from dataclasses import dataclass, field
from typing import Dict, List, Optional, Union

from ..base import BaseModelConfig


@dataclass
class ModelConfig(BaseModelConfig):
    model_type: str = "deepseek_v4"
    vocab_size: int = 129280
    hidden_size: int = 4096
    intermediate_size: int = 18432
    moe_intermediate_size: int = 2048
    num_hidden_layers: int = 43
    num_attention_heads: int = 64
    num_key_value_heads: int = 1
    n_shared_experts: int = 1
    n_routed_experts: int = 256
    routed_scaling_factor: float = 1.5
    q_lora_rank: int = 1024
    qk_rope_head_dim: int = 64
    num_experts_per_tok: int = 6
    norm_topk_prob: bool = True
    hidden_act: str = "silu"
    max_position_embeddings: int = 1048576
    rms_norm_eps: float = 1e-6
    rope_theta: float = 10000.0
    rope_scaling: Optional[Dict] = None
    attention_bias: bool = False
    attention_dropout: float = 0.0
    head_dim: int = 512
    scoring_func: str = "sqrtsoftplus"
    compress_ratios: List[int] = field(default_factory=list)
    compress_rope_theta: float = 160000.0
    hc_mult: int = 4
    hc_sinkhorn_iters: int = 20
    hc_eps: float = 1e-6
    num_hash_layers: int = 3
    swiglu_limit: float = 10.0
    sliding_window: int = 128
    o_groups: int = 8
    o_lora_rank: int = 1024
    index_n_heads: int = 64
    index_head_dim: int = 128
    index_topk: int = 512
    num_nextn_predict_layers: int = 1
    tie_word_embeddings: bool = False
    bos_token_id: Optional[int] = None
    eos_token_id: Optional[Union[int, List[int]]] = None
    pad_token_id: Optional[int] = None
    topk_method: str = "noaux_tc"

    def __post_init__(self):
        if not self.compress_ratios:
            n = self.num_hidden_layers
            self.compress_ratios = (
                [0]
                + [4 if i % 2 else 128 for i in range(max(n - 2, 0))]
                + ([0] if n >= 2 else [])
            )
        self.compress_ratios = list(self.compress_ratios[: self.num_hidden_layers])
        if len(self.compress_ratios) != self.num_hidden_layers:
            raise ValueError(
                "`compress_ratios` must have one entry per hidden layer, "
                f"got {len(self.compress_ratios)} for {self.num_hidden_layers} layers."
            )
        bad = [r for r in self.compress_ratios if r not in (0, 4, 128)]
        if bad:
            raise ValueError(f"Unsupported DeepSeek-V4 compress ratios: {bad}")