"""
Chess Transformer Model for the Chess Challenge.

This module provides a simple GPT-style transformer architecture
designed to fit within the 1M parameter constraint.

Key components:
- ChessConfig: Configuration class for model hyperparameters
- ChessForCausalLM: The main model class for next-move prediction
"""

from __future__ import annotations

import math
from dataclasses import dataclass
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast


class RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def forward(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight

class RotaryEmbedding(nn.Module):
    def __init__(self, dim, max_seq_len=256):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, x, seq_len):
        t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)
        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        return emb[None, :, None, :]

def apply_rotary_emb(q, k, freqs):
    def rotate_half(x):
        x1, x2 = x.chunk(2, dim=-1)
        return torch.cat((-x2, x1), dim=-1)
    
    q_rot = (q * freqs.cos()) + (rotate_half(q) * freqs.sin())
    k_rot = (k * freqs.cos()) + (rotate_half(k) * freqs.sin())
    return q_rot, k_rot

class SwiGLU(nn.Module):
    def __init__(self, dim: int, inner_dim: int, dropout: float):
        super().__init__()
        self.w1 = nn.Linear(dim, inner_dim, bias=False)
        self.w2 = nn.Linear(inner_dim, dim, bias=False)
        self.w3 = nn.Linear(dim, inner_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        # L'essence de SwiGLU : (SiLU(W1x) * W3x) * W2
        return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x)))

class ModernAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.n_head = config.n_head
        self.head_dim = config.n_embd // config.n_head
        
        self.wq = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.wk = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.wv = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.wo = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x, freqs, mask=None):
        bsz, seqlen, _ = x.shape
        q, k, v = self.wq(x), self.wk(x), self.wv(x)

        q = q.view(bsz, seqlen, self.n_head, self.head_dim)
        k = k.view(bsz, seqlen, self.n_head, self.head_dim)
        v = v.view(bsz, seqlen, self.n_head, self.head_dim)

        q, k = apply_rotary_emb(q, k, freqs)

        scores = torch.matmul(q.transpose(1, 2), k.transpose(1, 2).transpose(-2, -1)) / math.sqrt(self.head_dim)
        
        if mask is not None:
            scores = scores + mask[:, :, :seqlen, :seqlen]

        scores = F.softmax(scores.float(), dim=-1).type_as(q)
        output = torch.matmul(scores, v.transpose(1, 2))
        output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
        return self.dropout(self.wo(output))

class ModernBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.attention = ModernAttention(config)
        self.feed_forward = SwiGLU(config.n_embd, config.n_inner, config.dropout)
        self.attention_norm = RMSNorm(config.n_embd)
        self.ffn_norm = RMSNorm(config.n_embd)

    def forward(self, x, freqs, mask):
        x = x + self.attention(self.attention_norm(x), freqs, mask)
        x = x + self.feed_forward(self.ffn_norm(x))
        return x





class ChessConfig(PretrainedConfig):
    """
    Configuration class for the Chess Transformer model.
    
    This configuration is designed for a ~1M parameter model.
    Students can adjust these values to explore different architectures.
    
    Parameter budget breakdown (with default values):
    - Embeddings (vocab): 1200 x 128 = 153,600
    - Position Embeddings: 256 x 128 = 32,768
    - Transformer Layers: 6 x ~120,000 = ~720,000
    - LM Head (with weight tying): 0 (shared with embeddings)
    - Total: ~906,000 parameters
    
    Attributes:
        vocab_size: Size of the vocabulary (number of unique moves).
        n_embd: Embedding dimension (d_model).
        n_layer: Number of transformer layers.
        n_head: Number of attention heads.
        n_ctx: Maximum sequence length (context window).
        n_inner: Feed-forward inner dimension (default: 3 * n_embd).
        dropout: Dropout probability.
        layer_norm_epsilon: Epsilon for layer normalization.
        tie_weights: Whether to tie embedding and output weights.
    """
    
    model_type = "chess_transformer"
    
    def __init__(
        self,
        vocab_size: int = 1200,
        n_embd: int = 128,
        n_layer: int = 6,
        n_head: int = 8,
        n_ctx: int = 256,
        n_inner: Optional[int] = None,
        dropout: float = 0.1,
        layer_norm_epsilon: float = 1e-5,
        tie_weights: bool = True,
        pad_token_id: int = 0,
        bos_token_id: int = 1,
        eos_token_id: int = 2,
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            **kwargs,
        )
        
        self.vocab_size = vocab_size
        self.n_embd = n_embd
        self.n_layer = n_layer
        self.n_head = n_head
        self.n_ctx = n_ctx
        self.n_inner = n_inner if n_inner is not None else 3 * n_embd  # Reduced from 4x to 3x
        self.dropout = dropout
        self.layer_norm_epsilon = layer_norm_epsilon
        self.tie_weights = tie_weights
        # Inform HF base class about tying behavior
        self.tie_word_embeddings = bool(tie_weights)


class MultiHeadAttention(nn.Module):
    """
    Multi-head self-attention module.
    
    This is a standard scaled dot-product attention implementation
    with causal masking for autoregressive generation.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        assert config.n_embd % config.n_head == 0, \
            f"n_embd ({config.n_embd}) must be divisible by n_head ({config.n_head})"
        
        self.n_head = config.n_head
        self.n_embd = config.n_embd
        self.head_dim = config.n_embd // config.n_head
        
        # Combined QKV projection for efficiency
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
        
        self.dropout = nn.Dropout(config.dropout)
        
        # Causal mask (will be created on first forward pass)
        self.register_buffer(
            "bias",
            torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(
                1, 1, config.n_ctx, config.n_ctx
            ),
            persistent=False,
        )
    
    def forward(
        self,
        x: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        batch_size, seq_len, _ = x.size()
        
        # Compute Q, K, V
        qkv = self.c_attn(x)
        q, k, v = qkv.split(self.n_embd, dim=2)
        
        # Reshape for multi-head attention
        q = q.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
        k = k.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
        v = v.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
        
        # Scaled dot-product attention
        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        
        # Apply causal mask
        causal_mask = self.bias[:, :, :seq_len, :seq_len]
        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
        
        # Apply attention mask (for padding)
        if attention_mask is not None:
            # attention_mask shape: (batch_size, seq_len) -> (batch_size, 1, 1, seq_len)
            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
            attn_weights = attn_weights.masked_fill(attention_mask == 0, float("-inf"))
        
        attn_weights = F.softmax(attn_weights, dim=-1)
        attn_weights = self.dropout(attn_weights)
        
        # Apply attention to values
        attn_output = torch.matmul(attn_weights, v)
        
        # Reshape back
        attn_output = attn_output.transpose(1, 2).contiguous().view(
            batch_size, seq_len, self.n_embd
        )
        
        # Output projection
        attn_output = self.c_proj(attn_output)
        
        return attn_output


class FeedForward(nn.Module):
    """
    Feed-forward network (MLP) module.
    
    Standard two-layer MLP with GELU activation.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        self.c_fc = nn.Linear(config.n_embd, config.n_inner)
        self.c_proj = nn.Linear(config.n_inner, config.n_embd)
        self.dropout = nn.Dropout(config.dropout)
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.c_fc(x)
        x = F.gelu(x)
        x = self.c_proj(x)
        x = self.dropout(x)
        return x


class TransformerBlock(nn.Module):
    """
    A single transformer block with attention and feed-forward layers.
    
    Uses pre-normalization (LayerNorm before attention/FFN) for better
    training stability.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        self.ln_1 = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.attn = MultiHeadAttention(config)
        self.ln_2 = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.mlp = FeedForward(config)
    
    def forward(
        self,
        x: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        # Pre-norm attention
        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
        # Pre-norm FFN
        x = x + self.mlp(self.ln_2(x))
        return x


class ChessForCausalLM(PreTrainedModel):
    config_class = ChessConfig
    _tied_weights_keys = ["lm_head.weight"]
    
    def __init__(self, config: ChessConfig):
        super().__init__(config)
        
        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        
        self.rope = RotaryEmbedding(config.n_embd // config.n_head)
        
        self.drop = nn.Dropout(config.dropout)
        self.h = nn.ModuleList([ModernBlock(config) for _ in range(config.n_layer)])
        self.ln_f = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

        self.post_init()
        if config.tie_weights:
            self.tie_weights()
            

    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        batch_size, seq_len = input_ids.size()
        device = input_ids.device
        
        freqs = self.rope(input_ids, seq_len)
        
        mask = torch.full((seq_len, seq_len), float("-inf"), device=device)
        mask = torch.triu(mask, diagonal=1)
        mask = mask.view(1, 1, seq_len, seq_len)

        hidden_states = self.drop(self.wte(input_ids))
        
        for block in self.h:
            hidden_states = block(hidden_states, freqs, mask)
        
        hidden_states = self.ln_f(hidden_states)
        logits = self.lm_head(hidden_states)
        
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss = F.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
        
        if not return_dict:
            output = (logits,)
            return ((loss,) + output) if loss is not None else output
            
        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=None,
            hidden_states=None,
            attentions=None,
        )
    
    def get_input_embeddings(self):
        return self.wte

    def set_input_embeddings(self, value):
        self.wte = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def tie_weights(self):
        """
        C'est cette méthode que HF appelle automatiquement si 
        config.tie_word_embeddings est True.
        """
        self._tie_or_clone_weights(self.lm_head, self.wte)


# Register the model with Auto classes for easy loading
from transformers import AutoConfig, AutoModelForCausalLM

AutoConfig.register("chess_transformer", ChessConfig)
AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)