"""
GPT configuration (Hugging Face `transformers`-compatible).

This is shipped INSIDE the model repo and loaded via `trust_remote_code=True`:

    from transformers import AutoConfig
    cfg = AutoConfig.from_pretrained("<repo>", trust_remote_code=True)

It mirrors the training architecture: a GPT-style decoder with RMSNorm, ReZero
residual gates, tied token/LM-head weights, rotary position embeddings (RoPE),
and bias-free attention/MLP Linears.
"""

from transformers import PretrainedConfig


class GPTConfig(PretrainedConfig):
    model_type = "gpt"

    # Names transformers utilities sometimes read generically. We also keep the
    # project's own names (n_embd / n_head / n_layer / block_size) as the
    # authoritative ones used by modeling_gpt.py.
    attribute_map = {
        "hidden_size": "n_embd",
        "num_attention_heads": "n_head",
        "num_hidden_layers": "n_layer",
        "max_position_embeddings": "block_size",
    }

    def __init__(
        self,
        vocab_size=28000,
        n_embd=1664,
        n_head=26,
        n_layer=27,
        block_size=32768,
        dropout=0.0,
        tie_weights=True,
        rezero_scale_init=0.01,
        rms_norm_eps=1e-5,
        rope_theta=1_000_000.0,
        # Special token ids (verified from the project's SentencePiece tokenizer).
        # 0 <unk>/<pad>, 1 <s>, 2 </s>, 3 <|im_start|>, 4 <|im_end|>
        im_start_id=3,
        im_end_id=4,
        unk_pad_id=0,
        bos_token_id=1,
        eos_token_id=2,
        pad_token_id=0,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.n_embd = n_embd
        self.n_head = n_head
        self.n_layer = n_layer
        self.block_size = block_size
        self.dropout = dropout
        self.tie_weights = tie_weights
        self.rezero_scale_init = rezero_scale_init
        self.rms_norm_eps = rms_norm_eps
        self.rope_theta = rope_theta
        self.im_start_id = im_start_id
        self.im_end_id = im_end_id
        self.unk_pad_id = unk_pad_id

        # transformers ties via this flag + get_input/output_embeddings().
        kwargs.setdefault("tie_word_embeddings", tie_weights)

        super().__init__(
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            pad_token_id=pad_token_id,
            **kwargs,
        )