pre_call_hook, moderation_hook, post_call_success_hook not called when using Python SDK (litellm.completion)

[ayrayane]

When using the liteLLM Python SDK directly (not the proxy), custom guardrails registered via
litellm.guardrail_name_config_map and litellm.callbacks are never triggered.

Expected behavior:

apply_guardrail and async_pre_call_hook should be called when passing : extra_body={"guardrails": ["my_guardrail"]} to litellm.completion

Actual behavior:

Neither apply_guardrail nor async_pre_call_hook are called. Only init is called at startup.

Reproduction :

form litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.types.guardrails import GuardrailItem
import litellm

Class MyGuardrail(CustomGuardrail)
async def apply_guardrail(self, input, request_data, input_type, logging_obj=None):
print("CALLED")
return inputs
guardrail = MyGuardrail(guardrail_name="my_guardrail")
litellm.callbacks = [guardrail]
litellm.guardrail_name_config_map = {
"my_guardrail": GuardrailItem(
guardrail_name="my_guardrail",
callbacks=["my_guardrail"]
)
}
litellm.completion(
mode_"openai/gpt-4o",
messages=[{"role":"user", "content":"hello"}],
extra_body={"guardrails": ["my_guardrail"]}
)

Question:

Is apply_guardrail and other hook supported in the python SDK or only in the proxy ? What is the recommended way to apply custom guardrails with SDK ?

[smqd19]

The hooks like apply_guardrail and async_pre_call_hook you’ve defined won’t be triggered when using the Python SDK (litellm.completion) because they’re currently designed to work in the proxy-based flow. The SDK bypasses the guardrail orchestration logic that the proxy enforces, so custom hooks tied to guardrails won't execute unless explicitly integrated into the SDK's processing pipeline.

If you want to use guardrails in the SDK, you’ll need to explicitly call them within your wrapper logic. For example, you can manually invoke apply_guardrail before or after the SDK call. Here’s how you could structure it:

from litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.types.guardrails import GuardrailItem
import litellm

class MyGuardrail(CustomGuardrail):
    async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
        print("CALLED")
        return inputs

guardrail = MyGuardrail(guardrail_name="my_guardrail")
litellm.guardrail_name_config_map = {
    "my_guardrail": GuardrailItem(
        guardrail_name="my_guardrail",
        callbacks=["my_guardrail"]
    )
}

# Custom logic to apply guardrails manually
async def guarded_completion():
    inputs = {"messages": [{"role": "user", "content": "hello"}]}
    preprocessed_inputs = await guardrail.apply_guardrail(inputs, {}, "input_type")
    response = litellm.completion(
        mode="openai/gpt-4o",
        **preprocessed_inputs,
        extra_body={"guardrails": ["my_guardrail"]}  # This won’t trigger the guardrail
    )
    return response

If this is a core feature you need, you might want to open a PR or raise an issue requesting tighter SDK integration with litellm.guardrail_name_config_map and callbacks. Otherwise, manually invoking guardrails like shown above is your best call for now.