Update app.py

app.py

@@ -8,6 +8,7 @@ import time
 import os
 from typing import Iterator, List, Tuple
 import spaces
+import threading
 
 
 
@@ -133,7 +134,9 @@ class SinaReasonMedicalChat:
 
         # Start generation directly.
         # This will return immediately and the streamer will be populated in the background.
-        self.model.generate(**generation_kwargs)
+        #self.model.generate(**generation_kwargs)
+        thread = threading.Thread(target=self.model.generate, kwargs=generation_kwargs)
+        thread.start()
 
         # Stream the response
         partial_response = ""
@@ -143,7 +146,7 @@ class SinaReasonMedicalChat:
         for new_token in streamer:
             partial_response += new_token
 
-            prinr(partial_response)
+            print(partial_response)
             # Extract thinking and response
             thinking, response = self.extract_thinking_and_response(partial_response)