if self.topk_method == "noaux_tc":
    scores_for_choice = scores.view(bsz * seq_len, -1) + self.e_score_correction_bias.unsqueeze(0)
    group_scores = (
    scores_for_choice.view(bsz * seq_len, self.n_group, -1).topk(2, dim=-1)[0].sum(dim = -1)

    num_group = (
        parallel_state.get_expert_model_parallel_world_size()
    )  # num_group equals to expert parallel size
    group_scores = scores.view(num_tokens, num_group, -1).max(dim=-1).values
    group_idx = torch.topk(group_scores, k=moe_router_topk_limited_devices, dim=-1, sorted=False)[1]