feat: lazy load heavy imports to reduce memory usage at import time (#18592)

- Lazy load remove_index_from_tool_calls via __getattr__
- Lazy load _service_logger module via __getattr__ in __init__.py
- Lazy load audio_utils.utils with module-level caching in utils.py
- Remove direct imports that trigger heavy module loading

These changes reduce import-time memory usage by deferring imports
until they are actually needed, while maintaining performance through
caching for subsequent accesses.

litellm/__init__.py

@@ -1058,7 +1058,7 @@ openai_video_generation_models = ["sora-2"]
 
 # timeout is lazy-loaded via __getattr__
 # get_llm_provider is lazy-loaded via __getattr__
-from litellm.litellm_core_utils.core_helpers import remove_index_from_tool_calls
+# remove_index_from_tool_calls is lazy-loaded via __getattr__
 
 # Import KeyManagementSettings here (before utils import) because _key_management_settings
 # is accessed during import time in secret_managers/main.py (via dd_tracing -> datadog -> _service_logger -> utils)
@@ -1507,6 +1507,7 @@ if TYPE_CHECKING:
     get_first_chars_messages: Callable[..., str]
     get_provider_fields: Callable[..., List]
     get_valid_models: Callable[..., list]
+    remove_index_from_tool_calls: Callable[..., None]
 
     # Response types - truly lazy loaded only (not in main.py or elsewhere)
     ModelResponseListIterator: Type[Any]
@@ -1658,6 +1659,17 @@ def __getattr__(name: str) -> Any:
             LoggingCallbackManager = __getattr__("LoggingCallbackManager")
             _globals["logging_callback_manager"] = LoggingCallbackManager()
         return _globals["logging_callback_manager"]
+    
+    # Lazy load _service_logger module
+    if name == "_service_logger":
+        from ._lazy_imports import _get_litellm_globals
+        _globals = _get_litellm_globals()
+        # Check if already cached
+        if "_service_logger" not in _globals:
+            # Import the module lazily
+            import litellm._service_logger
+            _globals["_service_logger"] = litellm._service_logger
+        return _globals["_service_logger"]
 
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 

litellm/_lazy_imports_registry.py

@@ -290,6 +290,7 @@ TYPES_NAMES = (
 # LLM provider logic names that support lazy loading via _lazy_import_llm_provider_logic
 LLM_PROVIDER_LOGIC_NAMES = (
     "get_llm_provider",
+    "remove_index_from_tool_calls",
 )
 
 # Import maps for registry pattern - reduces repetition
@@ -393,6 +394,7 @@ _TYPES_IMPORT_MAP = {
 
 _LLM_PROVIDER_LOGIC_IMPORT_MAP = {
     "get_llm_provider": ("litellm.litellm_core_utils.get_llm_provider_logic", "get_llm_provider"),
+    "remove_index_from_tool_calls": ("litellm.litellm_core_utils.core_helpers", "remove_index_from_tool_calls"),
 }
 
 _LLM_CONFIGS_IMPORT_MAP = {

litellm/utils.py

@@ -48,9 +48,9 @@ from tiktoken import Encoding
 from tokenizers import Tokenizer
 
 import litellm
-import litellm._service_logger  # for storing API inputs, outputs, and metadata
+
 import litellm.litellm_core_utils
-import litellm.litellm_core_utils.audio_utils.utils
+# audio_utils.utils is lazy-loaded - only imported when needed for transcription calls
 import litellm.litellm_core_utils.json_validation_rule
 import litellm.llms
 import litellm.llms.gemini
@@ -157,6 +157,24 @@ from litellm.router_utils.get_retry_from_policy import (
     reset_retry_policy,
 )
 from litellm.secret_managers.main import get_secret
+
+# Cached lazy import for audio_utils.utils
+# Module-level cache to avoid repeated imports while preserving memory benefits
+_audio_utils_module = None
+
+
+def _get_cached_audio_utils():
+    """
+    Get cached audio_utils.utils module.
+    Lazy imports on first call to avoid loading audio_utils.utils at import time.
+    Subsequent calls use cached module for better performance.
+    """
+    global _audio_utils_module
+    if _audio_utils_module is None:
+        import litellm.litellm_core_utils.audio_utils.utils
+        _audio_utils_module = litellm.litellm_core_utils.audio_utils.utils
+    return _audio_utils_module
+
 from litellm.types.llms.anthropic import (
     ANTHROPIC_API_ONLY_HEADERS,
     AnthropicThinkingParam,
@@ -965,7 +983,9 @@ def function_setup(  # noqa: PLR0915
             or call_type == CallTypes.transcription.value
         ):
             _file_obj: FileTypes = args[1] if len(args) > 1 else kwargs["file"]
-            file_checksum = litellm.litellm_core_utils.audio_utils.utils.get_audio_file_content_hash(
+            # Lazy import audio_utils.utils only when needed for transcription calls
+            audio_utils = _get_cached_audio_utils()
+            file_checksum = audio_utils.get_audio_file_content_hash(
                 file_obj=_file_obj
             )
             if "metadata" in kwargs: