Fix LSTM: better loader fallbacks, verify tf-keras

Dockerfile

@@ -10,7 +10,8 @@ RUN pip install --no-cache-dir fastapi==0.111.0 uvicorn==0.30.1 pydantic==2.7.4
     scipy==1.13.1 nltk==3.8.1 tensorflow==2.16.1 keras==3.3.3 \
     transformers==4.41.2 joblib==1.4.2 && \
     pip install --no-cache-dir torch==2.3.1 --index-url https://download.pytorch.org/whl/cpu && \
-    pip install --no-cache-dir tf-keras
+    pip install --no-cache-dir tf-keras && \
+    python -c "import tf_keras; print('tf-keras OK')"
 
 RUN python -c "import nltk; nltk.download('stopwords', quiet=True); nltk.download('punkt', quiet=True); nltk.download('wordnet', quiet=True)"
 

ml_service/app/services/model_loader.py

@@ -112,11 +112,13 @@ class ModelManager:
                         continue
                 if not loaded:
                     raise RuntimeError("Could not load LSTM model with any backend")
-                
-                @tf.function(reduce_retracing=True)
-                def fast_predict(x):
-                    return self.lstm_model(x, training=False)
-                self.lstm_fast_predict = fast_predict
+                if hasattr(self.lstm_model, "signatures") and "serving_default" in self.lstm_model.signatures:
+                    self.lstm_fast_predict = self.lstm_model.signatures["serving_default"]
+                else:
+                    @tf.function(reduce_retracing=True)
+                    def fast_predict(x):
+                        return self.lstm_model(x, training=False)
+                    self.lstm_fast_predict = fast_predict
                 
                 with open(tok_path, "rb") as f:
                     self.lstm_tokenizer = pickle.load(f)