model.py

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


def precompute_rope_(head_dim: int, max_seq_length: int, base=10000):
    # Calculate theta_i
    inv_freq = 1/(base ** (torch.arange(0, head_dim, 2).float() / head_dim)) # [max_len/2]
    # create m
    t = torch.arange(max_seq_length, dtype=torch.float32) # [max_len]
    # m * theta_i
    freqs = torch.einsum("i,j->ij", t, inv_freq)
    # repeat
    emb = freqs.repeat_interleave(2, dim=-1)

    return emb

def apply_rope_(x, rope_emb):
    # Align shapes
    seq_len = x.shape[1]
    rope_emb_sliced = rope_emb[:seq_len, :]
    # Reshape for broadcasting
    # Shape (1, seq_len, 1, head_dim)
    emb = rope_emb_sliced.unsqueeze(0).unsqueeze(2)

    cos = emb.cos()
    sin = emb.sin()

    # Create partnet
    x_reshaped = x.float().reshape(*x.shape[:-1], -1, 2)
    x_partner = torch.stack([-x_reshaped[..., 1], x_reshaped[...,0]], dim=-1)
    x_partnet = x_partner.flatten(-2)

    return (x*cos + x_partner*sin).type_as(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 = 4,
        n_ctx: int = 256,
        n_inner: Optional[int] = None,
        dropout: float = 0.1,
        layer_norm_epsilon: float = 1e-5,
        tie_weights: bool = False,
        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 LlamaRotaryEmbedding(nn.Module):
    # inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: ChessConfig, device=None):
        super().__init__()
        self.max_seq_len_cached = config.n_ctx
        self.original_max_seq_len = config.n_ctx

        self.config = config

        rope_init_fn = self.compute_default_rope_parameters
        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)

        self.register_buffer("inv_freq", inv_freq, persistent=False)

    @staticmethod
    def compute_default_rope_parameters(
        config: ChessConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = 10_000.0
        dim = config.n_embd // config.n_head

        attention_factor = 1.0  # Unused in this type of RoPE

        # Compute the inverse frequencies
        inv_freq = 1.0 / (
            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device).float() / dim)
        )
        return inv_freq, attention_factor

    @torch.no_grad()
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos() * self.attention_scaling
        sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


def rotate_half(x):
    """Rotates 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, k, cos, sin, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    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 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
    

        # Using QK-norm
        self.q_norm = nn.RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.k_norm = nn.RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
        # 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,
        position_embeds = 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)
        
        q = self.q_norm(q)
        k = self.k_norm(k)


        # 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)
        
        cos, sin = position_embeds
        #  Positional Embedding
        q, k = apply_rotary_pos_emb(q, k, cos, sin)

        # 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 = nn.RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.attn = MultiHeadAttention(config)
        self.ln_2 = nn.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,
        position_embeds = None,
    ) -> torch.Tensor:
        pass
        # Attention -> norm -> residual
        x = self.ln_1(self.attn(x, attention_mask=attention_mask, position_embeds=position_embeds)) + x
        # Feed-forward -> norm -> residual
        x = self.ln_2(self.mlp(x)) + x
        return x



class ChessForCausalLM(PreTrainedModel):
    """
    Chess Transformer for Causal Language Modeling (next-move prediction).
    
    This model is designed to predict the next chess move given a sequence
    of previous moves. It uses a GPT-style architecture with:
    - Token embeddings for chess moves
    - Learned positional embeddings
    - Stacked transformer blocks
    - Linear head for next-token prediction
    
    The model supports weight tying between the embedding layer and the
    output projection to save parameters.
    
    Example:
        >>> config = ChessConfig(vocab_size=1200, n_embd=128, n_layer=6)
        >>> model = ChessForCausalLM(config)
        >>> inputs = {"input_ids": torch.tensor([[1, 42, 87]])}
        >>> outputs = model(**inputs)
        >>> next_move_logits = outputs.logits[:, -1, :]
    """
    
    config_class = ChessConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    # Suppress missing-key warning for tied lm_head when loading
    keys_to_ignore_on_load_missing = ["lm_head.weight"]
    
    def __init__(self, config: ChessConfig):
        super().__init__(config)
        
        # Token and position embeddings
        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        # self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
        self.wpe = LlamaRotaryEmbedding(config)
        self.drop = nn.Dropout(config.dropout)
        
        # Transformer blocks
        self.h = nn.ModuleList([
            TransformerBlock(config) for _ in range(config.n_layer)
        ])
        
        # Final layer norm
        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        
        # Output head
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        
        # Declare tied weights for proper serialization
        if config.tie_weights:
            self._tied_weights_keys = ["lm_head.weight"]
        
        # 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):
        # Use HF helper to tie or clone depending on config
        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 following GPT-2 style."""
        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]:
        """
        Forward pass of the model.
        
        Args:
            input_ids: Token IDs of shape (batch_size, seq_len).
            attention_mask: Attention mask of shape (batch_size, seq_len).
            position_ids: Position IDs of shape (batch_size, seq_len).
            labels: Labels for language modeling loss.
            return_dict: Whether to return a ModelOutput object.
        
        Returns:
            CausalLMOutputWithPast containing loss (if labels provided) and logits.
        """
        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
        
        # Create position IDs if not provided
        if position_ids is None:
            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
        
        # Get embeddings
        token_embeds = self.wte(input_ids)
        position_embeds = self.wpe(token_embeds, position_ids)
        hidden_states = self.drop(token_embeds)
        
        # Pass through transformer blocks
        for block in self.h:
            hidden_states = block(hidden_states, attention_mask=attention_mask, position_embeds=position_embeds)
        
        # Final layer norm
        hidden_states = self.ln_f(hidden_states)
        
        # Get logits
        logits = self.lm_head(hidden_states)
        
        # Compute loss if labels are provided
        loss = None
        if labels is not None:
            # Shift logits and labels for next-token prediction
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            
            # Flatten for cross-entropy
            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
            # loss_fct = nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
            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 given a sequence of moves.
        
        Args:
            input_ids: Token IDs of shape (1, seq_len).
            temperature: Sampling temperature (1.0 = no change).
            top_k: If set, only sample from top k tokens.
            top_p: If set, use nucleus sampling with this threshold.
        
        Returns:
            The token ID of the predicted next move.
        """
        self.eval()
        
        # Get logits for the last position
        outputs = self(input_ids)
        logits = outputs.logits[:, -1, :] / temperature
        
        # Apply top-k filtering
        if top_k is not None:
            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
            logits[indices_to_remove] = float("-inf")
        
        # Apply top-p (nucleus) filtering
        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)
            
            # Remove tokens with cumulative probability above the threshold
            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")
        
        # Sample from the distribution
        probs = F.softmax(logits, dim=-1)
        next_token = torch.multinomial(probs, num_samples=1)
        
        return next_token.item()


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

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