from transformers import PretrainedConfig


class AsterConfig(PretrainedConfig):
    model_type = "aster"

    def __init__(
        self, vocab_size=32000, hidden_size=1152, num_hidden_layers=22,
        num_attention_heads=18, num_key_value_heads=2, intermediate_size=5504,
        max_position_embeddings=4096, rms_norm_eps=1e-6, rope_theta=10000.0,
        gqa_implementation="native", bos_token_id=1, eos_token_id=2,
        pad_token_id=3, tie_word_embeddings=True, **kwargs,
    ):
        super().__init__(
            bos_token_id=bos_token_id, eos_token_id=eos_token_id,
            pad_token_id=pad_token_id, tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.hidden_size = hidden_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.intermediate_size = intermediate_size
        self.max_position_embeddings = max_position_embeddings
        self.rms_norm_eps = rms_norm_eps
        self.rope_theta = rope_theta
        self.gqa_implementation = gqa_implementation
        if hidden_size % num_attention_heads or num_attention_heads % num_key_value_heads:
            raise ValueError("Invalid attention head configuration")
        if hidden_size // num_attention_heads != 64:
            raise ValueError("Expected 64-dimensional attention heads for Stage 1")
        if gqa_implementation not in ("native", "expanded"):
            raise ValueError("gqa_implementation must be native or expanded")