修改启动命令为 Uvicorn，调整 Gunicorn 配置中的工作进程数；添加懒加载机制以优化模型和tokenizer的加载，确保线程安全

Dockerfile

@@ -45,4 +45,4 @@ RUN --mount=type=secret,id=HF_Token,mode=0444,required=true \
 
 
 # 修改启动命令，使用配置文件
-CMD ["gunicorn", "app:app", "-c", "gunicorn.conf.py"]
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", "--workers", "3"]

gunicorn.conf.py

@@ -5,7 +5,7 @@ bind = "0.0.0.0:7860"
 
 # 修改工作进程数配置策略
 # 对于CPU密集型应用，建议设置为 CPU核心数 + 1
-workers = multiprocessing.cpu_count() + 1
+workers = 3
 
 # 每个工作进程的线程数
 # 设置为2，增加并发处理能力

preprocess.py

@@ -51,14 +51,35 @@ print("Is NPL GPU used Preprocessing.py:", spacy.prefer_gpu())
 
 
 # 使用合适的模型和tokenizer
-tokenizer_one = AutoTokenizer.from_pretrained("ProsusAI/finbert")
-sa_model_one = AutoModelForSequenceClassification.from_pretrained("ProsusAI/finbert")
+# tokenizer_one = AutoTokenizer.from_pretrained("ProsusAI/finbert")
+# sa_model_one = AutoModelForSequenceClassification.from_pretrained("ProsusAI/finbert")
 
 
-tokenizer_two = BertTokenizer.from_pretrained('yiyanghkust/finbert-tone')
-sa_model_two = BertForSequenceClassification.from_pretrained('yiyanghkust/finbert-tone',num_labels=3)
+# tokenizer_two = BertTokenizer.from_pretrained('yiyanghkust/finbert-tone')
+# sa_model_two = BertForSequenceClassification.from_pretrained('yiyanghkust/finbert-tone',num_labels=3)
 
+import multiprocessing
 
+# 添加进程锁
+_tokenizer_lock = multiprocessing.Lock()
+_models = {}
+
+def get_tokenizer_and_model(model_type="one"):
+    """懒加载tokenizer和model"""
+    global _models
+    
+    if model_type not in _models:
+        with _tokenizer_lock:
+            if model_type not in _models:  # 双重检查锁定
+                if model_type == "one":
+                    tokenizer = AutoTokenizer.from_pretrained("ProsusAI/finbert")
+                    model = AutoModelForSequenceClassification.from_pretrained("ProsusAI/finbert")
+                else:
+                    tokenizer = BertTokenizer.from_pretrained('yiyanghkust/finbert-tone')
+                    model = BertForSequenceClassification.from_pretrained('yiyanghkust/finbert-tone', num_labels=3)
+                _models[model_type] = (tokenizer, model)
+                
+    return _models[model_type]
 
 index_us_stock_index_INX = ak.index_us_stock_sina(symbol=".INX")
 index_us_stock_index_DJI = ak.index_us_stock_sina(symbol=".DJI")
@@ -273,6 +294,10 @@ def get_sentiment_score(text):
     try:
         import torch
         
+        # 懒加载获取tokenizer和model
+        tokenizer_one, sa_model_one = get_tokenizer_and_model("one")
+        tokenizer_two, sa_model_two = get_tokenizer_and_model("two")
+        
         # 将长文本分段
         segments_one = process_long_text(text, tokenizer_one)
         segments_two = process_long_text(text, tokenizer_two)