从试点到生产：红帽与 NVIDIA AI 工厂如何让自主智能体真正跑起来

# nim-model-deployment.yaml
# 在 OpenShift 上部署一个 NVIDIA NIM 模型服务，供自主智能体调用
apiVersion: apps/v1
kind: Deployment
metadata:
  name: nim-llama3-8b-instruct
  namespace: ai-factory
  labels:
    app: nim-model
    model: llama3-8b-instruct
    version: v1.0.0
spec:
  replicas: 2
  selector:
    matchLabels:
      app: nim-model
      model: llama3-8b-instruct
  strategy:
    type: RollingUpdate
    rollingUpdate:
      maxSurge: 1
      maxUnavailable: 0        # 零中断滚动更新
  template:
    metadata:
      labels:
        app: nim-model
        model: llama3-8b-instruct
        version: v1.0.0
      annotations:
        audit.redhat.com/enabled: "true"   # 开启平台级审计
    spec:
      serviceAccountName: nim-model-sa     # 限制权限的专用 SA
      containers:
        - name: nim
          image: nvcr.io/nim/meta/llama3-8b-instruct:1.0.0
          ports:
            - containerPort: 8000
          resources:
            limits:
              nvidia.com/gpu: 1            # 每实例独占一块 GPU
            requests:
              nvidia.com/gpu: 1
          env:
            - name: NIM_MODEL_NAME
              value: "meta/llama3-8b-instruct"
            - name: NIM_MAX_BATCH_SIZE
              value: "32"
          volumeMounts:
            - name: model-cache
              mountPath: /model-cache
            - name: audit-log
              mountPath: /audit-log
      volumes:
        - name: model-cache
          persistentVolumeClaim:
            claimName: nim-model-cache-pvc
        - name: audit-log
          persistentVolumeClaim:
            claimName: audit-log-pvc        # 审计日志持久存储
---
apiVersion: v1
kind: Service
metadata:
  name: nim-llama3-8b-instruct
  namespace: ai-factory
spec:
  selector:
    app: nim-model
    model: llama3-8b-instruct
  ports:
    - port: 8000
      targetPort: 8000
---
# 智能体策略 ConfigMap——可独立更新，不影响模型服务
apiVersion: v1
kind: ConfigMap
metadata:
  name: agent-policy-finance-risk
  namespace: ai-factory
data:
  policy.yaml: |
    tools:
      - name: query_risk_db
        allowed: true
        audit: true
        max_calls_per_minute: 60
      - name: send_alert
        allowed: true
        audit: true
        requires_approval: false
    guardrails:
      max_context_tokens: 4096
      refuse_topics: ["personal_advice", "political"]
      output_format: "structured_json"
    compliance:
      log_retention_days: 365
      decision_trace: true

# 创建命名空间和 GPU Operator 已就绪的前提下
oc apply -f nim-model-deployment.yaml

# 查看模型服务状态
oc get pods -n ai-factory -l app=nim-model

# 查看审计日志输出
oc logs -n ai-factory deployment/nim-llama3-8b-instruct | grep audit

# 灰度发布新版本：先改 version 标签，用 canary 路由切 5% 流量
oc patch deployment nim-llama3-8b-instruct -n ai-factory \
  -p '{"spec":{"template":{"metadata":{"labels":{"version":"v1.1.0"}}}}}'

# agent_state_manager.py
# 智能体状态外置到 Redis，Pod 重启后可恢复
import redis
import json
from datetime import datetime

r = redis.Redis(host="redis.ai-factory", port=6379, decode_responses=True)

class AgentStateManager:
    def __init__(self, agent_id: str):
        self.key = f"agent:{agent_id}:state"

    def save_context(self, conversation_id: str, messages: list):
        """保存对话上下文，设置 24 小时过期"""
        payload = {
            "conversation_id": conversation_id,
            "messages": messages,
            "updated_at": datetime.utcnow().isoformat(),
        }
        r.hset(self.key, conversation_id, json.dumps(payload))
        r.expire(self.key, 86400)  # 24h TTL，自动清理过期对话

    def load_context(self, conversation_id: str) -> dict:
        """Pod 重启后恢复上下文"""
        raw = r.hget(self.key, conversation_id)
        if raw:
            return json.loads(raw)
        return {"conversation_id": conversation_id, "messages": [], "updated_at": None}

    def save_task_progress(self, task_id: str, step: int, result: dict):
        """记录任务执行进度"""
        r.hset(
            f"agent:{self.key}:tasks",
            task_id,
            json.dumps({"step": step, "result": result, "ts": datetime.utcnow().isoformat()})
        )

# 使用示例
state = AgentStateManager("finance-risk-agent-001")
state.save_context("conv-789", [{"role": "user", "content": "检查账户 A 风险"}])
# Pod 崩溃重启后
recovered = state.load_context("conv-789")
print(recovered["messages"])  # 上下文完整恢复