update initialize_model_once and create_llm_pipeline to handle error Failed to create pipeline: 'model'

app.py

@@ -224,21 +224,37 @@ def initialize_model_once(model_key):
                 print(f"Error loading model {model_name}: {str(e)}")
                 print(traceback.format_exc())
                 raise RuntimeError(f"Failed to load model {model_name}: {str(e)}")
-                
+
+        # Final verification that model loaded correctly
+        if MODEL_CACHE["model"] is None:
+            print(f"WARNING: Model {model_name} appears to be None after loading")
+            # Try to free memory before returning
+            torch.cuda.empty_cache() if torch.cuda.is_available() else None
+            gc.collect()
+        
     return MODEL_CACHE["tokenizer"], MODEL_CACHE["model"], MODEL_CACHE.get("is_gguf", False)
 
 def create_llm_pipeline(model_key):
-    """Create a new pipeline using the specified model"""
+    """Create a new pipeline using the specified model with better error handling"""
     try:
         print(f"Creating pipeline for model: {model_key}")
         tokenizer, model, is_gguf = initialize_model_once(model_key)
 
+        # Additional check to ensure model was properly loaded
+        if model is None:
+            print(f"Model is None for {model_key}, falling back to alternate model")
+            fallback_model = get_fallback_model(model_key)
+            if fallback_model != model_key:
+                print(f"Attempting to use fallback model: {fallback_model}")
+                tokenizer, model, is_gguf = initialize_model_once(fallback_model)
+                if model is None:
+                    raise ValueError(f"Both original and fallback models failed to load")
+            else:
+                raise ValueError(f"Model is None and no fallback available")
+        
         # Get the model info for reference
-        model_info = MODEL_CONFIG[model_key]
+         model_info = MODEL_CONFIG.get(model_key, MODEL_CONFIG.get(fallback_model, {}))
         
-        if model is None:
-            raise ValueError(f"Model is None for {model_key}")
-
         # For GGUF models from llama-cpp-python
         if is_gguf:
             # Create adapter to use GGUF model like HF pipeline