fix argument passing

utils/models.py

@@ -153,7 +153,7 @@ def run_inference(model_name, context, question):
         if generation_interrupt.is_set():
             return ""
 
-        outputs = pipe(text_input, max_new_tokens=512, generate_kwargs=generation_kwargs)
+        outputs = pipe(text_input, **generation_kwargs)
         result = outputs[0]['generated_text'][-1]['content']
 
     except Exception as e: