidia.com/gpu
            operator: Equal
            value: "true"
            effect: NoSchedule
        mainContainer:
          image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.2.0-deepseek-v4-cuda12-dev.3
          imagePullPolicy: Always
          workingDir: /workspace
          command:
            - python3
            - -m
            - dynamo.sglang
          args:
            - --model-path
            - deepseek-ai/DeepSeek-V4-Pro
            - --served-model-name
            - deepseek-ai/DeepSeek-V4-Pro
            - --trust-remote-code
            - --tp
            - "8"
            - --moe-runner-backend
            - flashinfer_mxfp4
            - --chunked-prefill-size
            - "4096"
            - --disable-flashinfer-autotune
            - --dyn-tool-call-parser
            - deepseek_v4
            - --dyn-reasoning-parser
            - deepseek_v4
          env:
            - name: HF_HOME
              value: /models
            - name: HF_HUB_OFFLINE
              value: "1"
            - name: CUDA_VISIBLE_DEVICES
              value: "0,1,2,3,4,5,6,7"
            - name: SGLANG_JIT_DEEPGEMM_PRECOMPILE
              value: "0"
            - name: SGLANG_JIT_DEEPGEMM_FAST_WARMUP
              value: "1"
            - name: NCCL_CUMEM_ENABLE
              value: "1"
            - name: GLOO_SOCKET_IFNAME
              value: eth0
          startupProbe:
            httpGet:
              path: /health
              port: 9090
            periodSeconds: 10
            timeoutSeconds: 10
            failureThreshold: 360

  pvcs:
    - name: shared-model-cache
      create: false