Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

Bug Description

Custom LLM providers registered via custom_provider_map are silently bypassed when the underlying model name (after prefix stripping by get_llm_provider()) matches a known built-in model. The request goes directly to the built-in provider (e.g. OpenAI) instead of the custom handler. No error is raised.

Fix

Move the _custom_providers dispatch block to the top of the if/elif chain — before all built-in provider checks. Explicitly registered custom providers should always take priority.

I have a PR ready with this fix.

Steps to Reproduce

1. Custom provider (mcs_provider.py)

from litellm.llms.custom_llm import CustomLLM
from litellm.utils import ModelResponse

class MCSProvider(CustomLLM):
    async def acompletion(self, model, messages, model_response, **kwargs) -> ModelResponse:
        # This method is NEVER called
        model_response.choices[0].message.content = "Hello from custom provider"
        model_response.choices[0].finish_reason = "stop"
        return model_response

mcs_handler = MCSProvider()

2. Config (config.yaml)

model_list:
  - model_name: "mcs/*"
    litellm_params:
      model: "mcs-provider/*"

litellm_settings:
  custom_provider_map:
    - {"provider": "mcs-provider", "custom_handler": "mcs_provider.mcs_handler"}

3. Test

litellm --config config.yaml --port 4000

curl http://localhost:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "mcs/gpt-5.4", "messages": [{"role": "user", "content": "Hi"}]}'

Expected: Response comes from the custom handler. Actual: Response comes directly from OpenAI. The custom handler is never called.

Root Cause

In litellm/main.py, the completion() function has a long if/elif dispatch chain. get_llm_provider() strips the custom provider prefix (mcs-provider/gpt-5.4 → gpt-5.4), and then the branch at ~line 2524:

elif (
    model in litellm.open_ai_chat_completion_models  # "gpt-5.4" matches here!
    or custom_llm_provider == "openai"
    ...
):

is reached before the custom provider check at ~line 4282:

elif (
    custom_llm_provider in litellm._custom_providers  # never reached
):

This affects any custom provider that routes to a model name known to any built-in provider (OpenAI, Anthropic, Cohere, etc.).

Relevant log output

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.82.1

Twitter / LinkedIn details

www.linkedin.com/in/dannygerst/