Does beam-search need to be adjusted?

[lijiabao2]

Thanks for your reply, I noticed my problem.
I have successfully run your code, but I noticed a problem. It seems that in beam-search mode, temperature, top_p and top_k need to be set to None. Do I need to modify base_hf_causal_model.py? The problem appears when running graph_constrained_decoding.sh.

    if self.args.generation_mode == "greedy":
        self.generation_cfg.do_sample = False
        self.generation_cfg.num_return_sequences = 1
    elif self.args.generation_mode == "sampling":
        self.generation_cfg.do_sample = True
        self.generation_cfg.num_return_sequences = self.args.k
    elif self.args.generation_mode == "beam":
        self.generation_cfg.do_sample = False
        self.generation_cfg.num_beams = self.args.k
        self.generation_cfg.num_return_sequences = self.args.k
    elif self.args.generation_mode == "beam-early-stopping":
        self.generation_cfg.do_sample = False
        self.generation_cfg.num_beams = self.args.k
        self.generation_cfg.num_return_sequences = self.args.k
        self.generation_cfg.early_stopping = True
    elif self.args.generation_mode == "group-beam":
        self.generation_cfg.do_sample = False
        self.generation_cfg.num_beams = self.args.k
        self.generation_cfg.num_return_sequences = self.args.k
        self.generation_cfg.num_beam_groups = self.args.k
        self.generation_cfg.diversity_penalty = 1.
    elif self.args.generation_mode == "group-beam-early-stopping":
        self.generation_cfg.do_sample = False
        self.generation_cfg.num_beams = self.args.k
        self.generation_cfg.num_return_sequences = self.args.k
        self.generation_cfg.num_beam_groups = self.args.k
        self.generation_cfg.early_stopping = True
        self.generation_cfg.diversity_penalty = 1.

Best regards，
lijiabao