fix(proxy): keep internal callbacks when litellm_settings.callbacks is a single string

When litellm_settings.callbacks was a single string instead of a list, the proxy
overwrote litellm.callbacks with just that callback, dropping the internal
callbacks registered at import, including the per-model max budget limiter. Keys
with model_max_budget then went over budget without ever being blocked. The
string form now appends, the same way the list form already extends

litellm/proxy/common_utils/callback_utils.py

@@ -404,7 +404,7 @@ def initialize_callbacks_on_proxy(
 
             PrometheusLogger._mount_metrics_endpoint()
     else:
-        litellm.callbacks = [
+        litellm.callbacks.append(
             _loaded_callback_or_raise(
                 entry=value,
                 loaded=get_instance_fn(
@@ -412,7 +412,7 @@ def initialize_callbacks_on_proxy(
                     config_file_path=config_file_path,
                 ),
             )
-        ]
+        )
     verbose_proxy_logger.debug("%s Initialized Callbacks - %s %s", blue_color_code, litellm.callbacks, reset_color_code)
 
 

tests/unit/proxy/common_utils/test_callback_utils.py

@@ -762,3 +762,15 @@ def test_initialize_callbacks_on_proxy_accepts_instance_non_list_value(probe_con
 
     assert len(litellm.callbacks) == 1
     assert isinstance(litellm.callbacks[0], CustomLogger)
+
+
+@pytest.mark.parametrize("as_list", [False, True], ids=["string_value", "list_value"])
+def test_initialize_callbacks_on_proxy_keeps_previously_registered_callbacks(probe_config_path: str, as_list: bool):
+    entry: Final = f"{_PROBE_MODULE_NAME}.proxy_handler_instance"
+    internal_callback: Final = CustomLogger()
+    litellm.callbacks.append(internal_callback)
+
+    _load_callbacks([entry] if as_list else entry, probe_config_path)
+
+    assert litellm.callbacks[0] is internal_callback
+    assert [type(cb).__name__ for cb in litellm.callbacks[1:]] == ["FloorMaxTokens"]