在 SageMaker JumpStart 上部署 NVIDIA Nemotron 3 Ultra：更快推理、更低成本的 Agentic AI 实践

import sagemaker
from sagemaker.jumpstart.model import JumpStartModel

# 创建 SageMaker 会话
role = sagemaker.get_execution_role()
session = sagemaker.Session()

# 从 JumpStart 模型目录选择 Nemotron 3 Ultra
# 模型 ID 可在 JumpStart 控制台搜索 "Nemotron" 确认
model_id = "nvidia-nemotron-3-ultra"  # 实际部署时请确认 JumpStart 中的精确 model_id

model = JumpStartModel(
    model_id=model_id,
    role=role,
    sagemaker_session=session,
)

# 部署为实时推理终端节点
# instance_type 根据负载选择，g5.12xlarge 是常见起步配置
predictor = model.deploy(
    initial_instance_count=1,
    instance_type="ml.g5.12xlarge",
)

print(f"终端节点名称: {predictor.endpoint_name}")

import json

# 构造 Agentic 场景的推理请求
payload = {
    "messages": [
        {
            "role": "system",
            "content": "你是一个任务规划 Agent。根据用户需求，拆解执行步骤并调用相应工具。"
        },
        {
            "role": "user",
            "content": "帮我分析最近一周的销售数据，找出异常波动，并给出可能的原因和应对建议。"
        }
    ],
    "max_tokens": 2048,
    "temperature": 0.3,  # Agentic 场景建议较低温度，减少随机性
}

response = predictor.predict(payload)
result = json.loads(response["body"].read().decode("utf-8"))
print(result["choices"][0]["message"]["content"])

predictor.delete_endpoint()

def agent_loop(user_task, predictor, max_rounds=10):
    context = [{"role": "user", "content": user_task}]

    for round_num in range(max_rounds):
        # 推理：生成下一步行动
        payload = {
            "messages": context,
            "max_tokens": 1024,
            "temperature": 0.2,
        }
        response = predictor.predict(payload)
        model_output = parse_response(response)

        # 解析行动类型
        if model_output["action"] == "finish":
            return model_output["final_answer"]
        elif model_output["action"] == "call_tool":
            tool_result = execute_tool(model_output["tool_name"], model_output["tool_args"])
            context.append({"role": "assistant", "content": model_output["reasoning"]})
            context.append({"role": "user", "content": f"工具返回结果：{tool_result}"})
        else:
            context.append({"role": "assistant", "content": model_output["content"]})

    return "达到最大轮数，任务未完成"