Need a better way to handle this then just to remove all variables until a common set is found.
e.g. Mistral models on Azure AI don't support frequency_penalty or presence_penalty.
There might be a way to query and endpoint first to determine what variables are supported.