[Feature] ensure greedy decoding with huggingface model

[Leymore]

Describe the feature

something like

from transformers.generation import GenerationConfig
self.model.generation_config = GenerationConfig.from_pretrained(path, trust_remote_code=True)
self.model.generation_config.do_sample = False

Will you implement it?

• [X] I would like to implement this feature and create a PR!