model.py

"""
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 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 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 = 4,
        n_kv_head: Optional[int] = None,  
        n_ctx: int = 256,
        n_inner: Optional[int] = None,
        dropout: float = 0.1,
        rms_norm_epsilon: float = 1e-6,
        tie_weights: bool = True,
        use_rope: bool = True,
        rope_theta: float = 10000.0,
        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_kv_head = n_kv_head if n_kv_head is not None else n_head
        self.n_ctx = n_ctx
        # SwiGLU typically uses 2/3 * 4 * n_embd, rounded to multiple of 64
        self.n_inner = n_inner if n_inner is not None else self._compute_swiglu_dim(n_embd)
        self.dropout = dropout
        self.rms_norm_epsilon = rms_norm_epsilon
        self.tie_weights = tie_weights
        self.tie_word_embeddings = bool(tie_weights)
        self.use_rope = use_rope
        self.rope_theta = rope_theta
        # For compatibility with src/utils.py parameter estimation
        self.layer_norm_epsilon = rms_norm_epsilon
    
    @staticmethod
    def _compute_swiglu_dim(n_embd: int) -> int:
        """Compute SwiGLU hidden dimension (typically 8/3 * n_embd, rounded)."""
        # Standard SwiGLU uses ~2.67x multiplier
        hidden = int(8 * n_embd / 3)
        # Round to multiple of 64 for efficiency (optional)
        return ((hidden + 63) // 64) * 64


class RMSNorm(nn.Module):
    """
    Root Mean Square Layer Normalization.
    
    Simpler and faster than LayerNorm - no mean centering, no bias.
    """
    
    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: torch.Tensor) -> torch.Tensor:
        # RMSNorm: x * weight / sqrt(mean(x^2) + eps)
        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
        return x * norm * self.weight


class RotaryEmbedding(nn.Module):
    """
    Rotary Position Embeddings (RoPE).
    
    Encodes position information directly into attention computation
    without learnable parameters.
    """
    
    def __init__(self, dim: int, max_seq_len: int = 512, theta: float = 10000.0):
        super().__init__()
        self.dim = dim
        self.max_seq_len = max_seq_len
        self.theta = theta
        
        # Precompute frequencies
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        
        # Precompute cos/sin cache
        self._build_cache(max_seq_len)
    
    def _build_cache(self, seq_len: int):
        """Build cos/sin cache for positions."""
        t = torch.arange(seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(t, self.inv_freq)
        # Concatenate to get full dim
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos_cached", emb.cos(), persistent=False)
        self.register_buffer("sin_cached", emb.sin(), persistent=False)
    
    def forward(self, seq_len: int, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
        """Return cos and sin for the given sequence length."""
        if seq_len > self.max_seq_len:
            self._build_cache(seq_len)
            self.max_seq_len = seq_len
        
        return (
            self.cos_cached[:seq_len].to(device),
            self.sin_cached[:seq_len].to(device),
        )


def rotate_half(x: torch.Tensor) -> torch.Tensor:
    """Rotate half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(
    q: torch.Tensor,
    k: torch.Tensor,
    cos: torch.Tensor,
    sin: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """Apply rotary position embeddings to query and key tensors."""
    # q, k: (batch, n_head, seq_len, head_dim)
    # cos, sin: (seq_len, head_dim)
    cos = cos.unsqueeze(0).unsqueeze(0)  # (1, 1, seq_len, head_dim)
    sin = sin.unsqueeze(0).unsqueeze(0)
    
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    
    return q_embed, k_embed


class MultiHeadAttention(nn.Module):
    """
    Multi-head self-attention with RoPE.
    
    Supports Grouped Query Attention (GQA) when n_kv_head < n_head.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        assert config.n_embd % config.n_head == 0
        
        self.n_head = config.n_head
        self.n_kv_head = config.n_kv_head
        self.n_embd = config.n_embd
        self.head_dim = config.n_embd // config.n_head
        self.n_rep = config.n_head // config.n_kv_head  # For GQA
        
        # Separate Q, K, V projections for clarity with GQA
        self.q_proj = nn.Linear(config.n_embd, config.n_head * self.head_dim, bias=False)
        self.k_proj = nn.Linear(config.n_embd, config.n_kv_head * self.head_dim, bias=False)
        self.v_proj = nn.Linear(config.n_embd, config.n_kv_head * self.head_dim, bias=False)
        self.o_proj = nn.Linear(config.n_head * self.head_dim, config.n_embd, bias=False)
        
        self.dropout = nn.Dropout(config.dropout)
        
        # RoPE
        self.rotary_emb = RotaryEmbedding(
            self.head_dim,
            max_seq_len=config.n_ctx,
            theta=config.rope_theta,
        )
        
        # Causal mask
        self.register_buffer(
            "causal_mask",
            torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(
                1, 1, config.n_ctx, config.n_ctx
            ),
            persistent=False,
        )
    
    def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
        """Repeat KV heads for GQA."""
        if self.n_rep == 1:
            return x
        batch, n_kv_head, seq_len, head_dim = x.shape
        x = x[:, :, None, :, :].expand(batch, n_kv_head, self.n_rep, seq_len, head_dim)
        return x.reshape(batch, n_kv_head * self.n_rep, seq_len, head_dim)
    
    def forward(
        self,
        x: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        batch_size, seq_len, _ = x.size()
        
        # Project Q, K, V
        q = self.q_proj(x)
        k = self.k_proj(x)
        v = self.v_proj(x)
        
        # Reshape for 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_kv_head, self.head_dim).transpose(1, 2)
        v = v.view(batch_size, seq_len, self.n_kv_head, self.head_dim).transpose(1, 2)
        
        # Apply RoPE
        cos, sin = self.rotary_emb(seq_len, x.device)
        q, k = apply_rotary_pos_emb(q, k, cos, sin)
        
        # Repeat KV for GQA
        k = self._repeat_kv(k)
        v = self._repeat_kv(v)
        
        # Scaled dot-product attention
        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        
        # Apply causal mask
        causal_mask = self.causal_mask[:, :, :seq_len, :seq_len]
        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
        
        # Apply padding mask
        if attention_mask is not None:
            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 and project output
        attn_output = attn_output.transpose(1, 2).contiguous().view(
            batch_size, seq_len, self.n_embd
        )
        attn_output = self.o_proj(attn_output)
        
        return attn_output


class SwiGLU(nn.Module):
    """
    SwiGLU Feed-Forward Network.
    
    SwiGLU(x) = (xW1 * SiLU(xW_gate)) @ W2
    
    More expressive than standard FFN with similar parameter count.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        hidden_dim = config.n_inner
        
        # Gate and up projections (can be fused for efficiency)
        self.gate_proj = nn.Linear(config.n_embd, hidden_dim, bias=False)
        self.up_proj = nn.Linear(config.n_embd, hidden_dim, bias=False)
        self.down_proj = nn.Linear(hidden_dim, config.n_embd, bias=False)
        self.dropout = nn.Dropout(config.dropout)
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # SwiGLU: SiLU(gate) * up, then down
        gate = F.silu(self.gate_proj(x))
        up = self.up_proj(x)
        x = gate * up
        x = self.down_proj(x)
        x = self.dropout(x)
        return x


class TransformerBlock(nn.Module):
    """
    Transformer block with RMSNorm, RoPE attention, and SwiGLU FFN.
    
    Uses pre-normalization for training stability.
    """
    
    def __init__(self, config: ChessConfig):
        super().__init__()
        
        self.ln_1 = RMSNorm(config.n_embd, eps=config.rms_norm_epsilon)
        self.attn = MultiHeadAttention(config)
        self.ln_2 = RMSNorm(config.n_embd, eps=config.rms_norm_epsilon)
        self.mlp = SwiGLU(config)
    
    def forward(
        self,
        x: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        # Pre-norm attention with residual
        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
        # Pre-norm FFN with residual
        x = x + self.mlp(self.ln_2(x))
        return x


class ChessForCausalLM(PreTrainedModel):
    """
    Chess Transformer for Causal Language Modeling.
    
    Modern architecture with RoPE, SwiGLU, and RMSNorm.
    """
    
    config_class = ChessConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    _tied_weights_keys = ["lm_head.weight"]
    keys_to_ignore_on_load_missing = ["lm_head.weight"]
    
    def __init__(self, config: ChessConfig):
        super().__init__(config)
        
        # Token embeddings (no position embeddings - using RoPE)
        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        
        self.drop = nn.Dropout(config.dropout)
        
        # Transformer blocks
        self.h = nn.ModuleList([
            TransformerBlock(config) for _ in range(config.n_layer)
        ])
        
        # Final RMSNorm
        self.ln_f = RMSNorm(config.n_embd, eps=config.rms_norm_epsilon)
        
        # Output head
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        
        # Initialize weights
        self.post_init()
        
        # Tie weights if configured
        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):
        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):
        """Initialize weights."""
        std = 0.02
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=std)
            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=std)
        elif isinstance(module, RMSNorm):
            torch.nn.init.ones_(module.weight)
    
    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
        
        # Get token embeddings (no position embeddings - RoPE handles position)
        hidden_states = self.wte(input_ids)
        hidden_states = self.drop(hidden_states)
        
        # Pass through transformer blocks
        for block in self.h:
            hidden_states = block(hidden_states, attention_mask=attention_mask)
        
        # Final norm and head
        hidden_states = self.ln_f(hidden_states)
        logits = self.lm_head(hidden_states)
        
        # Compute loss if labels provided
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            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:
        """Generate the next move token."""
        self.eval()
        
        outputs = self(input_ids)
        logits = outputs.logits[:, -1, :] / temperature
        
        if top_k is not None:
            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
            logits[indices_to_remove] = float("-inf")
        
        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)
            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")
        
        probs = F.softmax(logits, dim=-1)
        next_token = torch.multinomial(probs, num_samples=1)
        
        return next_token.item()