from transformers import PretrainedConfig


class TheoBertBaseConfig(PretrainedConfig):
    model_type = "theo_bert_base"

    def __init__(
        self,
        vocab_size=30522,
        n_layer=12,
        n_head=8,
        n_embd=768,
        seq_len=256,
        rope_base=10000,
        rope_cache_factor=10,
        pad_token_id=0,
        unk_token_id=100,
        cls_token_id=101,
        sep_token_id=102,
        mask_token_id=103,
        tokenizer_name="google-bert/bert-base-uncased",
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.n_layer = n_layer
        self.n_head = n_head
        self.n_embd = n_embd
        self.seq_len = seq_len
        self.rope_base = rope_base
        self.rope_cache_factor = rope_cache_factor
        self.tokenizer_name = tokenizer_name

        # Compatibility aliases expected by downstream tooling.
        self.hidden_size = n_embd
        self.num_hidden_layers = n_layer
        self.num_attention_heads = n_head
        self.max_position_embeddings = seq_len
        self.rope_cache_seq_len = seq_len * rope_cache_factor

        super().__init__(
            pad_token_id=pad_token_id,
            unk_token_id=unk_token_id,
            cls_token_id=cls_token_id,
            sep_token_id=sep_token_id,
            mask_token_id=mask_token_id,
            **kwargs,
        )