Describe the bug

We wanted to pass custom limits to the httpx client, so that the number of connections and max keepalive connections are changed. As bentoml does not support this, we subclasses httpx.AsyncClient and set the limits there. We apply the limits by overriding client_cls, but this does not work.

Bentoml uses client_cls is httpx.Client and client_cls is httpx.AsyncClient several times, but MyHttpxClient is not httpx.AsyncClient although it inherits from it. If bentoml would use issubclass, then everything would be fine.

Please not that this is not directly a feature request for custom limits directly (that would be also great!), but rather a request for extensibility of the bentoml client.

To reproduce

class MyHttpxClient(httpx.AsyncClient): ...


class MyClient(bentoml.AsyncHTTPClient):
    client_cls = MyHttpxClient


async def test_it() -> None:
    async with MyClient(...) as client:
        await client.predict(...)

This fails with RuntimeError: Attempted to send an sync request with an AsyncClient instance. because of

Expected behavior

Bentoml uses the overridden client class and it works.

Environment

bentoml: a567a3a
python: 3.12
OS: Arch Linux