model.py

"""
Modèle Transformer pour le Chess Challenge (1M paramètres).

Ce module fournit une architecture transformer de style GPT conçue
pour respecter la contrainte stricte de moins de 1 million de paramètres.

Composants clés :
- ChessConfig : Configuration des hyperparamètres.
- ChessForCausalLM : Le modèle principal pour la prédiction du prochain coup.
"""
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, AutoConfig, AutoModelForCausalLM
from transformers.modeling_outputs import CausalLMOutputWithPast

# -----------------------------------------------------------------------------
# Configuration
# -----------------------------------------------------------------------------

class ChessConfig(PretrainedConfig):
    """
    Classe de configuration pour le modèle Chess Transformer.
    
    Conçue pour un budget de paramètres très serré (< 1M).
    
    Répartition du budget (avec les valeurs par défaut de ton ami) :
    - Vocabulaire (Embeddings) : 72 * 92 = ~6,6k
    - Embeddings de position : 256 * 92 = ~23,5k
    - Couches Transformer : 11 couches * (~85k par couche) = ~935k
    - Tête LM (liée aux embeddings) : 0 paramètre supplémentaire
    - Total : ~970k paramètres (juste en dessous de 1M).
    """
    
    model_type = "chess_transformer"
    
    def __init__(
        self,
        vocab_size: int = 1200,
        n_embd: int = 128,
        n_layer: int = 6,
        n_head: int = 4,
        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
        self.dropout = dropout
        self.layer_norm_epsilon = layer_norm_epsilon
        self.tie_weights = tie_weights
        self.tie_word_embeddings = bool(tie_weights)


# -----------------------------------------------------------------------------
# Modules du Transformer
# -----------------------------------------------------------------------------

class MultiHeadAttention(nn.Module):
    """
    Module d'attention multi-têtes standard.
    Inclut le masquage causal pour empêcher le modèle de "voir le futur".
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        assert config.n_embd % config.n_head == 0, \
            f"n_embd ({config.n_embd}) doit être divisible par 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
        
        # Projection combinée Q, K, V pour l'efficacité
        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)
        
        # Masque causal (registre persistent=False pour ne pas le sauvegarder dans le checkpoint)
        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()
        
        # Calcul de Q, K, V
        qkv = self.c_attn(x)
        q, k, v = qkv.split(self.n_embd, dim=2)
        
        # Remodelage pour l'attention multi-têtes
        # (batch, seq_len, n_head, head_dim) -> (batch, n_head, seq_len, head_dim)
        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)
        
        # Attention produit scalaire (Scaled Dot-Product Attention)
        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        
        # Application du masque causal (le futur est masqué avec -inf)
        causal_mask = self.bias[:, :, :seq_len, :seq_len]
        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
        
        # Application du masque d'attention (pour le 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)
        
        # Application de l'attention aux valeurs
        attn_output = torch.matmul(attn_weights, v)
        
        # Remise en forme
        attn_output = attn_output.transpose(1, 2).contiguous().view(
            batch_size, seq_len, self.n_embd
        )
        
        # Projection de sortie
        attn_output = self.c_proj(attn_output)
        
        return attn_output


class FeedForward(nn.Module):
    """
    Réseau de neurones Feed-Forward (MLP).
    Deux couches linéaires avec une activation GELU entre les deux.
    """
    
    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):
    """
    Un bloc transformer unique contenant Attention et Feed-Forward.
    Utilise la "Pre-normalization" (LayerNorm avant l'attention/FFN) pour la stabilité.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.attn = MultiHeadAttention(config)
        self.ln_2 = nn.LayerNorm(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:
        # Connexion résiduelle + Pre-norm Attention
        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
        # Connexion résiduelle + Pre-norm FFN
        x = x + self.mlp(self.ln_2(x))
        return x


# -----------------------------------------------------------------------------
# Modèle Principal
# -----------------------------------------------------------------------------

class ChessForCausalLM(PreTrainedModel):
    """
    Modèle final pour la prédiction de coups (Causal Language Modeling).
    
    Architecture :
    1. Embeddings (Tokens + Position)
    2. Empilement de blocs Transformer
    3. Tête linéaire finale (Projection vers le vocabulaire)
    """
    
    config_class = ChessConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    
    # Ignore l'avertissement de clé manquante car lm_head partage les poids avec wte
    keys_to_ignore_on_load_missing = ["lm_head.weight"]
    
    def __init__(self, config: ChessConfig):
        super().__init__(config)
        
        # Embeddings
        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
        
        self.drop = nn.Dropout(config.dropout)
        
        # Blocs Transformer
        self.h = nn.ModuleList([
            TransformerBlock(config) for _ in range(config.n_layer)
        ])
        
        # LayerNorm final
        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        
        # Tête de sortie (sans biais pour économiser des paramètres)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        
        # Gestion du partage de poids (Weight Tying)
        if config.tie_weights:
            self._tied_weights_keys = ["lm_head.weight"]
        
        # Initialisation des poids
        self.post_init()
        
        # Forcer le lien des poids si configuré
        if config.tie_weights:
            self.tie_weights()

    def get_input_embeddings(self) -> nn.Module:
        return self.wte

    def set_input_embeddings(self, new_embeddings: nn.Module):
        self.wte = new_embeddings
        if getattr(self.config, "tie_weights", False):
            self.tie_weights()

    def get_output_embeddings(self) -> nn.Module:
        return self.lm_head

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

    def tie_weights(self):
        """Lie les poids de l'embedding d'entrée et de la tête de sortie."""
        if getattr(self.config, "tie_weights", False) or getattr(self.config, "tie_word_embeddings", False):
            self._tie_or_clone_weights(self.lm_head, self.wte)

    def _init_weights(self, module: nn.Module):
        """Initialisation des poids style GPT-2."""
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
        elif isinstance(module, nn.LayerNorm):
            torch.nn.init.ones_(module.weight)
            torch.nn.init.zeros_(module.bias)

    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        """
        Passe avant (Forward pass).
        Calcule les logits et, si des étiquettes (labels) sont fournies, la perte (loss).
        """
        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
        
        # Création des IDs de position s'ils ne sont pas fournis
        if position_ids is None:
            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
        
        # Calcul des embeddings (Token + Position)
        token_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        hidden_states = self.drop(token_embeds + position_embeds)
        
        # Passage dans les blocs Transformer
        for block in self.h:
            hidden_states = block(hidden_states, attention_mask=attention_mask)
        
        # Normalisation finale
        hidden_states = self.ln_f(hidden_states)
        
        # Calcul des logits (prédiction du prochain token)
        logits = self.lm_head(hidden_states)
        
        # Calcul de la perte (Training Loss)
        loss = None
        if labels is not None:
            # On décale les logits et les labels d'un cran
            # (le modèle doit prédire le token t+1 à partir du token t)
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            
            # Perte CrossEntropy standard
            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
            loss = loss_fct(
                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,
        )

    @torch.no_grad()
    def generate_move(
        self,
        input_ids: torch.LongTensor,
        temperature: float = 1.0,
        top_k: Optional[int] = None,
        top_p: Optional[float] = None,
    ) -> int:
        """
        Génère le prochain coup pour une séquence donnée.
        Utilisé pour l'inférence en jeu réel.
        """
        self.eval()
        
        # Récupère les logits pour la dernière position uniquement
        outputs = self(input_ids)
        logits = outputs.logits[:, -1, :] / temperature
        
        # Filtrage Top-K
        if top_k is not None:
            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
            logits[indices_to_remove] = float("-inf")
        
        # Filtrage Top-P (Nucleus Sampling)
        if top_p is not None:
            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
            cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
            
            # On retire les tokens qui sont au-dessus du seuil cumulatif
            sorted_indices_to_remove = cumulative_probs > top_p
            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
            sorted_indices_to_remove[..., 0] = 0
            
            indices_to_remove = sorted_indices_to_remove.scatter(
                dim=-1, index=sorted_indices, src=sorted_indices_to_remove
            )
            logits[indices_to_remove] = float("-inf")
        
        # Échantillonnage final
        probs = F.softmax(logits, dim=-1)
        next_token = torch.multinomial(probs, num_samples=1)
        
        return next_token.item()

# Enregistrement pour chargement automatique via AutoModel
AutoConfig.register("chess_transformer", ChessConfig)
AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)