[Feature] Support for decoding method that reduce Hallucinations

[zhly0]

Motivation

Recently，there are many good paper that try to alleviating hallucinations for large vision-language models during the decode process，like： OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation (CVPR 2024 Highlight)（code:https://github.com/shikiw/OPERA） Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs (ECCV 2024)(code:https://github.com/LALBJ/PAI) Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (code:https://github.com/huofushuo/SID) All these method and code were implemented during inference process. Is it possible to integrate one or more these method into InternVL2.0 series model decode in lmdeploy? Thank you!

Related resources

OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation (CVPR 2024 Highlight)（code:https://github.com/shikiw/OPERA） Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs (ECCV 2024)(code:https://github.com/LALBJ/PAI) Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (code:https://github.com/huofushuo/SID)

Additional context

No response

[lvhan028]

@irexyc may put this feature in the survey list