Refactoring the model to be a function not a generator

app.py

@@ -77,10 +77,14 @@ def spaces_model_predict(message: str, history: list[tuple[str, str]]):
     t = Thread(target=model.generate, kwargs=generate_kwargs)
     t.start()
 
-    partial_message = ""
+    generated_text = ""
     for new_token in streamer:
-        partial_message += new_token
-        yield partial_message
+        generated_text += new_token
+        if "<|user|>" in generated_text:
+            generated_text = generated_text.split("<|user|>")[0].strip()
+            break
+
+    return generated_text
 
 
 def predict(