Skip to content

The fastest, litest AI Gateway. Rust core with Python SDK. Call 100+ LLM APIs in OpenAI (or native) format with cost tracking, guardrails, load balancing, and logging [Bedrock, Azure, OpenAI, Anthropic, OpenAI, VertexAI, vLLM, Nvidia NIM]

5.5 KB
Raw Blame
import concurrent.futures
import json
from typing import Final

import httpx
import litellm
import pytest
import respx
from litellm.batch_completion.main import batch_completion_models_all_responses


def test_batch_completion_models_all_responses_submits_before_waiting(monkeypatch):
    """
    Regression test for issue #20704.
    Ensures all model calls are submitted to the thread pool before waiting on results.
    """
    models = ["model-a", "model-b", "model-c"]
    called_models = []

    class _AssertingFuture:
        def __init__(self, result, executor, expected_submissions):
            self._result = result
            self._executor = executor
            self._expected_submissions = expected_submissions

        def result(self):
            if self._executor.submit_count != self._expected_submissions:
                raise AssertionError(
                    "Not all model calls were submitted before waiting"
                )
            return self._result

    class _RecordingThreadPoolExecutor:
        def __init__(self, max_workers, *args, **kwargs):
            self.max_workers = max_workers
            self.submit_count = 0

        def __enter__(self):
            return self

        def __exit__(self, exc_type, exc, tb):
            return False

        def submit(self, fn, *args, **kwargs):
            self.submit_count += 1
            result = fn(*args, **kwargs)
            return _AssertingFuture(
                result=result,
                executor=self,
                expected_submissions=len(models),
            )

    def _mock_completion(*args, model, **kwargs):
        called_models.append(model)
        return {"model": model}

    monkeypatch.setattr(litellm, "completion", _mock_completion)
    monkeypatch.setattr(
        concurrent.futures, "ThreadPoolExecutor", _RecordingThreadPoolExecutor
    )

    responses = batch_completion_models_all_responses(
        models=models,
        messages=[{"role": "user", "content": "hello"}],
    )

    assert sorted(called_models) == sorted(models)
    assert len(responses) == len(models)
    assert sorted(response["model"] for response in responses) == sorted(models)


def test_batch_completion_models_all_responses_continues_on_model_error(monkeypatch):
    models = ["model-a", "model-error", "model-b"]

    def _mock_completion(*args, model, **kwargs):
        if model == "model-error":
            raise RuntimeError("simulated model failure")
        return {"model": model}

    monkeypatch.setattr(litellm, "completion", _mock_completion)

    responses = batch_completion_models_all_responses(
        models=models,
        messages=[{"role": "user", "content": "hello"}],
    )

    assert len(responses) == 2
    assert sorted(response["model"] for response in responses) == ["model-a", "model-b"]


def test_batch_completion_models_all_responses_returns_empty_for_empty_models(
    monkeypatch,
):
    called = False

    def _mock_completion(*args, model, **kwargs):
        nonlocal called
        called = True
        return {"model": model}

    monkeypatch.setattr(litellm, "completion", _mock_completion)

    responses = batch_completion_models_all_responses(
        models=[],
        messages=[{"role": "user", "content": "hello"}],
    )

    assert responses == []
    assert called is False


def test_batch_completion_models_all_responses_accepts_single_model_string(monkeypatch):
    called_models = []

    def _mock_completion(*args, model, **kwargs):
        called_models.append(model)
        return {"model": model}

    monkeypatch.setattr(litellm, "completion", _mock_completion)

    responses = batch_completion_models_all_responses(
        models="model-a",
        messages=[{"role": "user", "content": "hello"}],
    )

    assert called_models == ["model-a"]
    assert responses == [{"model": "model-a"}]


def test_batch_completion_models_all_responses_returns_each_model_response(
    respx_mock: respx.MockRouter,
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)

    def _response_for_model(request: httpx.Request) -> httpx.Response:
        requested_model: Final = str(json.loads(request.content)["model"])
        return httpx.Response(
            200,
            json={
                "id": f"chatcmpl-{requested_model}",
                "object": "chat.completion",
                "created": 1,
                "model": requested_model,
                "choices": [
                    {
                        "index": 0,
                        "message": {"role": "assistant", "content": f"answer from {requested_model}"},
                        "finish_reason": "stop",
                    }
                ],
                "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
            },
        )

    route: Final = respx_mock.post("https://api.openai.com/v1/chat/completions").mock(
        side_effect=_response_for_model
    )
    models: Final = ("openai/model-a", "openai/model-b")
    responses: Final = batch_completion_models_all_responses(
        models=models,
        messages=[{"role": "user", "content": "hello"}],
        api_key="test-key",
    )

    assert route.call_count == 2
    assert tuple(sorted(str(json.loads(call.request.content)["model"]) for call in route.calls)) == (
        "model-a",
        "model-b",
    )
    assert tuple(sorted(response.model for response in responses)) == ("model-a", "model-b")
    assert tuple(sorted(response.choices[0].message.content for response in responses)) == (
        "answer from model-a",
        "answer from model-b",
    )