开源模型如何成为各国实现 AI 主权的现实路径

# 安装 vLLM（需要 Python 3.9+，CUDA 12.1+）
pip install vllm

# 启动推理服务，开放 OpenAI 兼容 API
vllm serve Qwen/Qwen2.5-7B-Instruct \
  --api-key your-sovereign-key \
  --host 0.0.0.0 \
  --port 8000 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.9

from openai import OpenAI

# 指向你的主权推理服务，而非远端 API
client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="your-sovereign-key"
)

response = client.chat.completions.create(
    model="Qwen/Qwen2.5-7B-Instruct",
    messages=[
        {"role": "system", "content": "你是一个中国金融法规助手，只基于用户提供的上下文回答。"},
        {"role": "user", "content": "根据以下条文，企业发行债券的主体资格要求是什么？\n《公司法》第十六条：公司发行债券…"}
    ],
    temperature=0.3
)

print(response.choices[0].message.content)

# 安装微调工具链
pip install transformers peft datasets accelerate

# 下载基座模型权重
huggingface-cli download Qwen/Qwen2.5-7B-Instruct \
  --local-dir ./qwen2.5-7b-base

from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig, get_peft_model, TaskType
from datasets import load_dataset
from trl import SFTTrainer

model = AutoModelForCausalLM.from_pretrained(
    "./qwen2.5-7b-base",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("./qwen2.5-7b-base")

# LoRA 只训练 ~0.5% 的参数，显存需求从 ~80GB 降到 ~16GB
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,              # LoRA 秩
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# 输出类似：trainable params: 11,272,192 || all params: 7,615,616,000 || 0.15%

# 准备你的领域数据（JSONL 格式，每行一条对话）
# {"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
dataset = load_dataset("json", data_files="your_domain_data.jsonl", split="train")

training_args = TrainingArguments(
    output_dir="./qwen2.5-7b-lora-finetuned",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=10,
    save_steps=100
)

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=training_args,
    train_dataset=dataset,
    max_seq_length=2048,
    packing=False
)

trainer.train()

# 保存 LoRA 权重（仅几十 MB，而非完整模型的 ~15GB）
model.save_pretrained("./qwen2.5-7b-lora-finetuned")
tokenizer.save_pretrained("./qwen2.5-7b-lora-finetuned")

from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "./qwen2.5-7b-base", torch_dtype="auto", device_map="auto"
)
sovereign_model = PeftModel.from_pretrained(
    base_model, "./qwen2.5-7b-lora-finetuned"
)
# 现在你拥有一个注入了本国领域知识的、完全本地运行的模型

# MindIE 部署配置示例（昇腾 910B）
# mindie_config.yaml
server:
  host: 0.0.0.0
  port: 8000
  api_format: openai  # 保持 API 兼容，应用层无需改动

model:
  name: Qwen2.5-7B-Instruct
  path: ./qwen2.5-7b-base
  lora_path: ./qwen2.5-7b-lora-finetuned
  max_batch_size: 32
  max_seq_len: 8192

hardware:
  device: ascend
  npu_ids: [0, 1]  # 使用两张昇腾 910B