ChatGPT本地部署实战：从模型加载到API封装的技术解析

1次阅读

共计 2527 个字符，预计需要花费 7 分钟才能阅读完成。

在云端调用 ChatGPT API 时，开发者常遇到三个主要问题：

延迟不可控：网络传输和云端排队导致响应时间波动大
数据外流风险：敏感数据需要出境到第三方服务器
计费不可预测：突发流量可能导致意外的高额账单

本地部署能有效解决这些问题，同时还能实现：

定制化模型微调
私有知识库集成
硬件资源最大化利用

模型类型	显存占用	推理速度	精度损失
原始 FP16	13GB	1x	无
8bit 量化	7GB	1.2x	<1%
4bit 量化	4GB	1.5x	~3%

框架	QPS	平均延迟	内存占用
FastAPI	85	12ms	120MB
Flask	62	16ms	150MB

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

def load_model(model_path: str, quantize: bool = True):
    """
    加载量化模型并优化显存分配
    :param model_path: 模型本地路径
    :param quantize: 是否启用 8bit 量化
    """
    tokenizer = AutoTokenizer.from_pretrained(model_path)

    # 显存优化配置
    torch.backends.cudnn.benchmark = True
    torch.set_default_tensor_type(torch.cuda.HalfTensor)

    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        load_in_8bit=quantize,
        device_map='auto',  # 自动分配多 GPU
        torch_dtype=torch.float16
    )
    return model, tokenizer

from fastapi import Depends, HTTPException
from fastapi.security import HTTPBearer
import jwt

security = HTTPBearer()

async def verify_token(token: str = Depends(security)):
    try:
        payload = jwt.decode(
            token.credentials, 
            "your_secret_key",
            algorithms=["HS256"]
        )
        return payload
    except jwt.ExpiredSignatureError:
        raise HTTPException(status_code=403, detail="Token expired")
    except Exception:
        raise HTTPException(status_code=403, detail="Invalid token")

from fastapi import Response
from sse_starlette.sse import EventSourceResponse

async def stream_generator(prompt: str):
    for chunk in model.generate_stream(prompt):
        yield {
            "event": "text_chunk",
            "data": chunk
        }

@app.get("/stream")
async def stream_response(prompt: str):
    return EventSourceResponse(stream_generator(prompt))

# 构建阶段
FROM python:3.9-slim as builder

WORKDIR /app
COPY requirements.txt .
RUN pip install --user -r requirements.txt

# 运行阶段
FROM alpine:3.14

COPY --from=builder /root/.local /root/.local
COPY --from=builder /app .

# 优化 Alpine 镜像
RUN apk add --no-cache libstdc++ && \
    ln -s /root/.local/bin/* /usr/local/bin/

EXPOSE 8000
CMD ["uvicorn", "main:app", "--host", "0.0.0.0"]

resources:
  limits:
    nvidia.com/gpu: 1
    memory: 8Gi
  requests:
    memory: 6Gi

livenessProbe:
  httpGet:
    path: /healthz
    port: 8000
  initialDelaySeconds: 30

readinessProbe:
  tcpSocket:
    port: 8000