fix(proxy): model_call_details wins for timing; skip post-guardrail task when no chunks

- utils: let model_call_details override start_time/end_time in kwargs merge; add comment
- proxy_server: only create post-guardrail log task when _streaming_chunks_for_log is non-empty

Made-with: Cursor

litellm/proxy/proxy_server.py

@@ -5609,8 +5609,8 @@ async def async_data_generator(
             except Exception as e:
                 yield f"data: {str(e)}\n\n"
 
-        # Post-guardrail log: only for successful stream completion, not when we got an SSE error
-        if error_message is None:
+        # Post-guardrail log: only for successful stream completion with chunks to process
+        if error_message is None and _streaming_chunks_for_log:
             def _discard_task(t: asyncio.Task[None]) -> None:
                 _post_guardrail_log_tasks.discard(t)
 

litellm/proxy/utils.py

@@ -1994,7 +1994,12 @@ class ProxyLogging:
             if logging_obj is not None and getattr(
                 logging_obj, "model_call_details", None
             ):
+                # Merge so request data wins overall, but model_call_details wins for
+                # LLM-lifecycle timing fields so callbacks get accurate start/end times.
                 kwargs = {**logging_obj.model_call_details, **kwargs}
+                for key in ("start_time", "end_time"):
+                    if key in logging_obj.model_call_details:
+                        kwargs[key] = logging_obj.model_call_details[key]
             kwargs["user_api_key_dict"] = user_api_key_dict
             start_time = None
             if logging_obj is not None: