Update app.py

app.py

@@ -178,11 +178,15 @@ with gr.Blocks() as demo:
             keep_streaming = True
             s = requests.Session()
             with s.post(API_URL, headers=HEADERS, json=data, stream=True, timeout=30) as r:
-                time.sleep(0.3)
+                time.sleep(0.1)
                 for line in r.iter_lines():
+                    if time.time() - start_time > MAX_SEC:
+                        keep_streaming = False
+                        break
+                    
                     if line and keep_streaming:
                         if r.status_code != 200:
-                            raise Exception('status_code != 200')
+                            continue
                         json_response = json.loads(line)
         
                         try:
@@ -199,10 +203,6 @@ with gr.Blocks() as demo:
                         else:
                             history[-1][1] += delta
                         yield history
-
-                    if time.time() - start_time > MAX_SEC:
-                        keep_streaming = False
-                        break
                     
             if history[-1][1].endswith('</s>'):
                 history[-1][1] = history[-1][1][:-4]