from transformers.configuration_utils import PretrainedConfig

class ModeratoRRRMoeConfig(PretrainedConfig):
    model_type = "moderato_moe"
    keys_to_ignore_at_loading = ["rrr_controller"]

    def __init__(
        self,
        vocab_size=248320,
        hidden_size=5120,
        intermediate_size=17408,
        num_hidden_layers=64,
        num_attention_heads=40,
        num_key_value_heads=8,
        head_dim=128,
        num_experts=6,
        num_experts_per_tok=2,
        rrr_enabled=True,
        rrr_divergence_interval=64,
        rrr_divergence_threshold=0.3,
        rrr_confidence_threshold=0.5,
        max_position_embeddings=131072,
        rms_norm_eps=1e-6,
        rope_theta=1000000.0,
        tie_word_embeddings=False,
        **kwargs,
    ):
        super().__init__(
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.head_dim = head_dim
        self.num_experts = num_experts
        self.num_experts_per_tok = num_experts_per_tok
        self.rrr_enabled = rrr_enabled
        self.rrr_divergence_interval = rrr_divergence_interval
        self.rrr_divergence_threshold = rrr_divergence_threshold
        self.rrr_confidence_threshold = rrr_confidence_threshold
        self.max_position_embeddings = max_position_embeddings
        self.rms_norm_eps = rms_norm_eps
        self.rope_theta = rope_theta