fix: set litellm.drop_params to support non-OpenAI providers (#63)

[smwaqas89]

Summary

Fixes #63. Also addresses the issue described in #32.

Problem

When using non-OpenAI providers (Ollama, Groq, local models) as the weak model, litellm raises UnsupportedParamsError because parameters like presence_penalty and frequency_penalty are not supported by all providers.

The ChatCompletionRequest in openai_server.py defaults these to 0.0, and model_dump(exclude_none=True) always includes them since they're not None. When litellm forwards these to providers that don't support them, it crashes.

Solution

Set litellm.drop_params = True at module level in controller.py. This is litellm's built-in mechanism for handling provider compatibility — unsupported parameters are silently dropped rather than raising errors.

This enables RouteLLM to work seamlessly with any provider supported by litellm (Ollama, Groq, Together AI, local models, etc.) without requiring users to manually configure parameter handling.

Changes

• routellm/controller.py: Added import litellm and set litellm.drop_params = True

Testing

• Verified drop_params flag is set on module load
• Tested with Ollama as weak model — no more UnsupportedParamsError
• No impact on OpenAI/Anyscale providers (they already support these params)