You can later instantiate them with [GenerationConfig.from_pretrained]. This is useful if you want to
store several generation configurations for a single model (e.g. one for creative text generation with sampling, and
one for summarization with beam search).