Ascend 910B K8s部署脚本分享（vllm-http-client模式，两个 Deployment，各自一个 Service）

[shaoqing404]

docker file

FROM quay.m.daocloud.io/ascend/vllm-ascend:v0.13.0rc1

# Install runtime deps
RUN apt-get update && \
    apt-get install -y \
    fonts-noto-core \
    fonts-noto-cjk \
    fontconfig \
    libgl1 \
    libglib2.0-0 && \
    fc-cache -fv && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

# Install MinerU
RUN python3 -m pip install -U pip -i https://mirrors.aliyun.com/pypi/simple && \
    python3 -m pip install \
      'mineru[all]==2.7.1' \
      numpy==1.26.4 \
      opencv-python==4.11.0.86 \
      -i https://mirrors.aliyun.com/pypi/simple && \
    python3 -m pip cache purge

# Download models into default path (official route)
RUN TORCH_DEVICE_BACKEND_AUTOLOAD=0 /bin/bash -c "mineru-models-download -s modelscope -m all"

# Runtime defaults (can be overridden by K8s env)
ENV MINERU_MODEL_SOURCE=local
ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
ENV MINERU_LMDEPLOY_DEVICE=ascend
ENV MINERU_HOST=0.0.0.0
ENV MINERU_OPENAI_PORT=30000
ENV MINERU_API_PORT=8000
ENV MINERU_MODEL_SOURCE=modelscope
ENV MINERU_VL_SERVER=http://mineru-openai:30000

# Provide a single entrypoint that can start either service
COPY entrypoint.sh /usr/local/bin/entrypoint.sh
RUN chmod +x /usr/local/bin/entrypoint.sh

EXPOSE 30000 8000

ENTRYPOINT ["/usr/local/bin/entrypoint.sh"]
# default mode: openai server
CMD ["openai"]

entrypoint.sh

`#!/usr/bin/env bash
set -euo pipefail

MODE="${1:-openai}"

# Common env defaults (already set in Dockerfile ENV, but keep safe)
: "${MINERU_MODEL_SOURCE:=local}"
: "${VLLM_WORKER_MULTIPROC_METHOD:=spawn}"
: "${MINERU_LMDEPLOY_DEVICE:=ascend}"
: "${MINERU_HOST:=0.0.0.0}"
: "${MINERU_OPENAI_PORT:=30000}"
: "${MINERU_API_PORT:=8000}"
: "${MINERU_MODEL_SOURCE:=modelscope}"
: "${MINERU_VL_SERVER:=http://mineru-openai:30000}"

export MINERU_MODEL_SOURCE VLLM_WORKER_MULTIPROC_METHOD MINERU_LMDEPLOY_DEVICE MINERU_MODEL_SOURCE MINERU_APISERVER

echo "MODE=${MODE} MINERU_VL_SERVER=${MINERU_VL_SERVER}"

case "${MODE}" in
  openai|server)
    exec mineru-openai-server --host "${MINERU_HOST}" --port "${MINERU_OPENAI_PORT}"
    ;;
  api|client)
    exec mineru-api --host "${MINERU_HOST}" --port "${MINERU_API_PORT}"
    ;;
  bash|shell)
    exec /bin/bash
    ;;
  *)
    echo "Unknown MODE: ${MODE}"
    echo "Usage: entrypoint.sh [openai|api|bash]"
    exit 2
    ;;
esac

K8s启动脚本（请考虑实际情况修改，基准模版由DeepSeek生成）

！！！请务必在运维人员监督下使用！！！
！！！请务必在运维人员监督下使用！！！
！！！请务必在运维人员监督下使用！！！

apiVersion: v1
kind: Namespace
metadata:
  name: mineru
---
# OpenAI/VLM Server Service (30000)
apiVersion: v1
kind: Service
metadata:
  name: mineru-openai
  namespace: mineru
spec:
  selector:
    app: mineru-openai
  ports:
    - name: openai
      port: 30000
      targetPort: 30000
  type: ClusterIP
---
# OpenAI/VLM Server Deployment
apiVersion: apps/v1
kind: Deployment
metadata:
  name: mineru-openai
  namespace: mineru
spec:
  replicas: 1
  selector:
    matchLabels:
      app: mineru-openai
  template:
    metadata:
      labels:
        app: mineru-openai
    spec:
      containers:
        - name: openai
          image: your-registry/mineru:npu-vllm-latest
          imagePullPolicy: IfNotPresent
          args: ["openai"]
          env:
            - name: MINERU_MODEL_SOURCE
              value: "local"
            - name: MINERU_HOST
              value: "0.0.0.0"
            - name: MINERU_OPENAI_PORT
              value: "30000"
            - name: VLLM_WORKER_MULTIPROC_METHOD
              value: "spawn"
            - name: MINERU_LMDEPLOY_DEVICE
              value: "ascend"
          ports:
            - containerPort: 30000
              name: openai
          # 如果你们的 device plugin 支持 resource request，就在这里填
          resources:
            limits:
              # 替换成你们集群实际的 Ascend 资源名，例如：
              # huawei.com/Ascend910: "1"
              # ascend.com/npu: "1"
              example.com/ascend-npu: "1"
          readinessProbe:
            httpGet:
              path: /docs
              port: 30000
            initialDelaySeconds: 20
            periodSeconds: 10
            timeoutSeconds: 3
            failureThreshold: 12
          livenessProbe:
            httpGet:
              path: /docs
              port: 30000
            initialDelaySeconds: 60
            periodSeconds: 20
            timeoutSeconds: 3
            failureThreshold: 6
---
# API Service (8000)
apiVersion: v1
kind: Service
metadata:
  name: mineru-api
  namespace: mineru
spec:
  selector:
    app: mineru-api
  ports:
    - name: api
      port: 8000
      targetPort: 8000
  type: ClusterIP
---
# API Deployment
apiVersion: apps/v1
kind: Deployment
metadata:
  name: mineru-api
  namespace: mineru
spec:
  replicas: 1
  selector:
    matchLabels:
      app: mineru-api
  template:
    metadata:
      labels:
        app: mineru-api
    spec:
      containers:
        - name: api
          image: your-registry/mineru:npu-vllm-latest
          imagePullPolicy: IfNotPresent
          args: ["api"]
          env:
            - name: MINERU_MODEL_SOURCE
              value: "local"
            - name: MINERU_HOST
              value: "0.0.0.0"
            - name: MINERU_API_PORT
              value: "8000"
            # 关键：API 通过 service DNS 访问 openai server
            - name: MINERU_VL_SERVER
              value: "http://mineru-openai.mineru.svc.cluster.local:30000"
          ports:
            - containerPort: 8000
              name: api
          readinessProbe:
            httpGet:
              path: /docs
              port: 8000
            initialDelaySeconds: 10
            periodSeconds: 10
            timeoutSeconds: 3
            failureThreshold: 12
          livenessProbe:
            httpGet:
              path: /docs
              port: 8000
            initialDelaySeconds: 30
            periodSeconds: 20
            timeoutSeconds: 3
            failureThreshold: 6