优化应用初始化逻辑，使用异步上下文管理器处理生命周期；改进模型加载机制，添加线程锁以确保线程安全；更新 Gunicorn 配置以提高性能和稳定性

app.py

@@ -1,16 +1,37 @@
 import os
 from fastapi import FastAPI
 from pydantic import BaseModel
-from fastapi.middleware.wsgi import WSGIMiddleware
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.middleware.trustedhost import TrustedHostMiddleware
-
-from transformers import pipeline
+import asyncio
+from contextlib import asynccontextmanager
 
 from RequestModel import PredictRequest
-from us_stock import fetch_symbols
 
-app = FastAPI()  # 创建 FastAPI 应用
+# 全局变量，用于跟踪初始化状态
+is_initialized = False
+initialization_lock = asyncio.Lock()
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    # 启动时运行
+    global is_initialized
+    async with initialization_lock:
+        if not is_initialized:
+            await initialize_application()
+            is_initialized = True
+    yield
+    # 关闭时运行
+    # cleanup_code_here()
+
+async def initialize_application():
+    # 在这里进行所有需要的初始化
+    from us_stock import fetch_symbols
+
+    await fetch_symbols()
+    # 其他初始化代码...
+
+app = FastAPI(lifespan=lifespan)
 
 # 添加 CORS 中间件和限流配置
 app.add_middleware(
@@ -55,25 +76,12 @@ async def api_bbb(request: TextRequest):
     result = request.text + 'bbb'
     return {"result": result}
 
-
-@app.on_event("startup")
-async def initialize_symbols():
-    # 在 FastAPI 启动时初始化变量
-    await fetch_symbols()
-
 # 优化预测路由
 @app.post("/api/predict")
 async def predict(request: PredictRequest):
     from blkeras import predict
-
     try:
-        # 使用 asyncio.to_thread 将同步操作转换为异步
-        import asyncio
-        result = await asyncio.to_thread(
-            predict, 
-            request.text, 
-            request.stock_codes
-        )
+        result = await asyncio.to_thread(predict, request.text, request.stock_codes)
         return result
     except Exception as e:
         return []

blkeras.py

@@ -27,35 +27,48 @@ os.environ["TOKENIZERS_PARALLELISM"] = "false"
 # 设置环境变量，指定 Hugging Face 缓存路径
 os.environ["HF_HOME"] = "/tmp/huggingface"
 
+import threading
+
+# 添加线程锁
+model_lock = threading.Lock()
+model_initialized = False
+
 # 加载模型
 model = None
-if model is None:
-    # 从环境变量中获取 Hugging Face token
-    hf_token = os.environ.get("HF_Token")
 
-    
-    # 使用 Hugging Face API token 登录 (确保只读权限)
-    if hf_token:
-        login(token=hf_token)
-    else:
-        raise ValueError("Hugging Face token not found in environment variables.")
+def get_model():
+    global model, model_initialized
+    if not model_initialized:
+        with model_lock:
+            if not model_initialized:  # 双重检查锁定
+                # 从环境变量中获取 Hugging Face token
+                hf_token = os.environ.get("HF_Token")
+
+                
+                # 使用 Hugging Face API token 登录 (确保只读权限)
+                if hf_token:
+                    login(token=hf_token)
+                else:
+                    raise ValueError("Hugging Face token not found in environment variables.")
 
-    # 下载模型到本地
-    model_path = hf_hub_download(repo_id="parkerjj/BuckLake-Stock-Model", 
-                                filename="stock_prediction_model_1118_final.keras", 
-                                use_auth_token=hf_token)
+                # 下载模型到本地
+                model_path = hf_hub_download(repo_id="parkerjj/BuckLake-Stock-Model", 
+                                            filename="stock_prediction_model_1118_final.keras", 
+                                            use_auth_token=hf_token)
 
-    # 使用 Keras 加载模型
-    os.environ["KERAS_BACKEND"] = "jax"
-    print(f"Loading saved model from {model_path}...")
-    from model_build import TransformerEncoder, ExpandDimension, ConcatenateTimesteps
-    model = keras.saving.load_model(model_path, custom_objects={
-        "TransformerEncoder": TransformerEncoder, 
-        "ExpandDimension": ExpandDimension,
-        "ConcatenateTimesteps": ConcatenateTimesteps
-    })
+                # 使用 Keras 加载模型
+                os.environ["KERAS_BACKEND"] = "jax"
+                print(f"Loading saved model from {model_path}...")
+                from model_build import TransformerEncoder, ExpandDimension, ConcatenateTimesteps
+                model = keras.saving.load_model(model_path, custom_objects={
+                    "TransformerEncoder": TransformerEncoder, 
+                    "ExpandDimension": ExpandDimension,
+                    "ConcatenateTimesteps": ConcatenateTimesteps
+                })
 
-    model.summary()
+                model.summary()
+                model_initialized = True
+    return model
 
 
 
@@ -106,6 +119,7 @@ def predict(text: str, stock_codes: list):
             
         print(f"Input Text Length: {len(text)}, Start with: {text[:200] if len(text) > 200 else text}")
         print("Input stock codes:", stock_codes)
+        print("Current Time:", datetime.now())
 
         start_time = datetime.now()
         input_text = text
@@ -230,6 +244,7 @@ def predict(text: str, stock_codes: list):
             #     print(f"模型所需的输入层 {layer.name},   形状: {layer.shape}")
 
             # 使用模型进行预测
+            model = get_model()
             predictions = model.predict(features)
 
             # 生成伪精准度值

gunicorn.conf.py

@@ -11,6 +11,9 @@ workers = multiprocessing.cpu_count() + 1
 # 设置为2，增加并发处理能力
 threads = 2
 
+# 请求超时时间
+timeout = 600
+
 # 工作方式
 worker_class = "uvicorn.workers.UvicornWorker"
 
@@ -27,7 +30,20 @@ worker_connections = 2000
 
 # 工作模式
 worker_tmp_dir = "/dev/shm"  # 使用内存文件系统提高性能
-preload_app = True           # 预加载应用，减少启动时间
+preload_app = False          # 修改为 False，避免重复加载
+
+# 添加新的配置
+reload = False              # 禁用自动重载
+daemon = False             # 非守护进程模式运行
+
+# 添加应用初始化钩子
+def when_ready(server):
+    # 当 Gunicorn 准备好时执行
+    server.log.info("Server is ready. Doing nothing.")
+
+def post_fork(server, worker):
+    # 当 worker 进程被 fork 后执行
+    server.log.info(f"Worker spawned (pid: {worker.pid})")
 
 # 进程名称前缀
 proc_name = 'gunicorn_fastapi'