fix: resolve memory leak caused by Pydantic 2.11+ deprecation warnings (#16110)

- Access model_fields on class instead of instance to avoid deprecation warnings
- Add warning filter to suppress Pydantic instance attribute access warnings
- Fix affects streaming operations where warnings accumulate over time
- Updated cost_calculator.py, core_helpers.py, and model_response_utils.py

The previous implementation accessed model_fields on Pydantic model instances,
which triggers deprecation warnings in Pydantic 2.11+. These warnings accumulate
during streaming operations, causing memory leaks. Now accessing model_fields
via type(instance).model_fields to prevent warning generation.

litellm/__init__.py

@@ -2,6 +2,9 @@
 import warnings
 
 warnings.filterwarnings("ignore", message=".*conflict with protected namespace.*")
+# Suppress Pydantic 2.11+ deprecation warning about accessing model_fields on instances
+# This warning can accumulate during streaming and cause memory leaks
+warnings.filterwarnings("ignore", message=".*Accessing the.*attribute on the instance is deprecated.*")
 ### INIT VARIABLES ######################
 import threading
 import os

litellm/cost_calculator.py

@@ -1595,7 +1595,8 @@ class BaseTokenUsageProcessor:
                     combined.prompt_tokens_details = PromptTokensDetailsWrapper()
 
                 # Check what keys exist in the model's prompt_tokens_details
-                for attr in usage.prompt_tokens_details.model_fields:
+                # Access model_fields on the class, not the instance, to avoid Pydantic 2.11+ deprecation warnings
+                for attr in type(usage.prompt_tokens_details).model_fields:
                     if (
                         hasattr(usage.prompt_tokens_details, attr)
                         and not attr.startswith("_")
@@ -1626,7 +1627,8 @@ class BaseTokenUsageProcessor:
                     )
 
                 # Check what keys exist in the model's completion_tokens_details
-                for attr in usage.completion_tokens_details.model_fields:
+                # Access model_fields on the class, not the instance, to avoid Pydantic 2.11+ deprecation warnings
+                for attr in type(usage.completion_tokens_details).model_fields:
                     if not attr.startswith("_") and not callable(
                         getattr(usage.completion_tokens_details, attr)
                     ):

litellm/litellm_core_utils/core_helpers.py

@@ -234,7 +234,8 @@ def preserve_upstream_non_openai_attributes(
     """
     Preserve non-OpenAI attributes from the original chunk.
     """
-    expected_keys = set(model_response.model_fields.keys()).union({"usage"})
+    # Access model_fields on the class, not the instance, to avoid Pydantic 2.11+ deprecation warnings
+    expected_keys = set(type(model_response).model_fields.keys()).union({"usage"})
     for key, value in original_chunk.model_dump().items():
         if key not in expected_keys:
             setattr(model_response, key, value)

litellm/litellm_core_utils/model_response_utils.py

@@ -46,7 +46,8 @@ def is_model_response_stream_empty(model_response: ModelResponseStream) -> bool:
                 return False
 
     # Check for any non-base fields that are set
-    for model_response_field in model_response.model_fields.keys():
+    # Access model_fields on the class, not the instance, to avoid Pydantic 2.11+ deprecation warnings
+    for model_response_field in type(model_response).model_fields.keys():
         # Skip base fields that are always set
         if model_response_field in BASE_FIELDS:
             continue