Update app.py

app.py

@@ -84,38 +84,7 @@ class SafeGeocoder:
 
 
 # Corrected model loading function based on official usage example
-def load_model():
-    global tokenizer, model
-    try:
-        if model is None:
-            # Only load the tokenizer first (no CUDA initialization)
-            try:
-                from modelscope import AutoTokenizer as MSAutoTokenizer
-                tokenizer = MSAutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
-                print("Loaded tokenizer using modelscope AutoTokenizer")
-            except:
-                # Fall back to regular tokenizer
-                tokenizer = AutoTokenizer.from_pretrained(
-                    MODEL_NAME, 
-                    trust_remote_code=True,
-                    revision="main"
-                )
-                print("Loaded tokenizer using standard AutoTokenizer")
-            
-            # For the model, we'll only create a loading configuration but not actually load it yet
-            # This avoids CUDA initialization in the main process
-            print(f"Tokenizer successfully loaded, model will be loaded when needed")
-            return "✅ Tokenizer erfolgreich geladen. Model wird bei Bedarf geladen."
-            
-    except Exception as e:
-        import traceback
-        trace = traceback.format_exc()
-        print(f"Error loading tokenizer: {e}\n{trace}")
-        return f"❌ Fehler beim Laden des Tokenizers: {str(e)}"
 
-# Then, modify your extract_info function to load the model on first use
-@spaces.GPU
-@spaces.GPU
 def extract_info(template, text):
     global tokenizer, model