OpenAI 发布开源模型 GPT-OSS：开发者本地部署与使用完全指南

# 安装 Ollama
curl -fsSL https://ollama.com/install.sh | sh

# 拉取并运行 GPT-OSS-12B
ollama run gpt-oss-12b

# 拉取 70B 版本（需要足够的显存或内存）
ollama run gpt-oss-70b

# 安装 vLLM
pip install vllm

# 启动 GPT-OSS-70B 推理服务
python -m vllm.entrypoints.openai.api_server \
  --model openai/gpt-oss-70b \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 2

# 测试调用
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-oss-70b",
    "messages": [{"role": "user", "content": "Hello, who are you?"}],
    "max_tokens": 100
  }'

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "openai/gpt-oss-12b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

messages = [{"role": "user", "content": "写一首关于代码的诗"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

import openai

# 指向本地 Ollama
client = openai.Client(base_url="http://localhost:11434/v1", api_key="ollama")

response = client.chat.completions.create(
    model="gpt-oss-12b",
    messages=[
        {"role": "system", "content": "你是一位资深代码审查专家。"},
        {"role": "user", "content": """审查这段 Python 代码，找出潜在的安全漏洞：
        import os
        def process_user_input(user_input):
            os.system(f"echo {user_input}")
        """}
    ],
    max_tokens=500
)
print(response.choices[0].message.content)

# 用 LangChain + GPT-OSS 搭建离线问答
from langchain.llms import Ollama
from langchain.chains import RetrievalQA
from langchain.vectorstores import FAISS

llm = Ollama(model="gpt-oss-70b", temperature=0)
vectorstore = FAISS.load_local("docs_embeddings", allow_dangerous_deserialization=True)
qa_chain = RetrievalQA.from_chain_type(llm=llm, retriever=vectorstore.as_retriever())

result = qa_chain.invoke("我们的数据保留政策是什么？")
print(result["result"])

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "获取指定城市的天气",
            "parameters": {
                "type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"]
            }
        }
    }
]

response = client.chat.completions.create(
    model="gpt-oss-12b",
    messages=[{"role": "user", "content": "北京今天天气怎么样？"}],
    tools=tools,
    tool_choice="auto"
)

# 模型返回 tool_calls，你执行函数后回填结果
call = response.choices[0].message.tool_calls[0]
print(f"调用函数: {call.function.name}")
print(f"参数: {call.function.arguments}")