Update app.py

app.py

@@ -79,6 +79,11 @@ def safe_load_pretrained_model(model_path, model_base=None, model_name=None, **k
             tokenizer.add_special_tokens({'pad_token': '[PAD]'})
             print('[INFO] Added [PAD] token to tokenizer')
 
+    # Set padding_side to left for Llama 3 (prevents empty generation)
+    if hasattr(tokenizer, 'padding_side'):
+        tokenizer.padding_side = "left"
+        print('[INFO] Set tokenizer.padding_side to "left" for proper generation')
+
     # Force all model components to CPU (keep original dtype if possible, fallback to float32)
     print('[INFO] Ensuring all components are on CPU...')
     try: