Fix "Event loop Closed" error (#188)

By default httpx will keep the last few connections open in a pool for
re-use. Since we create the httpx connection inside an asyncio task, it
is bound to the lifetime of the task rather than the httpx client. Thus
when the client attempts to reuse the connection in a new task it fails.
There may be performance implications for this change, but since the
default is only to re-use the last 5 connections, its likely minor.
([ref](encode/httpx#2959))

Fixes: #147

---------

Co-authored-by: Mark Kurtz <mark.j.kurtz@gmail.com>

Author: sjmonson

src/guidellm/backend/backend.py

@@ -110,6 +110,14 @@ def info(self) -> dict[str, Any]:
         """
         ...
 
+    @abstractmethod
+    async def reset(self) -> None:
+        """
+        Reset the connection object. This is useful for backends that
+        reuse connections or have state that needs to be cleared.
+        """
+        ...
+
     async def validate(self):
         """
         Handle final setup and validate the backend is ready for use.
@@ -126,6 +134,8 @@ async def validate(self):
         ):  # type: ignore[attr-defined]
             pass
 
+        await self.reset()
+
     @abstractmethod
     async def check_setup(self):
         """

src/guidellm/backend/openai.py

@@ -157,6 +157,15 @@ def info(self) -> dict[str, Any]:
             "chat_completions_path": CHAT_COMPLETIONS_PATH,
         }
 
+    async def reset(self) -> None:
+        """
+        Reset the connection object. This is useful for backends that
+        reuse connections or have state that needs to be cleared.
+        For this backend, it closes the async client if it exists.
+        """
+        if self._async_client is not None:
+            await self._async_client.aclose()
+
     async def check_setup(self):
         """
         Check if the backend is setup correctly and can be used for requests.
@@ -361,7 +370,7 @@ def _get_async_client(self) -> httpx.AsyncClient:
 
         :return: The async HTTP client.
         """
-        if self._async_client is None:
+        if self._async_client is None or self._async_client.is_closed:
             client = httpx.AsyncClient(
                 http2=self.http2,
                 timeout=self.timeout,

tests/unit/mock_backend.py

@@ -41,6 +41,9 @@ def model(self) -> Optional[str]:
     def info(self) -> dict[str, Any]:
         return {}
 
+    async def reset(self) -> None:
+        pass
+
     async def prepare_multiprocessing(self):
         pass