Fine-tuning/backend/main.py

import os

# 禁用 FlashAttention，解决沐曦显卡共享内存不足问题
# 必须放在最开头，在任何库导入之前设置
os.environ["PYTORCH_NO_FLASH"] = "1"
os.environ["FLASH_ATTENTION_ENABLED"] = "0"
os.environ["USE_FLASH_ATTENTION"] = "0"
os.environ["TORCH_FLASH_ATTN"] = "0"

from contextlib import asynccontextmanager

from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware

from app.config import get_settings

settings = get_settings()


@asynccontextmanager
async def lifespan(app: FastAPI):
    # 启动时：确保数据目录存在 + 初始化数据库 + 启动 JobQueue
    settings.ensure_dirs()
    from app.core.db import init_db

    await init_db()

    from app.core.job_queue import job_queue
    from app.services.training_service import update_job_in_db

    job_queue.register_callback(update_job_in_db)
    await job_queue.start()

    yield

    # 关闭时：停止 JobQueue
    await job_queue.stop()


def create_app() -> FastAPI:
    app = FastAPI(
        title="PEFT Fine-Tuning Platform",
        version="0.1.0",
        lifespan=lifespan,
    )

    # CORS 中间件
    app.add_middleware(
        CORSMiddleware,
        allow_origins=settings.backend_cors_origins,
        allow_credentials=True,
        allow_methods=["*"],
        allow_headers=["*"],
    )

    # 挂载路由
    from app.api import models as models_api
    from app.api import datasets as datasets_api
    from app.api import training as training_api
    from app.api import evaluation as evaluation_api
    from app.api import deployment as deployment_api
    from app.api import inference as inference_api

    app.include_router(models_api.router, prefix="/api/v1/models", tags=["models"])
    app.include_router(datasets_api.router, prefix="/api/v1/datasets", tags=["datasets"])
    app.include_router(training_api.router, prefix="/api/v1/training", tags=["training"])
    app.include_router(evaluation_api.router, prefix="/api/v1/evaluation", tags=["evaluation"])
    app.include_router(deployment_api.router, prefix="/api/v1/deployment", tags=["deployment"])
    app.include_router(inference_api.router, prefix="/api/v1/inference", tags=["inference"])

    # WebSocket
    from app.core.websocket import router as ws_router

    app.include_router(ws_router)

    @app.get("/health")
    async def health_check():
        return {"status": "ok", "env": settings.backend_env}

    return app


app = create_app()