Changing the loop methodology

main.py

@@ -65,10 +65,15 @@ def _generate(history):
         add_generation_prompt=True,
     )
 
+    try:
+        loop = asyncio.get_event_loop()
+    except:
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+
     model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
     streamer = AsyncTextIteratorStreamer(tokenizer, skip_special_tokens=True)
 
-    loop = asyncio.new_event_loop()
     task = loop.run_in_executor(
         None,
         functools.partial(
@@ -78,7 +83,7 @@ def _generate(history):
             **model_inputs,
         ),
     )
-    return loop, task, streamer
+    return task, streamer
 
 
 async def chat(prompt, history):
@@ -92,7 +97,7 @@ async def chat(prompt, history):
     history = [] if history is None else history
     message_list = history + [message]
 
-    loop, task, streamer = _generate(message_list)
+    task, streamer = _generate(message_list)
 
     buffer = ""
     reasoning = ""
@@ -127,8 +132,6 @@ async def chat(prompt, history):
         print("Signal sent")
         raise
 
-    loop.close()
-
 
 chat_bot = gr.Chatbot(
     latex_delimiters=[