import torch
from transformers import PretrainedConfig, AutoConfig

class MiniSTUConfig(PretrainedConfig):
    model_type = "ministu"

    def __init__(
    self,
    bsz: int = 1,
    dim: int = 896,
    num_heads: int = 8,
    num_layers: int = 12,
    seq_len: int = 8192,
    weight_tying: bool = False,
    window_size: int = 1024,
    vocab_size: int = 200064,
    mlp_scale: int = 12,
    bias: bool = False,
    dropout: float = 0.0,
    num_eigh: int = 24,
    use_hankel_L: bool = False,
    use_flash_fft: bool = True,
    use_approx: bool = True,
    use_attn: bool = True,
    softcap: float = 50.0,
    theta: float = 10_000.0,
    use_alibi: bool = False,
    dilation: int = 2,
    torch_dtype: torch.dtype = torch.bfloat16,
    device: torch.device = None,
    **kwargs,
    ):
        super().__init__(**kwargs)
        self.bsz = bsz
        self.dim = dim
        self.num_heads = num_heads
        self.num_layers = num_layers
        self.seq_len = seq_len
        self.weight_tying = weight_tying
        self.window_size = window_size
        self.vocab_size = vocab_size
        self.hidden_size = dim
        self.mlp_scale = mlp_scale
        self.intermediate_size = self.hidden_size * self.mlp_scale
        self.bias = bias
        self.dropout = dropout
        self.num_eigh = num_eigh
        self.use_hankel_L = use_hankel_L
        self.use_flash_fft = use_flash_fft
        self.use_approx = use_approx
        self.use_attn = use_attn
        self.softcap = softcap
        self.theta = theta
        self.use_alibi = use_alibi
        self.torch_dtype = torch_dtype
        self.device = self.device = device or ('cuda' if torch.cuda.is_available() else 'cpu')  # Store as string