Merge 214ca1c205 into 719f29fb29

api/responder.py

@@ -7,6 +7,7 @@ import aiohttp
 import asyncio
 import starlette
 
+from typing import Set
 from rich import print
 from dotenv import load_dotenv
 
@@ -23,6 +24,8 @@ CRITICAL_API_ERRORS = ['invalid_api_key', 'account_deactivated']
 
 keymanager = providerkeys.manager
 
+background_tasks: Set[asyncio.Task] = set()
+
 async def respond(
     path: str='/v1/chat/completions',
     user: dict=None,
@@ -166,7 +169,7 @@ async def respond(
     if (not is_stream) and server_json_response:
         yield json.dumps(server_json_response)
 
-    asyncio.create_task(
+    task = asyncio.create_task(
         after_request.after_request(
             incoming_request=incoming_request,
             target_request=target_request,
@@ -178,3 +181,5 @@ async def respond(
             model=model,
         )
     )
+    background_tasks.add(task)
+    task.add_done_callback(background_tasks.discard)