Can the LLM engine be replaced by litellm?

[linhaowei1]

Different LLMs have varying APIs. For example, the o-series requires the "max_completion_tokens" parameter and does not support "temperature," but this is currently hard-coded in the code. When I tested GPT-5, it followed the same format as the o-series, so we had to add a specific rule to accommodate it. If users rely on self-hosted models or APIs from specialized providers, managing the LLM engines could become quite messy. I suggest that OpenEvolve could integrate a unified LLM proxy like LiteLLM, which can automatically handle different request formats appropriately.

[codelion]

what do we need to do to support gpt-5? does it use the same max_completion_tokens?

We can consider adding support for LiteLLM, for the time been you can use it or OptiLLM as a proxy and point the base url to http://localhost/v1.

[linhaowei1]

GPT-5 seems to use the same interface as the "o-series" models (e.g., similar to GPT-4o), including the max_completion_tokens parameter. To add support for it, simply update the model detection logic by adding the following check after if str(model_names).startswith('o'). (e.g., ... or if 'gpt-5' in str(model_names)). This should suffice for basic compatibility.

OptiLLM seems like a great option. We could also add some guidelines in the documentation for setting up and managing LLM engines to make this process smoother for users. Add many "if-else" may be not elegant (though it's very fast..)?

[codelion]

@linhaowei1 yes good idea, happy to review a PR and add support gpt-5 for now and update the documentation. In future I will review the LLM engine implementation to see how to support more LLM providers better.