from transformers import PretrainedConfig


class RSLMConfig(PretrainedConfig):
    model_type = "rslm"

    def __init__(
        self,
        hidden_size=2048,
        num_layers=24,
        num_q_heads=16,
        num_kv_heads=1,
        head_dim=128,
        intermediate_size=4352,
        vocab_size=65536,
        max_position_embeddings=262144,
        original_max_position_embeddings=8192,
        rope_theta=1000000.0,
        rope_scaling=None,
        window_size=4096,
        global_layers_0idx=(5, 11, 17, 23),
        evict_local_kv=True,
        local_cache_keep=4096,
        parallel_block=True,
        rms_norm_eps=1e-6,
        hidden_act="swiglu",
        tie_word_embeddings=True,
        bos_token_id=1,
        eos_token_id=2,
        pad_token_id=0,
        **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.hidden_size = hidden_size
        self.num_layers = num_layers
        self.num_q_heads = num_q_heads
        self.num_kv_heads = num_kv_heads
        self.head_dim = head_dim
        self.intermediate_size = intermediate_size
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.original_max_position_embeddings = original_max_position_embeddings
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.window_size = window_size
        self.global_layers_0idx = list(global_layers_0idx)
        self.evict_local_kv = evict_local_kv
        self.local_cache_keep = local_cache_keep
        self.parallel_block = parallel_block
        self.rms_norm_eps = rms_norm_eps
        self.hidden_act = hidden_act