"""Configuration for the Q-50M architecture (replica of q-project/Q-50M-Base).

512-wide, 10-layer decoder with GQA (8 q / 2 kv heads), per-head QK-Norm,
content-dependent scalar residual gates on both branches, tied embeddings.
"""

from transformers import PretrainedConfig


class Q50MConfig(PretrainedConfig):
    model_type = "q50m"

    def __init__(
        self,
        vocab_size: int = 32768,
        hidden_size: int = 512,
        intermediate_size: int = 1792,
        num_hidden_layers: int = 10,
        num_attention_heads: int = 8,
        num_key_value_heads: int = 2,
        head_dim: int = 64,
        max_position_embeddings: int = 2048,
        rope_theta: float = 10000.0,
        rms_norm_eps: float = 1e-5,
        qk_norm: bool = True,
        attention_scalar_gate: bool = True,
        mlp_scalar_gate: bool = True,
        gate_multiplier: float = 2.0,
        nope_every_n: int | None = None,
        nope_layers: list[int] | None = None,
        attention_dropout: float = 0.0,
        tie_word_embeddings: bool = True,
        bos_token_id: int = 1,
        eos_token_id: int = 2,
        pad_token_id: int = 0,
        **kwargs,
    ):
        super().__init__(
            tie_word_embeddings=tie_word_embeddings,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            pad_token_id=pad_token_id,
            **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.max_position_embeddings = max_position_embeddings
        self.rope_theta = rope_theta
        self.rms_norm_eps = rms_norm_eps
        self.qk_norm = qk_norm
        self.attention_scalar_gate = attention_scalar_gate
        self.mlp_scalar_gate = mlp_scalar_gate
        self.gate_multiplier = gate_multiplier
        self.nope_every_n = nope_every_n
        if nope_layers is not None:
            self.nope_layers = list(nope_layers)
        elif nope_every_n:
            self.nope_layers = [
                i for i in range(num_hidden_layers) if (i + 1) % nope_every_n == 0
            ]
        else:
            self.nope_layers = []
        self.attention_dropout = attention_dropout