Update app.py

app.py

@@ -21,7 +21,10 @@ def model_inference(
     input_dict, history, decoding_strategy, temperature, max_new_tokens,
     repetition_penalty, top_p, progress=gr.Progress(track_tqdm=True)
 ): 
-    print("Processing")
+    buffer = "Thinking... "
+
+    yield buffer
+    
     t0 = time.perf_counter()
     text = input_dict["text"]
     print(input_dict["files"])
@@ -79,8 +82,6 @@ def model_inference(
     thread.start()
     thread.join()
     
-    buffer = ""
-    
       
     for new_text in streamer: