Merge remote-tracking branch 'origin/main' into litellm_migrate_tests_p12

tests/test_litellm/llms/openai/evals/__init__.py

@@ -1 +0,0 @@
-"""OpenAI Evals API tests"""

tests/test_litellm/llms/openai_like/embedding/__init__.py

@@ -1 +0,0 @@
-# Test module for OpenAI-like embedding handler

tests/test_litellm/llms/watsonx/audio_transcription/test_watsonx_audio_transcription_transformation.py

@@ -1,337 +0,0 @@
-"""
-Tests for IBM WatsonX Audio Transcription.
-
-Validates that litellm.transcription transforms requests correctly for WatsonX.
-"""
-
-import json
-from unittest.mock import AsyncMock, MagicMock, patch
-
-import pytest
-
-
-import litellm
-from litellm.llms.watsonx.audio_transcription.transformation import (
-    IBMWatsonXAudioTranscriptionConfig,
-)
-from litellm.types.utils import TranscriptionResponse
-
-
-class TestWatsonXAudioTranscription:
-    """Tests for WatsonX audio transcription via litellm.transcription."""
-
-    @pytest.mark.asyncio
-    async def test_watsonx_transcription_url_and_headers(self):
-        """
-        Test that litellm.transcription sends request to correct WatsonX URL with proper headers.
-        """
-        captured_request = {}
-
-        async def mock_post(*args, **kwargs):
-            captured_request["url"] = str(kwargs.get("url", args[0] if args else None))
-            captured_request["headers"] = kwargs.get("headers", {})
-            captured_request["data"] = kwargs.get("data", {})
-            captured_request["files"] = kwargs.get("files", {})
-
-            mock_response = MagicMock()
-            mock_response.json.return_value = {
-                "text": "test transcription",
-                "duration": 1.0,
-            }
-            mock_response.status_code = 200
-            return mock_response
-
-        with patch(
-            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
-            new=mock_post,
-        ):
-            try:
-                await litellm.atranscription(
-                    model="watsonx/whisper-large-v3-turbo",
-                    file=b"fake_audio_data",
-                    api_base="https://us-south.ml.cloud.ibm.com",
-                    api_key="test-api-key",
-                    project_id="test-project-123",
-                    token="test-bearer-token",
-                )
-            except Exception:
-                pass  # We just want to capture the request
-
-        # Validate URL contains WatsonX audio transcription endpoint
-        assert "/ml/v1/audio/transcriptions" in captured_request["url"]
-        assert "version=" in captured_request["url"]
-        # project_id should NOT be in URL (it should be in form data instead)
-        assert "project_id=test-project-123" not in captured_request["url"]
-
-        # Validate headers contain WatsonX auth
-        assert "Authorization" in captured_request["headers"]
-        assert (
-            "Bearer test-bearer-token" in captured_request["headers"]["Authorization"]
-        )
-
-        # Validate Content-Type is NOT set (httpx sets multipart/form-data automatically)
-        assert "Content-Type" not in captured_request["headers"]
-
-        # Validate project_id is in form data, not URL
-        assert captured_request["data"].get("project_id") == "test-project-123"
-
-        # Validate file is in files dict
-        assert "file" in captured_request["files"]
-
-    @pytest.mark.asyncio
-    async def test_watsonx_transcription_request_body(self):
-        """
-        Test that litellm.transcription sends correct request body for WatsonX.
-
-        Validates that:
-        - Request uses multipart/form-data (data + files)
-        - Model name has watsonx/ prefix removed
-        - project_id is in form data, not URL
-        - Audio file is in files dict
-        - OpenAI params are included in form data
-        """
-        captured_request = {}
-
-        async def mock_post(*args, **kwargs):
-            captured_request["data"] = kwargs.get("data", {})
-            captured_request["files"] = kwargs.get("files", {})
-
-            mock_response = MagicMock()
-            mock_response.json.return_value = {
-                "text": "test transcription",
-                "duration": 1.0,
-            }
-            mock_response.status_code = 200
-            return mock_response
-
-        with patch(
-            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
-            new=mock_post,
-        ):
-            try:
-                await litellm.atranscription(
-                    model="watsonx/whisper-large-v3-turbo",
-                    file=b"fake_audio_data",
-                    api_base="https://us-south.ml.cloud.ibm.com",
-                    api_key="test-api-key",
-                    project_id="test-project-123",
-                    token="test-bearer-token",
-                    language="en",
-                    temperature=0.5,
-                )
-            except Exception:
-                pass  # We just want to capture the request
-
-        # Validate form data contains expected fields
-        data = captured_request.get("data", {})
-
-        print("JSON DUMPS captured_request:")
-        print(json.dumps(captured_request, indent=4, default=str))
-
-        # Model name should NOT have watsonx/ prefix
-        assert data.get("model") == "whisper-large-v3-turbo"
-
-        # project_id should be in form data
-        assert data.get("project_id") == "test-project-123"
-
-        # OpenAI params should be in form data
-        assert data.get("language") == "en"
-        assert data.get("temperature") == 0.5
-        # response_format should NOT be set by default - only send what user specifies
-        assert "response_format" not in data
-
-        # Validate file is in files dict (multipart/form-data)
-        files = captured_request.get("files", {})
-        assert "file" in files
-        assert isinstance(
-            files["file"], tuple
-        )  # Should be (filename, content, content_type)
-
-    @pytest.mark.asyncio
-    async def test_watsonx_transcription_only_user_params_sent_with_project_id(self):
-        """
-        Test that only user-specified params are sent in request body to WatsonX.
-
-        LiteLLM should NOT add extra params like response_format if user didn't specify them.
-        """
-        captured_request = {}
-
-        async def mock_post(*args, **kwargs):
-            captured_request["data"] = kwargs.get("data", {})
-            captured_request["files"] = kwargs.get("files", {})
-
-            mock_response = MagicMock()
-            mock_response.json.return_value = {
-                "text": "test transcription",
-                "duration": 1.0,
-            }
-            mock_response.status_code = 200
-            return mock_response
-
-        with patch(
-            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
-            new=mock_post,
-        ):
-            try:
-                # Minimal request - only required params
-                await litellm.atranscription(
-                    model="watsonx/whisper-large-v3-turbo",
-                    file=b"fake_audio_data",
-                    api_base="https://us-south.ml.cloud.ibm.com",
-                    api_key="test-api-key",
-                    project_id="test-project-123",
-                    token="test-bearer-token",
-                )
-            except Exception:
-                pass  # We just want to capture the request
-
-        data = captured_request.get("data", {})
-
-        # These are the ONLY keys that should be in data
-        expected_keys = {"model", "project_id"}
-        actual_keys = set(data.keys())
-
-        assert actual_keys == expected_keys, (
-            f"Request body should only contain {expected_keys}, "
-            f"but got {actual_keys}. "
-            f"Extra keys: {actual_keys - expected_keys}"
-        )
-
-        # Specifically verify response_format is NOT added
-        assert (
-            "response_format" not in data
-        ), "response_format should NOT be added by default"
-
-        # Verify file is sent separately
-        files = captured_request.get("files", {})
-        assert "file" in files
-
-    @pytest.mark.asyncio
-    async def test_watsonx_transcription_only_user_params_sent_with_space_id(self):
-        """
-        Test that only user-specified params are sent in request body to WatsonX.
-
-        LiteLLM should NOT add extra params like response_format if user didn't specify them.
-        """
-        captured_request = {}
-
-        async def mock_post(*args, **kwargs):
-            captured_request["data"] = kwargs.get("data", {})
-            captured_request["files"] = kwargs.get("files", {})
-
-            mock_response = MagicMock()
-            mock_response.json.return_value = {
-                "text": "test transcription",
-                "duration": 1.0,
-            }
-            mock_response.status_code = 200
-            return mock_response
-
-        with patch(
-            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
-            new=mock_post,
-        ):
-            try:
-                # Minimal request - only required params
-                await litellm.atranscription(
-                    model="watsonx/whisper-large-v3-turbo",
-                    file=b"fake_audio_data",
-                    api_base="https://us-south.ml.cloud.ibm.com",
-                    api_key="test-api-key",
-                    space_id="test-space_id-123",
-                    token="test-bearer-token",
-                )
-            except Exception:
-                pass  # We just want to capture the request
-
-        data = captured_request.get("data", {})
-
-        # These are the ONLY keys that should be in data
-        expected_keys = {"model", "space_id"}
-        actual_keys = set(data.keys())
-
-        assert actual_keys == expected_keys, (
-            f"Request body should only contain {expected_keys}, "
-            f"but got {actual_keys}. "
-            f"Extra keys: {actual_keys - expected_keys}"
-        )
-
-        # Specifically verify response_format is NOT added
-        assert (
-            "response_format" not in data
-        ), "response_format should NOT be added by default"
-
-        # Verify file is sent separately
-        files = captured_request.get("files", {})
-        assert "file" in files
-
-    def test_transform_audio_transcription_response_removes_model_field(self):
-        """
-        Test that transform_audio_transcription_response removes the 'model' field
-        from WatsonX response before creating TranscriptionResponse.
-
-        This test ensures that when WatsonX returns a response with a 'model' field,
-        it is removed before creating the TranscriptionResponse object, since
-        TranscriptionResponse doesn't accept a 'model' parameter.
-        """
-        handler = IBMWatsonXAudioTranscriptionConfig()
-
-        # Mock response with 'model' field (as WatsonX may return)
-        mock_response = MagicMock()
-        mock_response.json.return_value = {
-            "text": "Hello, this is a test transcription.",
-            "model": "whisper-large-v3-turbo",  # This field should be removed
-            "duration": 5.5,
-        }
-        mock_response.text = '{"text": "Hello, this is a test transcription.", "model": "whisper-large-v3-turbo", "duration": 5.5}'
-
-        # This should not raise a TypeError - model field should be removed
-        result = handler.transform_audio_transcription_response(mock_response)
-
-        # Verify the result is a TranscriptionResponse
-        assert isinstance(result, TranscriptionResponse)
-
-        # Verify the text is correct
-        assert result.text == "Hello, this is a test transcription."
-
-        # Verify duration is set via dictionary assignment
-        assert result["duration"] == 5.5
-
-        # Verify the model field is NOT in the serialized result
-        # Check via model_dump() or dict() to ensure it's not in the output
-        try:
-            result_dict = result.model_dump()
-        except AttributeError:
-            # Fallback for pydantic v1
-            result_dict = result.dict()
-
-        # The 'model' field should not be in the result
-        assert "model" not in result_dict, "Model field should be removed from response"
-
-    def test_transform_audio_transcription_response_without_model_field(self):
-        """
-        Test that transform_audio_transcription_response works correctly
-        when WatsonX response doesn't include a 'model' field.
-        """
-        handler = IBMWatsonXAudioTranscriptionConfig()
-
-        # Mock response without 'model' field
-        mock_response = MagicMock()
-        mock_response.json.return_value = {
-            "text": "Hello, this is a test transcription.",
-            "duration": 5.5,
-        }
-        mock_response.text = (
-            '{"text": "Hello, this is a test transcription.", "duration": 5.5}'
-        )
-
-        result = handler.transform_audio_transcription_response(mock_response)
-
-        # Verify the result is a TranscriptionResponse
-        assert isinstance(result, TranscriptionResponse)
-
-        # Verify the text is correct
-        assert result.text == "Hello, this is a test transcription."
-
-        # Verify duration is set via dictionary assignment
-        assert result["duration"] == 5.5

tests/test_litellm/llms/watsonx/test_watsonx.py

@@ -1,577 +0,0 @@
-import json
-
-from typing import Optional
-from unittest.mock import Mock, patch
-
-import pytest
-
-import litellm
-from litellm import completion
-from litellm.llms.custom_httpx.http_handler import HTTPHandler
-
-
-@pytest.fixture
-def watsonx_chat_completion_call():
-    def _call(
-        model="watsonx/my-test-model",
-        messages=None,
-        api_key="test_api_key",
-        space_id: Optional[str] = None,
-        headers=None,
-        client=None,
-        patch_token_call=True,
-    ):
-        if messages is None:
-            messages = [{"role": "user", "content": "Hello, how are you?"}]
-        if client is None:
-            client = HTTPHandler()
-
-        if patch_token_call:
-            mock_response = Mock()
-            mock_response.json.return_value = {
-                "access_token": "mock_access_token",
-                "expires_in": 3600,
-            }
-            mock_response.raise_for_status = Mock()  # No-op to simulate no exception
-
-            with (
-                patch.object(client, "post") as mock_post,
-                patch.object(
-                    litellm.module_level_client, "post", return_value=mock_response
-                ) as mock_get,
-            ):
-                try:
-                    completion(
-                        model=model,
-                        messages=messages,
-                        api_key=api_key,
-                        headers=headers or {},
-                        client=client,
-                        space_id=space_id,
-                    )
-                except Exception as e:
-                    print(e)
-
-                return mock_post, mock_get
-        else:
-            with patch.object(client, "post") as mock_post:
-                try:
-                    completion(
-                        model=model,
-                        messages=messages,
-                        api_key=api_key,
-                        headers=headers or {},
-                        client=client,
-                        space_id=space_id,
-                    )
-                except Exception as e:
-                    print(e)
-                return mock_post, None
-
-    return _call
-
-
-def test_watsonx_deployment_model_id_not_in_payload(
-    monkeypatch, watsonx_chat_completion_call
-):
-    """Test that deployment models do not include 'model_id' in the request payload"""
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-    model = "watsonx/deployment/test-deployment-id"
-    messages = [{"role": "user", "content": "Test message"}]
-
-    mock_post, _ = watsonx_chat_completion_call(model=model, messages=messages)
-
-    assert mock_post.call_count == 1
-    json_data = json.loads(mock_post.call_args.kwargs["data"])
-    # Ensure model_id is not in the payload for deployment models
-    assert "model_id" not in json_data or json_data["model_id"] is None
-    # Ensure project_id is also not in the payload for deployment models
-    assert "project_id" not in json_data or json_data["project_id"] is None
-
-
-def test_watsonx_regular_model_includes_model_id(
-    monkeypatch, watsonx_chat_completion_call
-):
-    """Test that regular models include 'model_id' in the request payload"""
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-    model = "watsonx/regular-model"
-    messages = [{"role": "user", "content": "Test message"}]
-
-    mock_post, _ = watsonx_chat_completion_call(model=model, messages=messages)
-
-    assert mock_post.call_count == 1
-    json_data = json.loads(mock_post.call_args.kwargs["data"])
-    # Ensure model_id is included in the payload for regular models
-    assert "model_id" in json_data
-    assert json_data["model_id"] == "regular-model"  # Provider prefix is stripped
-    # Ensure project_id is also included for regular models
-    assert "project_id" in json_data
-
-
-@pytest.fixture
-def watsonx_completion_call():
-    def _call(
-        model="watsonx_text/my-test-model",
-        prompt="Hello, how are you?",
-        api_key="test_api_key",
-        space_id: Optional[str] = None,
-        headers=None,
-        client=None,
-        patch_token_call=True,
-    ):
-        if client is None:
-            client = HTTPHandler()
-
-        if patch_token_call:
-            mock_response = Mock()
-            mock_response.json.return_value = {
-                "access_token": "mock_access_token",
-                "expires_in": 3600,
-            }
-            mock_response.raise_for_status = Mock()
-
-            with (
-                patch.object(client, "post") as mock_post,
-                patch.object(
-                    litellm.module_level_client, "post", return_value=mock_response
-                ) as mock_get,
-            ):
-                try:
-                    litellm.text_completion(
-                        model=model,
-                        prompt=prompt,
-                        api_key=api_key,
-                        headers=headers or {},
-                        client=client,
-                        space_id=space_id,
-                    )
-                except Exception as e:
-                    print(e)
-
-                return mock_post, mock_get
-        else:
-            with patch.object(client, "post") as mock_post:
-                try:
-                    litellm.text_completion(
-                        model=model,
-                        prompt=prompt,
-                        api_key=api_key,
-                        headers=headers or {},
-                        client=client,
-                        space_id=space_id,
-                    )
-                except Exception as e:
-                    print(e)
-                return mock_post, None
-
-    return _call
-
-
-def test_watsonx_completion_deployment_model_id_not_in_payload(
-    monkeypatch, watsonx_completion_call
-):
-    """Test that deployment models do not include 'model_id' in completion request payload"""
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-    model = "watsonx_text/deployment/test-deployment-id"
-    prompt = "Test prompt"
-
-    mock_post, _ = watsonx_completion_call(model=model, prompt=prompt)
-
-    assert mock_post.call_count == 1
-    json_data = json.loads(mock_post.call_args.kwargs["data"])
-    # Ensure model_id is not in the payload for deployment models
-    assert "model_id" not in json_data
-    # Ensure project_id is also not in the payload for deployment models
-    assert "project_id" not in json_data
-
-
-def test_watsonx_completion_regular_model_includes_model_id(
-    monkeypatch, watsonx_completion_call
-):
-    """Test that regular models include 'model_id' in completion request payload"""
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-    model = "watsonx_text/regular-model"
-    prompt = "Test prompt"
-
-    mock_post, _ = watsonx_completion_call(model=model, prompt=prompt)
-
-    assert mock_post.call_count == 1
-    json_data = json.loads(mock_post.call_args.kwargs["data"])
-    # Ensure model_id is included in the payload for regular models
-    assert "model_id" in json_data
-    assert json_data["model_id"] == "regular-model"  # Provider prefix is stripped
-    # Ensure project_id is also included for regular models
-    assert "project_id" in json_data
-
-
-def test_watsonx_gpt_oss_prompt_transformation(monkeypatch):
-    """
-    Test that gpt-oss-120b model transforms messages to proper format instead of simple concatenation.
-
-    This test calls litellm.completion (sync) and verifies what gets sent in the final POST request body.
-    Input messages should be transformed using the HuggingFace chat template from openai/gpt-oss-120b,
-    not just concatenated as "You are chatgpt Hi there".
-    """
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-
-    # Test with gpt-oss model using watsonx_text provider (text generation endpoint)
-    model = "watsonx_text/openai/gpt-oss-120b"
-
-    # Input messages
-    messages = [
-        {"role": "system", "content": "You are chatgpt"},
-        {"role": "user", "content": "Hi there"},
-    ]
-
-    client = HTTPHandler()
-
-    # Mock HuggingFace template fetch to make test deterministic and avoid network flakiness.
-    # The test verifies that prompt transformation occurs (not simple concatenation), not the exact
-    # HuggingFace template format. Using a mock template that produces the correct format is sufficient.
-    #
-    # Mock template that produces gpt-oss-120b-like format.
-    # Note: This is a simplified version of the actual template. The real template is more complex
-    # (adds metadata, handles tools, thinking messages, etc.), but this captures the key aspects:
-    # - Converts system role to developer (matching real template behavior)
-    # - Uses the same tag structure (<|start|>, <|message|>, <|end|>)
-    # - Preserves message content
-    mock_tokenizer_config = {
-        "status": "success",
-        "tokenizer": {
-            "chat_template": "{% for message in messages %}{% if message['role'] == 'system' %}<|start|>developer<|message|>{% else %}<|start|>{{ message['role'] }}<|message|>{% endif %}{{ message['content'] }}<|end|>{% endfor %}",
-            "bos_token": None,
-            "eos_token": None,
-        },
-    }
-
-    # Isolate known_tokenizer_config so parallel tests don't interfere.
-    # monkeypatch.setitem restores the original value on teardown.
-    hf_model = "openai/gpt-oss-120b"
-    monkeypatch.setitem(litellm.known_tokenizer_config, hf_model, mock_tokenizer_config)
-
-    # Mock IAM token generation to avoid real HTTP calls.
-    mock_token_response = Mock()
-    mock_token_response.json.return_value = {
-        "access_token": "mock_access_token",
-        "expires_in": 3600,
-    }
-    mock_token_response.raise_for_status = Mock()
-
-    with (
-        patch.object(client, "post") as mock_post,
-        patch.object(
-            litellm.module_level_client, "post", return_value=mock_token_response
-        ),
-    ):
-        try:
-            completion(
-                model=model,
-                messages=messages,
-                api_key="test_api_key",
-                client=client,
-            )
-        except Exception as e:
-            print(f"Caught expected exception: {e}")
-
-    # Verify the POST was called
-    assert (
-        mock_post.call_count == 1
-    ), f"POST should have been called exactly once, got {mock_post.call_count}"
-
-    # Get the request body
-    call_args = mock_post.call_args
-    assert "data" in call_args.kwargs, "call_args.kwargs should contain 'data'"
-    json_data = json.loads(call_args.kwargs["data"])
-
-    # Verify the transformed input is in the request
-    assert "input" in json_data, "Request should have 'input' field"
-    transformed_prompt = json_data["input"]
-
-    # Verify it's NOT simple concatenation
-    simple_concat = "You are chatgpt Hi there"
-    assert transformed_prompt != simple_concat, (
-        f"Prompt should not be simple concatenation.\n"
-        f"Expected: Chat template with <|start|> tags\n"
-        f"Got: {transformed_prompt}"
-    )
-
-    # Verify it contains proper chat template formatting
-    assert "<|start|>" in transformed_prompt, "Prompt should contain <|start|> tag"
-    assert "<|message|>" in transformed_prompt, "Prompt should contain <|message|> tag"
-    assert "<|end|>" in transformed_prompt, "Prompt should contain <|end|> tag"
-    assert (
-        "You are chatgpt" in transformed_prompt
-    ), "Prompt should contain system message content"
-    assert (
-        "Hi there" in transformed_prompt
-    ), "Prompt should contain user message content"
-
-
-@pytest.mark.asyncio
-@pytest.mark.xdist_group("watsonx_heavy")
-async def test_watsonx_gpt_oss_uses_async_http_handler():
-    """
-    Test that verifies async HTTP client is used when fetching HuggingFace templates.
-    """
-    from unittest.mock import AsyncMock, MagicMock, patch
-
-    from litellm.litellm_core_utils.prompt_templates.huggingface_template_handler import (
-        _aget_chat_template_file,
-    )
-
-    # Mock the async HTTP client
-    mock_async_client = MagicMock()
-    mock_get = AsyncMock()
-    mock_async_client.get = mock_get
-
-    # Create mock response for chat template file
-    mock_response = MagicMock()
-    mock_response.status_code = 200
-    mock_response.content = b"test template content"
-    mock_get.return_value = mock_response
-
-    # Test the async function directly
-    with patch(
-        "litellm.litellm_core_utils.prompt_templates.huggingface_template_handler.get_async_httpx_client",
-        return_value=mock_async_client,
-    ):
-        result = await _aget_chat_template_file(hf_model_name="test/model")
-
-        # Verify async HTTP client was called
-        assert mock_get.called, "Async HTTP client's get method should be called"
-        assert mock_get.await_count > 0, "Async HTTP client's get should be awaited"
-
-        # Verify it was called with HuggingFace URL
-        call_args = mock_get.call_args
-        assert call_args is not None, "get should have been called with arguments"
-        called_url = call_args.kwargs.get("url", "")
-        assert (
-            "huggingface.co/test/model" in called_url
-        ), f"Should call HuggingFace API for test/model, got: {called_url}"
-        assert result["status"] == "success", "Should return success status"
-
-
-@pytest.mark.parametrize("tokenizer_config_cached", [False, True], ids=["tokenizer_config", "cached_config_jinja"])
-async def test_watsonx_text_gpt_oss_async_completion_fetches_hf_template_off_the_event_loop(
-    monkeypatch, tokenizer_config_cached
-):
-    import httpx
-
-    from litellm._uuid import uuid
-    from litellm.litellm_core_utils.prompt_templates import huggingface_template_handler
-    from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
-
-    hf_model = f"openai/gpt-oss-{uuid.uuid4()}"
-    chat_template = "{% for m in messages %}<|{{ m['role'] }}|>{{ m['content'] }}{% endfor %}"
-    if tokenizer_config_cached:
-        cached_config = {"status": "success", "tokenizer": {"bos_token": None, "eos_token": None}}
-        monkeypatch.setattr(litellm, "known_tokenizer_config", {hf_model: cached_config})
-        expected_fetch = f"https://huggingface.co/{hf_model}/raw/main/chat_template.jinja"
-    else:
-        monkeypatch.setattr(litellm, "known_tokenizer_config", {})
-        expected_fetch = f"https://huggingface.co/{hf_model}/raw/main/tokenizer_config.json"
-    hf_fetched = []
-    captured = {}
-
-    def forbid_sync_client():
-        raise AssertionError("sync HuggingFace fetch ran on the request path")
-
-    async def serve_hf_file(url, **kwargs):
-        hf_fetched.append(url)
-        if url.endswith(".jinja"):
-            return httpx.Response(200, content=chat_template.encode())
-        return httpx.Response(200, json={"chat_template": chat_template, "bos_token": None, "eos_token": None})
-
-    monkeypatch.setattr(huggingface_template_handler, "_get_httpx_client", forbid_sync_client)
-    monkeypatch.setattr(huggingface_template_handler, "get_async_httpx_client", lambda **kwargs: Mock(get=serve_hf_file))
-
-    def handle(request):
-        captured["body"] = json.loads(request.content)
-        return httpx.Response(
-            200,
-            json={
-                "model_id": hf_model,
-                "results": [
-                    {
-                        "generated_text": "Hi",
-                        "generated_token_count": 1,
-                        "input_token_count": 1,
-                        "stop_reason": "eos_token",
-                    }
-                ],
-            },
-        )
-
-    client = AsyncHTTPHandler()
-    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
-
-    response = await litellm.acompletion(
-        model=f"watsonx_text/{hf_model}",
-        messages=[{"role": "user", "content": "Hi there"}],
-        api_base="https://test-api.watsonx.ai",
-        project_id="test-project-id",
-        token="test-token",
-        client=client,
-    )
-
-    assert response.choices[0].message.content == "Hi"
-    assert hf_fetched == [expected_fetch]
-    assert captured["body"]["input"] == "<|user|>Hi there"
-
-
-def test_watsonx_chat_completion_with_reasoning_effort(monkeypatch):
-    """
-    Test that 'reasoning_effort' is correctly passed through to the WatsonX API payload.
-    """
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-
-    model = "watsonx/openai/gpt-oss-120b"
-    messages = [{"role": "user", "content": "Test message"}]
-
-    client = HTTPHandler()
-
-    # Mock the token generation call
-    mock_token_response = Mock()
-    mock_token_response.json.return_value = {
-        "access_token": "mock_access_token",
-        "expires_in": 3600,
-    }
-    mock_token_response.raise_for_status = Mock()
-
-    # Call litellm.completion with the new parameter
-    with (
-        patch.object(client, "post") as mock_post,
-        patch.object(
-            litellm.module_level_client, "post", return_value=mock_token_response
-        ),
-    ):
-        try:
-            completion(
-                model=model,
-                messages=messages,
-                api_key="test_api_key",
-                client=client,
-                reasoning_effort="low",
-            )
-        except Exception as e:
-            print(f"Caught expected exception: {e}")
-
-    # Verify the parameter is in the final request payload
-    assert (
-        mock_post.call_count == 1
-    ), "The completion endpoint should have been called once."
-
-    # Get the JSON data sent in the POST request
-    request_kwargs = mock_post.call_args.kwargs
-    json_data = json.loads(request_kwargs["data"])
-
-    print("\nRequest payload sent to WatsonX API:")
-    print(json.dumps(json_data, indent=2))
-
-    # Check for the parameter at the top level of the payload
-    assert (
-        "reasoning_effort" in json_data
-    ), "'reasoning_effort' should be at the top level of the payload."
-    assert (
-        json_data["reasoning_effort"] == "low"
-    ), "The value of 'reasoning_effort' should be 'low'."
-
-
-def test_watsonx_zen_api_key_from_client(monkeypatch, watsonx_chat_completion_call):
-    """
-    Test that zen_api_key can be passed from client code and is used in Authorization header.
-    """
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-
-    model = "watsonx/ibm/granite-3-3-8b-instruct"
-    messages = [{"role": "user", "content": "What is your favorite color?"}]
-
-    client = HTTPHandler()
-
-    zen_api_key = "U1ZDLWQo="
-
-    # No need to patch token call since zen_api_key should skip token generation
-    with patch.object(client, "post") as mock_post:
-        try:
-            completion(
-                model=model,
-                messages=messages,
-                api_key="test_api_key",
-                client=client,
-                zen_api_key=zen_api_key,
-            )
-        except Exception as e:
-            print(f"Caught expected exception: {e}")
-
-    # Verify the request was made
-    assert (
-        mock_post.call_count == 1
-    ), "The completion endpoint should have been called once."
-
-    # Get the headers sent in the POST request
-    request_kwargs = mock_post.call_args.kwargs
-    headers = request_kwargs["headers"]
-
-    print("\nHeaders sent to WatsonX API:")
-    print(json.dumps(dict(headers), indent=2))
-
-    # Verify Authorization header uses ZenApiKey format
-    assert "Authorization" in headers, "Authorization header should be present."
-    assert headers["Authorization"] == f"ZenApiKey {zen_api_key}", (
-        f"Authorization header should use ZenApiKey format. "
-        f"Expected: 'ZenApiKey {zen_api_key}', Got: '{headers['Authorization']}'"
-    )
-
-
-def test_watsonx_zen_api_key_from_env(monkeypatch, watsonx_chat_completion_call):
-    """
-    Test that zen_api_key from environment variable is used in Authorization header.
-    """
-    monkeypatch.setenv("WATSONX_PROJECT_ID", "test-project-id")
-    monkeypatch.setenv("WATSONX_API_BASE", "https://test-api.watsonx.ai")
-
-    zen_api_key = "U1ZDLWxpdG--==="
-    monkeypatch.setenv("WATSONX_ZENAPIKEY", zen_api_key)
-
-    model = "watsonx/ibm/granite-3-3-8b-instruct"
-    messages = [{"role": "user", "content": "What is your favorite color?"}]
-
-    client = HTTPHandler()
-
-    # No need to patch token call since zen_api_key should skip token generation
-    with patch.object(client, "post") as mock_post:
-        try:
-            completion(
-                model=model,
-                messages=messages,
-                api_key="test_api_key",
-                client=client,
-            )
-        except Exception as e:
-            print(f"Caught expected exception: {e}")
-
-    # Verify the request was made
-    assert (
-        mock_post.call_count == 1
-    ), "The completion endpoint should have been called once."
-
-    # Get the headers sent in the POST request
-    request_kwargs = mock_post.call_args.kwargs
-    headers = request_kwargs["headers"]
-
-    print("\nHeaders sent to WatsonX API:")
-    print(json.dumps(dict(headers), indent=2))
-
-    # Verify Authorization header uses ZenApiKey format
-    assert "Authorization" in headers, "Authorization header should be present."
-    assert headers["Authorization"] == f"ZenApiKey {zen_api_key}", (
-        f"Authorization header should use ZenApiKey format. "
-        f"Expected: 'ZenApiKey {zen_api_key}', Got: '{headers['Authorization']}'"
-    )

tests/test_litellm/llms/xai/xai_responses/__init__.py

@@ -1 +0,0 @@
-# XAI Responses API tests

tests/test_litellm/llms/xai/xai_responses/test_transformation.py

@@ -1,105 +0,0 @@
-"""
-Tests for XAI Responses API transformation
-
-Tests the XAIResponsesAPIConfig class that handles XAI-specific
-transformations for the Responses API.
-
-Source: litellm/llms/xai/responses/transformation.py
-"""
-
-
-
-import pytest
-from litellm.types.utils import LlmProviders
-from litellm.utils import ProviderConfigManager
-from litellm.llms.xai.responses.transformation import XAIResponsesAPIConfig
-from litellm.types.llms.openai import ResponsesAPIOptionalRequestParams
-
-
-class TestXAIResponsesAPITransformation:
-    """Test XAI Responses API configuration and transformations"""
-
-    def test_xai_provider_config_registration(self):
-        """Test that XAI provider returns XAIResponsesAPIConfig"""
-        config = ProviderConfigManager.get_provider_responses_api_config(
-            model="xai/grok-4-fast",
-            provider=LlmProviders.XAI,
-        )
-
-        assert config is not None, "Config should not be None for XAI provider"
-        assert isinstance(
-            config, XAIResponsesAPIConfig
-        ), f"Expected XAIResponsesAPIConfig, got {type(config)}"
-        assert (
-            config.custom_llm_provider == LlmProviders.XAI
-        ), "custom_llm_provider should be XAI"
-
-    def test_code_interpreter_container_field_removed(self):
-        """Test that container field is removed from code_interpreter tools"""
-        config = XAIResponsesAPIConfig()
-
-        params = ResponsesAPIOptionalRequestParams(
-            tools=[{"type": "code_interpreter", "container": {"type": "auto"}}]
-        )
-
-        result = config.map_openai_params(
-            response_api_optional_params=params, model="grok-4-fast", drop_params=False
-        )
-
-        assert "tools" in result
-        assert len(result["tools"]) == 1
-        assert result["tools"][0]["type"] == "code_interpreter"
-        assert (
-            "container" not in result["tools"][0]
-        ), "Container field should be removed"
-
-    def test_instructions_parameter_forwarded(self):
-        """xAI supports 'instructions' on /v1/responses, so it must survive param mapping"""
-        config = XAIResponsesAPIConfig()
-
-        params = ResponsesAPIOptionalRequestParams(
-            instructions="You are a helpful assistant.", temperature=0.7
-        )
-
-        result = config.map_openai_params(
-            response_api_optional_params=params, model="grok-4-fast", drop_params=False
-        )
-
-        assert result.get("instructions") == "You are a helpful assistant."
-        assert result.get("temperature") == 0.7, "Other params should be preserved"
-
-    def test_supported_params_includes_instructions(self):
-        """A system message bridged to 'instructions' must not be rejected for xAI"""
-        config = XAIResponsesAPIConfig()
-        supported = config.get_supported_openai_params("grok-4-fast")
-
-        assert "instructions" in supported, "instructions should be supported"
-        assert "tools" in supported, "tools should be supported"
-        assert "temperature" in supported, "temperature should be supported"
-        assert "model" in supported, "model should be supported"
-
-    def test_xai_responses_endpoint_url(self):
-        """Test that get_complete_url returns correct XAI endpoint"""
-        config = XAIResponsesAPIConfig()
-
-        # Test with default XAI API base
-        url = config.get_complete_url(api_base=None, litellm_params={})
-        assert (
-            url == "https://api.x.ai/v1/responses"
-        ), f"Expected XAI responses endpoint, got {url}"
-
-        # Test with custom api_base
-        custom_url = config.get_complete_url(
-            api_base="https://custom.x.ai/v1", litellm_params={}
-        )
-        assert (
-            custom_url == "https://custom.x.ai/v1/responses"
-        ), f"Expected custom endpoint, got {custom_url}"
-
-        # Test with trailing slash
-        url_with_slash = config.get_complete_url(
-            api_base="https://api.x.ai/v1/", litellm_params={}
-        )
-        assert (
-            url_with_slash == "https://api.x.ai/v1/responses"
-        ), "Should handle trailing slash"

tests/unit/conftest.py

@@ -1,4 +1,5 @@
 import os
+from collections.abc import Iterator
 from typing import Final
 
 import litellm
@@ -7,7 +8,19 @@ from pytest_socket import enable_socket, socket_allow_hosts
 
 os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
 
+import litellm  # noqa: E402  # litellm reads LITELLM_LOCAL_MODEL_COST_MAP at import
+import litellm.router as litellm_router_module  # noqa: E402  # same import-time dependency
+import litellm.utils as litellm_utils_module  # noqa: E402  # same import-time dependency
+
 LOOPBACK_HOSTS: Final = ["127.0.0.1", "::1"]
+AMBIENT_AZURE_CREDENTIAL_ENV_VARS: Final = (
+    "AZURE_AD_TOKEN",
+    "AZURE_TENANT_ID",
+    "AZURE_CLIENT_ID",
+    "AZURE_CLIENT_SECRET",
+    "AZURE_USERNAME",
+    "AZURE_PASSWORD",
+)
 
 
 @pytest.fixture
@@ -35,5 +48,38 @@ def pytest_runtest_setup() -> None:
     _allow_loopback_only()
 
 
+@pytest.fixture(autouse=True)
+def isolate_router_model_cost_state() -> Iterator[None]:
+    original_live_routers: Final = frozenset(litellm_router_module._live_routers)
+    original_runtime_registered_model_cost: Final = {
+        model_key: dict(model_value)
+        for model_key, model_value in litellm_utils_module._runtime_registered_model_cost.items()
+    }
+    yield
+    for router in tuple(litellm_router_module._live_routers):
+        litellm_router_module._live_routers.discard(router)
+    for router in original_live_routers:
+        litellm_router_module._live_routers.add(router)
+    litellm_utils_module._runtime_registered_model_cost.clear()
+    litellm_utils_module._runtime_registered_model_cost.update(original_runtime_registered_model_cost)
+    litellm_utils_module._invalidate_model_cost_lowercase_map()
+    litellm.get_model_info.cache_clear()
+
+
+@pytest.fixture
+def local_model_cost_map(monkeypatch: pytest.MonkeyPatch) -> Iterator[None]:
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    litellm.get_model_info.cache_clear()
+    yield
+    litellm.get_model_info.cache_clear()
+
+
+@pytest.fixture
+def no_ambient_azure_credentials(monkeypatch: pytest.MonkeyPatch) -> None:
+    for name in AMBIENT_AZURE_CREDENTIAL_ENV_VARS:
+        monkeypatch.delenv(name, raising=False)
+
+
 def pytest_sessionfinish() -> None:
     enable_socket()

tests/unit/llms/azure/image_generation/test_azure_image_generation_init.py

@@ -133,88 +133,6 @@ def test_azure_image_generation_flattens_extra_body():
     assert data["size"] == "1024x1024"
 
 
-def test_azure_image_generation_creates_token_provider_from_credentials():
-    """
-    Test that azure_ad_token_provider is created from tenant_id, client_id, client_secret.
-
-    This test verifies the fix in images/main.py where we now create the
-    azure_ad_token_provider from credentials in litellm_params if it's not already provided.
-    """
-    # Simulate the fix in images/main.py
-    litellm_params_dict = {
-        "tenant_id": "test-tenant-id",
-        "client_id": "test-client-id",
-        "client_secret": "test-client-secret",
-        "azure_scope": None,
-    }
-
-    azure_ad_token_provider = None
-
-    # This is the logic we added in images/main.py
-    if azure_ad_token_provider is None:
-        tenant_id = litellm_params_dict.get("tenant_id")
-        client_id = litellm_params_dict.get("client_id")
-        client_secret = litellm_params_dict.get("client_secret")
-        azure_scope = (
-            litellm_params_dict.get("azure_scope")
-            or "https://cognitiveservices.azure.com/.default"
-        )
-
-        # Verify the credentials are extracted correctly
-        assert tenant_id == "test-tenant-id"
-        assert client_id == "test-client-id"
-        assert client_secret == "test-client-secret"
-        assert azure_scope == "https://cognitiveservices.azure.com/.default"
-
-        # Verify the condition to create token provider is met
-        assert (
-            tenant_id and client_id and client_secret
-        ), "Credentials should be present to create token provider"
-
-
-def test_azure_image_generation_headers_without_api_key():
-    """
-    Test that when api_key is None, the api-key header is not added to headers.
-
-    This prevents the httpx TypeError: "Header value must be str or bytes, not <class 'NoneType'>"
-    that was occurring when api_key was None and being set in headers.
-
-    This is a unit test for the fix in images/main.py where we now check:
-    if api_key is not None:
-        default_headers["api-key"] = api_key
-    """
-    from litellm.images.main import image_generation
-
-    # Test the header building logic directly
-    api_key = None
-
-    default_headers = {
-        "Content-Type": "application/json",
-    }
-
-    # This is the fix: only add api-key if it's not None
-    if api_key is not None:
-        default_headers["api-key"] = api_key
-
-    # Verify api-key is not in headers when api_key is None
-    assert "api-key" not in default_headers
-
-    # Verify Content-Type is still there
-    assert default_headers["Content-Type"] == "application/json"
-
-    # Test with a valid api_key
-    api_key = "valid-key-123"
-    default_headers_with_key = {
-        "Content-Type": "application/json",
-    }
-    if api_key is not None:
-        default_headers_with_key["api-key"] = api_key
-
-    # Verify api-key is added when api_key is valid
-    assert "api-key" in default_headers_with_key
-    assert default_headers_with_key["api-key"] == "valid-key-123"
-
-
 def test_azure_image_generation_drop_params_response_format():
     """
     Test that unsupported params like response_format are dropped when drop_params=True.

tests/unit/llms/azure/realtime/test_azure_realtime_handler.py

@@ -426,41 +426,6 @@ async def test_async_realtime_beta_without_api_version_raises():
         )
 
 
-@pytest.mark.asyncio
-async def test_realtime_protocol_env_var_fallback():
-    """
-    Test that LITELLM_AZURE_REALTIME_PROTOCOL env var is used as fallback.
-    Fixes #22127: no way to set realtime_protocol from config.
-    """
-    from litellm.realtime_api.main import _arealtime
-    from litellm.types.router import GenericLiteLLMParams
-
-    with patch.dict(os.environ, {"LITELLM_AZURE_REALTIME_PROTOCOL": "v1"}):
-        # Create a GenericLiteLLMParams without realtime_protocol
-        litellm_params = GenericLiteLLMParams()
-        # The env var should be picked up as fallback
-        realtime_protocol = (
-            {}.get("realtime_protocol")
-            or litellm_params.get("realtime_protocol")
-            or os.environ.get("LITELLM_AZURE_REALTIME_PROTOCOL")
-            or "beta"
-        )
-        assert realtime_protocol == "v1"
-
-
-@pytest.mark.asyncio
-async def test_realtime_protocol_from_litellm_params():
-    """
-    Test that realtime_protocol is read from litellm_params (config.yaml extra field).
-    Fixes #22127: realtime_protocol in litellm_params was not used.
-    """
-    from litellm.types.router import GenericLiteLLMParams
-
-    # Simulate config.yaml with realtime_protocol as an extra field
-    litellm_params = GenericLiteLLMParams(realtime_protocol="GA")
-    assert litellm_params.get("realtime_protocol") == "GA"
-
-
 @pytest.mark.asyncio
 async def test_arealtime_transcription_intent_defaults_to_ga(monkeypatch):
     """
@@ -742,7 +707,7 @@ async def test_realtime_health_check_uses_bearer_token_when_no_api_key(monkeypat
 
 
 @pytest.mark.asyncio
-async def test_arealtime_forwards_deployment_azure_ad_token(monkeypatch):
+async def test_arealtime_forwards_deployment_azure_ad_token(monkeypatch, no_ambient_azure_credentials):
     """
     The router binds a deployment's `azure_ad_token` to `_arealtime`'s named parameter rather than
     **kwargs, so it must still reach the handler.

tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py

@@ -352,21 +352,6 @@ def test_azure_model_router_stamps_selected_model_on_hidden_params():
     )
 
 
-def test_azure_model_router_stamp_does_not_leak_across_responses():
-    """
-    ModelResponse declares _hidden_params as a class-level dict, so the stamp has to be written
-    as a fresh dict. Mutating in place would bleed the selected model into unrelated responses.
-    """
-    from litellm.llms.azure_ai.common_utils import (
-        AZURE_MODEL_ROUTER_SELECTED_MODEL_KEY,
-    )
-    from litellm.types.utils import ModelResponse
-
-    untouched = ModelResponse()
-
-    assert AZURE_MODEL_ROUTER_SELECTED_MODEL_KEY not in (untouched._hidden_params or {})
-
-
 def test_drop_tool_level_extra_fields_strips_copilot_mcp_server_name():
     """
     Regression test: Azure AI returns 400 when tools contain copilot_mcp_server_name.

tests/unit/llms/azure_ai/image_edit/test_azure_ai_image_edit_transformation.py

@@ -41,7 +41,7 @@ def test_azure_ai_url_generation():
     assert complete_url == expected_url
 
 
-def test_azure_ai_validate_environment_with_entra_token(monkeypatch):
+def test_azure_ai_validate_environment_with_entra_token(monkeypatch, no_ambient_azure_credentials):
     monkeypatch.delenv("AZURE_AI_API_KEY", raising=False)
     monkeypatch.setattr(litellm, "api_key", None)
     config = AzureFoundryFluxImageEditConfig()
@@ -55,7 +55,7 @@ def test_azure_ai_validate_environment_with_entra_token(monkeypatch):
     assert headers == {"Authorization": "Bearer entra-token"}
 
 
-def test_flux2_validate_environment_with_entra_token(monkeypatch):
+def test_flux2_validate_environment_with_entra_token(monkeypatch, no_ambient_azure_credentials):
     monkeypatch.delenv("AZURE_AI_API_KEY", raising=False)
     monkeypatch.setattr(litellm, "api_key", None)
     config = AzureFoundryFlux2ImageEditConfig()

tests/unit/llms/azure_ai/image_edit/test_mai_image_edit_transformation.py

@@ -174,7 +174,7 @@ class TestAzureMAIImageEdit:
         assert image_response.usage.total_tokens == 1024
 
 
-def test_mai_validate_environment_with_entra_token(monkeypatch):
+def test_mai_validate_environment_with_entra_token(monkeypatch, no_ambient_azure_credentials):
     monkeypatch.delenv("AZURE_AI_API_KEY", raising=False)
     monkeypatch.setattr(litellm, "api_key", None)
 

tests/unit/llms/azure_ai/passthrough/test_azure_ai_passthrough_transformation.py

@@ -256,13 +256,13 @@ def test_serverless_host_gets_a_bearer_token():
     assert "api-key" not in headers
 
 
-def test_entra_token_is_used_when_the_deployment_has_no_api_key():
+def test_entra_token_is_used_when_the_deployment_has_no_api_key(no_ambient_azure_credentials):
     headers = _auth_headers(api_key=None, api_base=FOUNDRY_BASE, litellm_params={"azure_ad_token": "entra-token"})
 
     assert headers["Authorization"] == "Bearer entra-token"
 
 
-def test_no_credentials_at_all_raises():
+def test_no_credentials_at_all_raises(no_ambient_azure_credentials):
     with pytest.raises(ValueError, match="Missing Azure AI credentials"):
         _auth_headers(api_key=None, api_base=FOUNDRY_BASE)
 

tests/unit/llms/azure_ai/rerank/test_azure_ai_rerank_transformation.py

@@ -105,7 +105,7 @@ class TestAzureAIRerankConfigValidateEnvironment:
 
         assert headers["Authorization"] == "Bearer my-key"
 
-    def test_falls_back_to_entra_token(self, monkeypatch):
+    def test_falls_back_to_entra_token(self, monkeypatch, no_ambient_azure_credentials):
         monkeypatch.delenv("AZURE_AI_API_KEY", raising=False)
         monkeypatch.setattr(litellm, "azure_key", None)
 

tests/unit/llms/vertex_ai/vertex_ai_partner_models/count_tokens/test_count_tokens_location.py

@@ -98,6 +98,8 @@ class TestCountTokensLocationResolution:
         self, counter, monkeypatch
     ):
         """Claude models without any location should default to us-east5."""
+        monkeypatch.delenv("VERTEXAI_LOCATION", raising=False)
+        monkeypatch.delenv("VERTEX_LOCATION", raising=False)
         captured = {}
 
         async def fake_ensure_access_token(

tests/unit/llms/vertex_ai/vertex_ai_partner_models/mistral/test_vertex_ai_partner_models_mistral_transformation.py

@@ -1,6 +1,21 @@
+import pytest
+
 import litellm
 
 
+@pytest.fixture
+def local_model_cost_map(monkeypatch):
+    original_model_cost = litellm.model_cost
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+    litellm.get_model_info.cache_clear()
+    try:
+        yield
+    finally:
+        litellm.model_cost = original_model_cost
+        litellm.get_model_info.cache_clear()
+
+
 def test_reasoning_effort_stays_unsupported_on_vertex_partner_models(local_model_cost_map):
     assert "reasoning_effort" in litellm.get_supported_openai_params(
         model="mistral-medium-3", custom_llm_provider="mistral"