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]

23.8 KB
Raw Blame
#!/usr/bin/env python3
"""
Test to verify the Google GenAI generate_content handler functionality
"""

import asyncio
import json
import time
from unittest.mock import AsyncMock, MagicMock, patch
from typing import Final

import httpx
import pytest
import respx

import litellm
from litellm.integrations.custom_logger import CustomLogger

from litellm.google_genai.adapters.handler import GenerateContentToCompletionHandler
from litellm.google_genai.adapters.transformation import GoogleGenAIAdapter, GoogleGenAIStreamWrapper
from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices, Usage


def test_stream_response_when_stream_requested_sync():
    """
    Test that when a stream response is returned and streaming was requested,
    the sync handler correctly transforms it to generate_content streaming format.
    """
    # Mock a stream response
    mock_stream = MagicMock()
    mock_stream.__iter__ = MagicMock(return_value=iter([]))

    # Mock the GoogleGenAIAdapter's translate_completion_output_params_streaming method
    with patch.object(
        GoogleGenAIAdapter,
        "translate_completion_output_params_streaming",
        return_value=mock_stream,
    ) as mock_translate:
        with patch("litellm.completion", return_value=mock_stream):
            # Call the handler with stream=True
            result = GenerateContentToCompletionHandler.generate_content_handler(
                model="gemini-pro",
                contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                litellm_params={},  # Empty dict for params
                stream=True,
            )

            # Verify that translate_completion_output_params_streaming was called
            mock_translate.assert_called_once_with(mock_stream)
            # Verify the result is the transformed stream
            assert result == mock_stream


@pytest.mark.asyncio
async def test_stream_response_when_stream_requested_async():
    """
    Test that when a stream response is returned and streaming was requested,
    the async handler correctly transforms it to generate_content streaming format.
    """
    # Mock a stream response
    mock_stream = MagicMock()
    mock_stream.__aiter__ = AsyncMock(return_value=iter([]))  # Return an empty async iterator

    # Mock the GoogleGenAIAdapter's translate_completion_output_params_streaming method
    with patch.object(
        GoogleGenAIAdapter,
        "translate_completion_output_params_streaming",
        return_value=mock_stream,
    ) as mock_translate:
        with patch("litellm.acompletion", return_value=mock_stream):
            # Call the handler with stream=True
            result = await GenerateContentToCompletionHandler.async_generate_content_handler(
                model="gemini-pro",
                contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                litellm_params={},  # Empty dict for params
                stream=True,
            )

            # Verify that translate_completion_output_params_streaming was called
            mock_translate.assert_called_once_with(mock_stream)
            # Verify the result is the transformed stream
            assert result == mock_stream


def test_stream_transformation_error_sync():
    """
    Test that when a stream transformation fails, the sync handler raises a ValueError.
    """
    # Mock a stream response
    mock_stream = MagicMock()
    mock_stream.__iter__ = MagicMock(return_value=iter([]))

    # Mock the GoogleGenAIAdapter's translate_completion_output_params_streaming method to return None
    with patch.object(
        GoogleGenAIAdapter,
        "translate_completion_output_params_streaming",
        return_value=None,
    ):
        # Patch litellm.completion directly to prevent real API calls
        with patch("litellm.completion", return_value=mock_stream):
            # Call the handler with stream=True and expect a ValueError
            with pytest.raises(ValueError, match="Failed to transform streaming response"):
                GenerateContentToCompletionHandler.generate_content_handler(
                    model="gemini-pro",
                    contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                    litellm_params={},  # Empty dict for params
                    stream=True,
                )


@pytest.mark.asyncio
async def test_stream_transformation_error_async():
    """
    Test that when a stream transformation fails, the async handler raises a ValueError.
    """
    # Mock a stream response
    mock_stream = MagicMock()
    mock_stream.__aiter__ = AsyncMock(return_value=mock_stream)

    # Mock the GoogleGenAIAdapter's translate_completion_output_params_streaming method to return None
    with patch.object(
        GoogleGenAIAdapter,
        "translate_completion_output_params_streaming",
        return_value=None,
    ):
        # Mock litellm.acompletion at the module level where it's imported
        # We need to patch it in the handler module, not in litellm itself
        with patch("litellm.google_genai.adapters.handler.litellm") as mock_litellm:
            # Use AsyncMock for async function
            mock_litellm.acompletion = AsyncMock(return_value=mock_stream)
            # Call the handler with stream=True and expect a ValueError
            with pytest.raises(ValueError, match="Failed to transform streaming response"):
                await GenerateContentToCompletionHandler.async_generate_content_handler(
                    model="gemini-pro",
                    contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                    litellm_params={},  # Empty dict for params
                    stream=True,
                )


def test_citation_metadata_transformation():
    """
    Test that citationMetadata.citationSources is properly transformed to citationMetadata.citations
    to avoid Pydantic validation errors.
    """
    from unittest.mock import MagicMock

    import httpx

    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
    from litellm.llms.gemini.google_genai.transformation import GoogleGenAIConfig

    # Create a mock response with citationMetadata.citationSources (the problematic format)
    mock_response_data = {
        "candidates": [
            {
                "content": {
                    "parts": [{"text": "This is a video analysis response with citation metadata."}],
                    "role": "model",
                },
                "finishReason": "STOP",
                "index": 0,
                "safetyRatings": [],
                "citationMetadata": {
                    "citationSources": [
                        {
                            "startIndex": 5848,
                            "endIndex": 5900,
                            "uri": "https://example.com/video-source",
                            "license": "MIT",
                            "title": "Video Analysis Source",
                            "publicationDate": "2024-01-15",
                        },
                        {
                            "startIndex": 6200,
                            "endIndex": 6250,
                            "uri": "https://another-source.com/reference",
                            "license": "CC-BY",
                            "title": "Another Reference",
                            "publicationDate": "2024-02-01",
                        },
                    ]
                },
            }
        ],
        "usageMetadata": {
            "promptTokenCount": 150,
            "candidatesTokenCount": 200,
            "totalTokenCount": 350,
        },
        "responseId": "test-response-123",
    }

    # Create mock httpx response
    mock_httpx_response = MagicMock(spec=httpx.Response)
    mock_httpx_response.json.return_value = mock_response_data
    mock_httpx_response.status_code = 200
    mock_httpx_response.headers = {}

    # Create logging object
    logging_obj = LiteLLMLoggingObj(
        model="gemini-2.5-flash",
        messages=[],
        stream=False,
        call_type="generate_content",
        start_time=1234567890,
        litellm_call_id="test-call-123",
        function_id="test-function-123",
    )

    # Create GoogleGenAI config
    config = GoogleGenAIConfig()

    # Test the transformation
    try:
        result = config.transform_generate_content_response(
            model="gemini-2.5-flash",
            raw_response=mock_httpx_response,
            logging_obj=logging_obj,
        )

        # Verify the transformation worked
        assert result is not None

        # Check that citationSources was transformed to citations
        if hasattr(result, "candidates") and result.candidates:
            candidate = result.candidates[0]
            if hasattr(candidate, "citationMetadata") and candidate.citationMetadata:
                # The citationMetadata should now have 'citations' instead of 'citationSources'
                citation_metadata = candidate.citationMetadata

                # Check that citations field exists
                assert hasattr(citation_metadata, "citations"), "citations field should exist after transformation"

                # Verify the citations data is preserved
                if hasattr(citation_metadata, "citations") and citation_metadata.citations:
                    assert len(citation_metadata.citations) == 2, "Should have 2 citations"
                    assert citation_metadata.citations[0]["uri"] == "https://example.com/video-source"
                    assert citation_metadata.citations[1]["uri"] == "https://another-source.com/reference"

        print("✅ Citation metadata transformation test passed!")

    except Exception as e:
        pytest.fail(f"Citation metadata transformation failed: {e}")


@pytest.mark.asyncio
async def test_generate_content_adapter_preserves_proxy_server_request():
    """
    Ensure GenerateContentToCompletionHandler forwards proxy_server_request
    to the downstream completion call so proxy spend logging captures the request body.
    """
    from litellm.types.router import GenericLiteLLMParams
    from litellm.types.utils import Choices, Message, ModelResponse

    handler = GenerateContentToCompletionHandler()

    dummy_proxy_request: dict[str, object] = {
        "url": "http://localhost:4000/v1beta/models/gemini-2.0-flash:generateContent",
        "method": "POST",
        "headers": {"content-type": "application/json"},
        "body": {"contents": [{"role": "user", "parts": [{"text": "Hello, world!"}]}]},
    }

    gemini_data: list[dict[str, object]] = [{"role": "user", "parts": [{"text": "Hello, world!"}]}]

    mock_response = ModelResponse(choices=[Choices(message=Message(content="Hi!", role="assistant"))])

    with patch(
        "litellm.google_genai.adapters.handler.litellm.acompletion",
        new_callable=AsyncMock,
    ) as mock_acompletion:
        mock_acompletion.return_value = mock_response

        await handler.async_generate_content_handler(
            model="gemini-2.0-flash",
            contents=gemini_data,
            litellm_params=GenericLiteLLMParams(),
            proxy_server_request=dummy_proxy_request,
            metadata={"source": "unit_test"},
        )

        assert mock_acompletion.called, "Inner acompletion was not called"
        called_kwargs = mock_acompletion.call_args.kwargs

        assert "proxy_server_request" in called_kwargs, "proxy_server_request was dropped from completion_kwargs"
        assert called_kwargs["proxy_server_request"] == dummy_proxy_request


_CHAT_SSE = (
    'data: {"id":"c","object":"chat.completion.chunk","created":1,"model":"m",'
    '"choices":[{"index":0,"delta":{"role":"assistant","content":"144"},"finish_reason":null}]}\n\n'
    'data: {"id":"c","object":"chat.completion.chunk","created":1,"model":"m",'
    '"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}\n\n'
    "data: [DONE]\n\n"
)
_CHAT_BODY: dict[str, object] = {
    "id": "c",
    "object": "chat.completion",
    "created": 1,
    "model": "m",
    "choices": [{"index": 0, "message": {"role": "assistant", "content": "144"}, "finish_reason": "stop"}],
    "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4},
}


_CHAT_SSE_WITH_USAGE = _CHAT_SSE.replace(
    "data: [DONE]",
    'data: {"id":"c","object":"chat.completion.chunk","created":1,"model":"m","choices":[],'
    '"usage":{"prompt_tokens":340,"completion_tokens":1,"total_tokens":341}}\n\n'
    "data: [DONE]",
)

_ANTHROPIC_SSE: Final = "".join(
    f"event: {event}\ndata: {json.dumps(data)}\n\n"
    for event, data in (
        (
            "message_start",
            {
                "type": "message_start",
                "message": {
                    "id": "msg_1",
                    "type": "message",
                    "role": "assistant",
                    "model": "claude-haiku-4-5",
                    "content": [],
                    "stop_reason": None,
                    "usage": {"input_tokens": 20, "output_tokens": 1},
                },
            },
        ),
        (
            "content_block_start",
            {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}},
        ),
        (
            "content_block_delta",
            {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "144"}},
        ),
        ("content_block_stop", {"type": "content_block_stop", "index": 0}),
        (
            "message_delta",
            {"type": "message_delta", "delta": {"stop_reason": "end_turn"}, "usage": {"output_tokens": 5}},
        ),
        ("message_stop", {"type": "message_stop"}),
    )
)


class _FailureCallTypes(CustomLogger):
    def __init__(self) -> None:
        super().__init__()
        self.call_types: list[str] = []

    def log_failure_event(self, kwargs, response_obj, start_time, end_time):
        self.call_types.append(kwargs["call_type"])

    async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time):
        self.call_types.append(kwargs["call_type"])


class _SuccessPayloads(CustomLogger):
    def __init__(self) -> None:
        super().__init__()
        self.model_groups: list[str | None] = []

    async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
        self.model_groups.append(kwargs["standard_logging_object"]["model_group"])


@pytest.fixture
def _httpx_transport(monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)


async def _drain(stream: object) -> None:
    async for _ in stream:
        pass


@pytest.mark.asyncio
async def test_generate_content_adapter_keeps_provider_error_status(_httpx_transport: None):
    from litellm.types.router import GenericLiteLLMParams

    with respx.mock() as upstream:
        upstream.post("http://upstream.test/v1/chat/completions").mock(
            return_value=httpx.Response(400, json={"error": {"message": "max_tokens must be positive"}})
        )
        with pytest.raises(litellm.BadRequestError) as exc_info:
            await GenerateContentToCompletionHandler.async_generate_content_handler(
                model="hosted_vllm/grok",
                contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                litellm_params=GenericLiteLLMParams(api_base="http://upstream.test/v1", api_key="k"),
                config={"maxOutputTokens": -5},
            )

    assert exc_info.value.status_code == 400


@pytest.mark.asyncio
async def test_generate_content_adapter_logs_router_model_group(
    _httpx_transport: None, monkeypatch: pytest.MonkeyPatch
):
    from litellm.types.router import GenericLiteLLMParams

    success_payloads = _SuccessPayloads()
    monkeypatch.setattr(litellm, "callbacks", [success_payloads])

    with respx.mock() as upstream:
        upstream.post("http://upstream.test/v1/chat/completions").mock(
            return_value=httpx.Response(200, json=_CHAT_BODY)
        )
        await GenerateContentToCompletionHandler.async_generate_content_handler(
            model="hosted_vllm/grok",
            contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
            litellm_params=GenericLiteLLMParams(api_base="http://upstream.test/v1", api_key="k"),
            litellm_metadata={"model_group": "grok-group"},
        )
    await asyncio.sleep(0.5)

    assert "grok-group" in success_payloads.model_groups


@pytest.mark.parametrize(
    ("stream", "requested_stream_options", "expected_stream_options"),
    [
        (True, None, {"include_usage": True}),
        (True, {"include_usage": True}, {"include_usage": True}),
        (True, {"include_usage": False}, None),
        (False, None, None),
    ],
)
@pytest.mark.asyncio
async def test_generate_content_adapter_asks_openai_compatible_upstream_for_stream_usage(
    _httpx_transport: None,
    stream: bool,
    requested_stream_options: dict[str, bool] | None,
    expected_stream_options: dict[str, bool] | None,
):
    from litellm.types.router import GenericLiteLLMParams

    with respx.mock() as upstream:
        route = upstream.post("http://upstream.test/v1/chat/completions").mock(
            return_value=(
                httpx.Response(200, text=_CHAT_SSE, headers={"content-type": "text/event-stream"})
                if stream
                else httpx.Response(200, json=_CHAT_BODY)
            )
        )
        response = await GenerateContentToCompletionHandler.async_generate_content_handler(
            model="hosted_vllm/grok",
            contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
            litellm_params=GenericLiteLLMParams(api_base="http://upstream.test/v1", api_key="k"),
            stream=stream,
            **({} if requested_stream_options is None else {"stream_options": requested_stream_options}),
        )
        if stream:
            await _drain(response)

    assert json.loads(route.calls[0].request.content).get("stream_options") == expected_stream_options


@pytest.mark.parametrize(
    ("model", "api_base", "upstream_url", "upstream_sse", "prompt_tokens"),
    [
        (
            "hosted_vllm/grok",
            "http://upstream.test/v1",
            "http://upstream.test/v1/chat/completions",
            _CHAT_SSE_WITH_USAGE,
            340,
        ),
        ("anthropic/claude-haiku-4-5", "http://upstream.test", "http://upstream.test/v1/messages", _ANTHROPIC_SSE, 20),
    ],
    ids=["openai_compatible", "anthropic"],
)
@pytest.mark.asyncio
async def test_generate_content_adapter_stream_ends_with_upstream_usage(
    _httpx_transport: None, model: str, api_base: str, upstream_url: str, upstream_sse: str, prompt_tokens: int
):
    payloads: Final = await _adapter_stream_payloads(model, api_base, upstream_url, upstream_sse)

    assert payloads[-1]["candidates"][0]["finishReason"] == "STOP"
    assert payloads[-1]["usageMetadata"]["promptTokenCount"] == prompt_tokens


@pytest.mark.parametrize(
    ("model", "api_base", "upstream_url", "upstream_sse"),
    [
        ("hosted_vllm/grok", "http://upstream.test/v1", "http://upstream.test/v1/chat/completions", _CHAT_SSE),
        ("anthropic/claude-haiku-4-5", "http://upstream.test", "http://upstream.test/v1/messages", _ANTHROPIC_SSE),
    ],
    ids=["openai_compatible", "anthropic"],
)
@pytest.mark.asyncio
async def test_generate_content_adapter_stream_without_usage_still_ends_with_stop(
    _httpx_transport: None, model: str, api_base: str, upstream_url: str, upstream_sse: str
):
    payloads: Final = await _adapter_stream_payloads(
        model, api_base, upstream_url, upstream_sse, stream_options={"include_usage": False}
    )

    assert [payload["candidates"][0].get("finishReason") for payload in payloads][-1] == "STOP"


async def _adapter_stream_payloads(
    model: str, api_base: str, upstream_url: str, upstream_sse: str, **kwargs: object
):
    from litellm.types.router import GenericLiteLLMParams

    with respx.mock() as upstream:
        upstream.post(upstream_url).mock(
            return_value=httpx.Response(200, text=upstream_sse, headers={"content-type": "text/event-stream"})
        )
        response = await GenerateContentToCompletionHandler.async_generate_content_handler(
            model=model,
            contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
            litellm_params=GenericLiteLLMParams(api_base=api_base, api_key="k"),
            stream=True,
            **kwargs,
        )
        events: Final = [chunk async for chunk in response]

    return [
        json.loads(line.removeprefix("data: "))
        for line in b"".join(events).decode().splitlines()
        if line.startswith("data: {")
    ]


@pytest.mark.parametrize("entrypoint", ["agenerate_content", "agenerate_content_stream"])
@pytest.mark.asyncio
async def test_generate_content_adapter_failure_is_logged_once(
    _httpx_transport: None, monkeypatch: pytest.MonkeyPatch, entrypoint: str
):
    failures = _FailureCallTypes()
    monkeypatch.setattr(litellm, "callbacks", [failures])

    with respx.mock() as upstream:
        upstream.post("http://upstream.test/v1/chat/completions").mock(
            return_value=httpx.Response(400, json={"error": {"message": "max_tokens must be positive"}})
        )
        with pytest.raises(litellm.BadRequestError):
            await getattr(litellm.google_genai, entrypoint)(
                model="hosted_vllm/grok",
                contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                api_base="http://upstream.test/v1",
                api_key="k",
            )
    await asyncio.sleep(0.5)

    assert failures.call_types == ["acompletion"]


def test_sync_generate_content_adapter_failure_is_logged_once(_httpx_transport: None, monkeypatch: pytest.MonkeyPatch):
    failures = _FailureCallTypes()
    monkeypatch.setattr(litellm, "callbacks", [failures])

    with respx.mock() as upstream:
        upstream.post("http://upstream.test/v1/chat/completions").mock(
            return_value=httpx.Response(400, json={"error": {"message": "max_tokens must be positive"}})
        )
        with pytest.raises(litellm.BadRequestError):
            litellm.google_genai.generate_content(
                model="hosted_vllm/grok",
                contents=[{"role": "user", "parts": [{"text": "Hello"}]}],
                api_base="http://upstream.test/v1",
                api_key="k",
            )
    time.sleep(0.5)

    assert failures.call_types == ["completion"]


def _stream_chunk(content: str | None, finish_reason: str | None, usage: Usage | None = None) -> ModelResponseStream:
    return ModelResponseStream(
        choices=[StreamingChoices(index=0, delta=Delta(content=content), finish_reason=finish_reason)],
        **({} if usage is None else {"usage": usage}),
    )


_FINISH_WITHOUT_USAGE: Final = (_stream_chunk("144", None), _stream_chunk(None, "stop"))
_FINISH_WITH_USAGE: Final = (
    _stream_chunk("144", None),
    _stream_chunk(None, "stop", Usage(prompt_tokens=20, completion_tokens=1, total_tokens=21)),
)


@pytest.mark.parametrize(
    ("chunks", "prompt_tokens"), [(_FINISH_WITHOUT_USAGE, 0), (_FINISH_WITH_USAGE, 20)], ids=["no_usage", "usage"]
)
def test_stream_wrapper_iteration_ends_with_one_stop_event(chunks: tuple[ModelResponseStream, ...], prompt_tokens: int):
    events: Final = list(GoogleGenAIStreamWrapper(completion_stream=iter(chunks)))

    assert [event["candidates"][0].get("finishReason") for event in events] == [None, "STOP"]
    assert events[-1]["usageMetadata"]["promptTokenCount"] == prompt_tokens


@pytest.mark.parametrize(
    ("chunks", "prompt_tokens"), [(_FINISH_WITHOUT_USAGE, 0), (_FINISH_WITH_USAGE, 20)], ids=["no_usage", "usage"]
)
@pytest.mark.asyncio
async def test_stream_wrapper_async_iteration_ends_with_one_stop_event(
    chunks: tuple[ModelResponseStream, ...], prompt_tokens: int
):
    async def completion_stream():
        for chunk in chunks:
            yield chunk

    events: Final = [event async for event in GoogleGenAIStreamWrapper(completion_stream=completion_stream())]

    assert [event["candidates"][0].get("finishReason") for event in events] == [None, "STOP"]
    assert events[-1]["usageMetadata"]["promptTokenCount"] == prompt_tokens