Update app.py

app.py

@@ -42,7 +42,6 @@ model_configs = [
     {"repo_id": "Ffftdtd5dtft/Mistral-Nemo-Instruct-2407-Q2_K-GGUF", "filename": "mistral-nemo-instruct-2407-q2_k.gguf", "name": "Mistral Nemo Instruct 2407"}
 ]
 
-
 class ModelManager:
     def __init__(self):
         self.models = {}
@@ -113,14 +112,11 @@ async def handle_request(request):
 
 
 async def run_server(port: int):
-    async with httpx.AsyncClient(base_url=f"http://localhost:{port}") as client:
-        while True:
-            request = await client.get("/")  # You might need to adjust this based on your expected requests
-            response = await handle_request(request)
-            print(f"Received request: {request}")
-            print(f"Sending response: {response}")
-            await asyncio.sleep(1)  # Adjust the sleep duration as needed
+    async def serve_request(request: httpx.Request) -> httpx.Response:
+        return await handle_request(request)
 
+    server = await httpx.serve(serve_request, host="localhost", port=port) 
+    await server.aclose()
 
 def remove_repetitive_responses(responses):
     unique_responses = {}