Update app.py

app.py

@@ -263,16 +263,25 @@ def simple_chat(message, temperature: float = 0.8, max_length: int = 4096, top_p
     def generate_text():
         nonlocal buffer
         with torch.no_grad():
-            thread = Thread(target=model.generate, kwargs=gen_kwargs)
-            thread.start()
+            model.generate(**gen_kwargs)
             for new_text in streamer:
                 buffer += new_text
+            # thread = Thread(target=model.generate, kwargs=gen_kwargs)
+            # thread.start()
+            # for new_text in streamer:
+            #     buffer += new_text
 
-    # Start the generation in a separate thread
-    generate_text()
     
-    # Wait for the generation to finish
+    thread = Thread(target=generate_text)
+    thread.start()
+
     thread.join()
+    
+    # # Start the generation in a separate thread
+    # generate_text()
+    
+    # # Wait for the generation to finish
+    # thread.join()
 
     print("---------")
     print("Text: ")