chore: Defer import of live, Client and _transformers in google.genai

This is targeting the Client import mostly, but also prevents future latency increase if the other modules in genai adds more 3p dependencies. This change will make ADK only import `live`, `Client` and `_transformers` just-in-time, therefore cutting down cold start latency.

Co-authored-by: Liang Wu <wuliang@google.com>
PiperOrigin-RevId: 831124329

src/google/adk/flows/llm_flows/instructions.py

@@ -19,7 +19,6 @@ from __future__ import annotations
 from typing import AsyncGenerator
 from typing import TYPE_CHECKING
 
-from google.genai import _transformers
 from typing_extensions import override
 
 from ...agents.readonly_context import ReadonlyContext
@@ -85,6 +84,8 @@ class _InstructionsLlmRequestProcessor(BaseLlmRequestProcessor):
 
     # Handle static_instruction - add via append_instructions
     if agent.static_instruction:
+      from google.genai import _transformers
+
       # Convert ContentUnion to Content using genai transformer
       static_content = _transformers.t_content(agent.static_instruction)
       llm_request.append_instructions(static_content)

src/google/adk/models/apigee_llm.py

@@ -18,12 +18,10 @@ from __future__ import annotations
 from functools import cached_property
 import logging
 import os
-import re
 from typing import Optional
 from typing import TYPE_CHECKING
 
 from google.adk import version as adk_version
-from google.genai import Client
 from google.genai import types
 from typing_extensions import override
 
@@ -31,8 +29,11 @@ from ..utils.env_utils import is_env_enabled
 from .google_llm import Gemini
 
 if TYPE_CHECKING:
+  from google.genai import Client
+
   from .llm_request import LlmRequest
 
+
 logger = logging.getLogger('google_adk.' + __name__)
 
 _APIGEE_PROXY_URL_ENV_VARIABLE_NAME = 'APIGEE_PROXY_URL'
@@ -137,6 +138,7 @@ class ApigeeLlm(Gemini):
     Returns:
       The api client.
     """
+    from google.genai import Client
 
     kwargs_for_http_options = {}
     if self._api_version:

src/google/adk/models/gemini_context_cache_manager.py

@@ -21,8 +21,8 @@ import json
 import logging
 import time
 from typing import Optional
+from typing import TYPE_CHECKING
 
-from google.genai import Client
 from google.genai import types
 
 from ..utils.feature_decorator import experimental
@@ -32,6 +32,9 @@ from .llm_response import LlmResponse
 
 logger = logging.getLogger("google_adk." + __name__)
 
+if TYPE_CHECKING:
+  from google.genai import Client
+
 
 @experimental
 class GeminiContextCacheManager:

src/google/adk/models/gemini_llm_connection.py

@@ -18,7 +18,6 @@ import logging
 from typing import AsyncGenerator
 from typing import Union
 
-from google.genai import live
 from google.genai import types
 
 from ..utils.context_utils import Aclosing
@@ -28,6 +27,10 @@ from .llm_response import LlmResponse
 logger = logging.getLogger('google_adk.' + __name__)
 
 RealtimeInput = Union[types.Blob, types.ActivityStart, types.ActivityEnd]
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+  from google.genai import live
 
 
 class GeminiLlmConnection(BaseLlmConnection):

src/google/adk/models/google_llm.py

@@ -27,7 +27,6 @@ from typing import Optional
 from typing import TYPE_CHECKING
 from typing import Union
 
-from google.genai import Client
 from google.genai import types
 from typing_extensions import override
 
@@ -41,6 +40,8 @@ from .gemini_llm_connection import GeminiLlmConnection
 from .llm_response import LlmResponse
 
 if TYPE_CHECKING:
+  from google.genai import Client
+
   from .llm_request import LlmRequest
 
 logger = logging.getLogger('google_adk.' + __name__)
@@ -200,6 +201,8 @@ class Gemini(BaseLlm):
     Returns:
       The api client.
     """
+    from google.genai import Client
+
     return Client(
         http_options=types.HttpOptions(
             headers=self._tracking_headers,
@@ -239,6 +242,8 @@ class Gemini(BaseLlm):
 
   @cached_property
   def _live_api_client(self) -> Client:
+    from google.genai import Client
+
     return Client(
         http_options=types.HttpOptions(
             headers=self._tracking_headers, api_version=self._live_api_version

tests/unittests/models/test_apigee_llm.py

@@ -46,7 +46,7 @@ def llm_request():
 
 
 @pytest.mark.asyncio
-@mock.patch('google.adk.models.apigee_llm.Client')
+@mock.patch('google.genai.Client')
 async def test_generate_content_async_non_streaming(
     mock_client_constructor, llm_request
 ):
@@ -96,7 +96,7 @@ async def test_generate_content_async_non_streaming(
 
 
 @pytest.mark.asyncio
-@mock.patch('google.adk.models.apigee_llm.Client')
+@mock.patch('google.genai.Client')
 async def test_generate_content_async_streaming(
     mock_client_constructor, llm_request
 ):
@@ -167,7 +167,7 @@ async def test_generate_content_async_streaming(
 
 
 @pytest.mark.asyncio
-@mock.patch('google.adk.models.apigee_llm.Client')
+@mock.patch('google.genai.Client')
 async def test_generate_content_async_with_custom_headers(
     mock_client_constructor, llm_request
 ):
@@ -207,7 +207,7 @@ async def test_generate_content_async_with_custom_headers(
 
 
 @pytest.mark.asyncio
-@mock.patch('google.adk.models.apigee_llm.Client')
+@mock.patch('google.genai.Client')
 async def test_vertex_model_path_parsing(mock_client_constructor):
   """Tests that Vertex AI model paths are parsed correctly."""
   apigee_llm = ApigeeLlm(model=APIGEE_VERTEX_MODEL_ID, proxy_url=PROXY_URL)
@@ -249,7 +249,7 @@ async def test_vertex_model_path_parsing(mock_client_constructor):
 
 
 @pytest.mark.asyncio
-@mock.patch('google.adk.models.apigee_llm.Client')
+@mock.patch('google.genai.Client')
 async def test_proxy_url_from_env_variable(mock_client_constructor):
   """Tests that proxy_url is read from environment variable."""
   with mock.patch.dict(
@@ -381,7 +381,7 @@ def test_vertex_model_missing_project_or_location_raises_error(
         ),
     ],
 )
-@mock.patch('google.adk.models.apigee_llm.Client')
+@mock.patch('google.genai.Client')
 async def test_model_string_parsing_and_client_initialization(
     mock_client_constructor,
     model_string,