fix litellm_params accessor

litellm/litellm_core_utils/streaming_handler.py

@@ -15,6 +15,7 @@ from litellm import verbose_logger
 from litellm.litellm_core_utils.redact_messages import LiteLLMLoggingObject
 from litellm.litellm_core_utils.thread_pool_executor import executor
 from litellm.types.llms.openai import ChatCompletionChunk
+from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import Delta
 from litellm.types.utils import GenericStreamingChunk as GChunk
 from litellm.types.utils import (
@@ -71,6 +72,12 @@ class CustomStreamWrapper:
         self.sent_first_chunk = False
         self.sent_last_chunk = False
 
+        litellm_params: GenericLiteLLMParams = self.logging_obj.model_call_details.get(
+            "litellm_params", {}
+        )
+        self.merge_reasoning_content_in_choices = (
+            litellm_params.merge_reasoning_content_in_choices
+        )
         self.sent_first_thinking_block = False
         self.sent_last_thinking_block = False
         self.thinking_content = ""
@@ -92,12 +99,7 @@ class CustomStreamWrapper:
         self.holding_chunk = ""
         self.complete_response = ""
         self.response_uptil_now = ""
-        _model_info = (
-            self.logging_obj.model_call_details.get("litellm_params", {}).get(
-                "model_info", {}
-            )
-            or {}
-        )
+        _model_info: Dict = litellm_params.model_info or {}
 
         _api_base = get_api_base(
             model=model or "",

litellm/proxy/proxy_config.yaml

@@ -4,11 +4,9 @@ model_list:
       model: bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0
       thinking: {"type": "enabled", "budget_tokens": 1024}
       max_tokens: 1080
+      merge_reasoning_content_in_choices: true
 
 
 general_settings:
   store_model_in_db: true
   store_prompts_in_spend_logs: true
-
-litellm_settings:
-  merge_reasoning_content_in_choices: true

litellm/types/router.py

@@ -192,6 +192,8 @@ class GenericLiteLLMParams(BaseModel):
     budget_duration: Optional[str] = None
     use_in_pass_through: Optional[bool] = False
     model_config = ConfigDict(extra="allow", arbitrary_types_allowed=True)
+    merge_reasoning_content_in_choices: Optional[bool] = False
+    model_info: Optional[Dict] = None
 
     def __init__(
         self,
@@ -231,6 +233,9 @@ class GenericLiteLLMParams(BaseModel):
         budget_duration: Optional[str] = None,
         # Pass through params
         use_in_pass_through: Optional[bool] = False,
+        # This will merge the reasoning content in the choices
+        merge_reasoning_content_in_choices: Optional[bool] = False,
+        model_info: Optional[Dict] = None,
         **params,
     ):
         args = locals()

litellm/types/utils.py

@@ -1803,6 +1803,7 @@ all_litellm_params = [
     "max_budget",
     "budget_duration",
     "use_in_pass_through",
+    "merge_reasoning_content_in_choices",
 ] + list(StandardCallbackDynamicParams.__annotations__.keys())