fix(wrapper): drop mutable call-kwargs tracker, assert exceptions at dispatch time

litellm/utils.py

@@ -2198,7 +2198,6 @@ def client(original_function):
         kwargs.pop("_is_litellm_internal_call", None)  # discard if injected
         _is_litellm_internal_call: Final = is_internal_call.get()
         _deployment_call_end_time: datetime.datetime | None = None
-        uncached_call_kwargs: dict | None = None
 
         try:
             if logging_obj is None:
@@ -2309,8 +2308,6 @@ def client(original_function):
                 and _caching_handler_response.embedding_uncached_input is not None
                 else kwargs
             )
-            if call_kwargs is not kwargs:
-                uncached_call_kwargs = call_kwargs
             try:
                 result = await original_function(*args, **call_kwargs)
             except Exception as deployment_error:
@@ -2323,6 +2320,11 @@ def client(original_function):
                     )
                 except BaseException:  # noqa: S110, BLE001  # hook dispatch - including cancellation mid-await - must never replace the real deployment failure, so there is nothing to do with what it raises
                     pass
+                if call_kwargs is not kwargs:
+                    # The uncached-input dict is a wrapper-private copy of kwargs whose
+                    # request never left this frame except as part of the failure
+                    # traceback below, so its logging reference is released here.
+                    call_kwargs.pop("litellm_logging_obj", None)
                 raise
             end_time = datetime.datetime.now()
 
@@ -2505,8 +2507,6 @@ def client(original_function):
             if logging_obj is not None and not _is_litellm_internal_call:
                 logging_obj.model_call_details.pop("exception", None)
             kwargs.pop("litellm_logging_obj", None)
-            if uncached_call_kwargs is not None:
-                uncached_call_kwargs.pop("litellm_logging_obj", None)
             raise e
 
         finally:

tests/unit/integrations/test_custom_callback_input.py

@@ -222,6 +222,9 @@ class _CallbackCapture(CustomLogger):
         self.async_success: Final = AsyncMock()
         self.sync_failure: Final = Mock()
         self.async_failure: Final = AsyncMock()
+        # Exception seen at dispatch time; litellm releases it from the call-details
+        # dict once the failure handlers have run (issue #48649).
+        self.async_failure_exceptions: Final = []
 
     def log_success_event(self, kwargs, response_obj, start_time, end_time) -> None:
         self.sync_success(kwargs=kwargs, response_obj=response_obj)
@@ -234,6 +237,7 @@ class _CallbackCapture(CustomLogger):
         self.sync_failure(kwargs=kwargs, response_obj=response_obj)
 
     async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time) -> None:
+        self.async_failure_exceptions.append(kwargs.get("exception"))
         await self.async_failure(kwargs=kwargs, response_obj=response_obj)
 
 
@@ -501,7 +505,7 @@ async def test_async_embedding_failure_reaches_failure_callback(
     )
     assert len(failure_events) == 1
     assert failure_events[0]["model"] == "text-embedding-3-small"
-    assert isinstance(failure_events[0]["exception"], litellm.AuthenticationError)
+    assert isinstance(callback_capture.async_failure_exceptions[0], litellm.AuthenticationError)
     assert callback_capture.async_success.call_args_list == []
 
 

tests/unit/test_router/test_router_callback_hook_sequence.py

@@ -85,6 +85,7 @@ class _HookEvent(NamedTuple):
     model: object
     kwargs: Mapping[str, object]
     response: object
+    exception: object
 
 
 def _router_context_problems(kwargs: Mapping[str, object]) -> tuple[str, ...]:
@@ -156,7 +157,9 @@ class _HookRecorder(CustomLogger):
         if not _is_from_this_test(kwargs):
             return
         self.errors.extend(f"{state}: {problem}" for problem in problems)
-        self.events.append(_HookEvent(state, model, kwargs, response))
+        # Snapshot the exception at dispatch time: litellm no longer keeps it on the
+        # call-details dict after the failure handlers have run (issue #48649).
+        self.events.append(_HookEvent(state, model, kwargs, response, kwargs.get("exception")))
         if self.loop is not None:
             self.loop.call_soon_threadsafe(self._notify)
 
@@ -344,7 +347,7 @@ async def test_router_chat_success_streaming_and_failure_fire_the_hooks_in_order
     failed: Final = recorder.events[6:]
     assert [event.state for event in failed] == ["sync_pre_api_call", "post_api_call", "async_failure"]
     assert failed[2].response is None
-    assert isinstance(failed[2].kwargs["exception"], litellm.AuthenticationError)
+    assert isinstance(failed[2].exception, litellm.AuthenticationError)
     assert recorder.errors == []
 
 
@@ -372,7 +375,7 @@ async def test_router_embedding_success_and_failure_fire_the_hooks_in_order(
     await recorder.until(_settled(2, posts=2))
     assert recorder.states[3:] == ["sync_pre_api_call", "post_api_call", "async_failure"]
     assert recorder.events[5].response is None
-    assert isinstance(recorder.events[5].kwargs["exception"], litellm.AuthenticationError)
+    assert isinstance(recorder.events[5].exception, litellm.AuthenticationError)
     assert recorder.errors == []
 
 
@@ -425,7 +428,7 @@ async def test_router_fallback_fires_failure_then_success_hooks(
     ]
     assert [_model_group(event) for event in recorder.events] == ["primary-group"] * 3 + ["fallback-group"] * 3
     assert _model_id(recorder.events[0]) != _model_id(recorder.events[3])
-    assert isinstance(recorder.events[2].kwargs["exception"], litellm.AuthenticationError)
+    assert isinstance(recorder.events[2].exception, litellm.AuthenticationError)
     assert _completion(recorder.events[5]).choices[0].message.content == "hello"
     assert recorder.errors == []