"""
Configuration class for Ivme-Conversate-S-v1.

Mirrors train_ivme_s_v1.ModelConfig exactly (same field names, same defaults),
so config.json produced from a real training run's ModelConfig round-trips
into this class with no field remapping needed.
"""

from transformers import PretrainedConfig


class IvmeConversateSConfig(PretrainedConfig):
    model_type = "ivme_conversate_s"

    def __init__(
        self,
        vocab_size: int = 8000,
        embed_rank: int = 48,
        d_model: int = 256,
        n_unique_layers: int = 14,
        share_factor: int = 2,
        n_heads: int = 8,
        n_kv_heads: int = 2,
        d_ff: int = 512,
        n_meta_tokens: int = 4,
        max_seq_len: int = 768,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.embed_rank = embed_rank
        self.d_model = d_model
        self.n_unique_layers = n_unique_layers
        self.share_factor = share_factor
        self.n_heads = n_heads
        self.n_kv_heads = n_kv_heads
        self.d_ff = d_ff
        self.n_meta_tokens = n_meta_tokens
        self.max_seq_len = max_seq_len
        super().__init__(**kwargs)

    @property
    def n_layers_effective(self):
        return self.n_unique_layers * self.share_factor