Upload model.py

model.py

@@ -0,0 +1,817 @@
+"""
+Chess Transformer Model for the Chess Challenge.
+
+This module provides a modular 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
+
+Modular options:
+- Attention: MHA (standard), GQA (grouped query), MQA (multi-query)
+- Position encoding: learned, rope (rotary), alibi
+- FFN activation: gelu, swiglu
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union, Literal
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PretrainedConfig, PreTrainedModel
+try:
+    from transformers.generation.utils import GenerationMixin
+except ImportError:  # Fallback for older transformers
+    from transformers import GenerationMixin
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+# Type aliases for configuration options
+AttentionType = Literal["mha", "gqa", "mqa"]
+PositionEncoding = Literal["learned", "rope", "alibi"]
+FFNType = Literal["gelu", "swiglu"]
+
+
+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 (0 with rope/alibi)
+    - 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_kv_heads: Number of key-value heads (for GQA/MQA). None = same as n_head.
+        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.
+        attention_type: Type of attention mechanism ("mha", "gqa", "mqa").
+        pos_encoding: Type of position encoding ("learned", "rope", "alibi").
+        ffn_type: Type of FFN activation ("gelu", "swiglu").
+        rope_theta: Base frequency for RoPE (default 10000.0).
+        legal_loss_weight: Auxiliary legal-move loss weight (default 0.0).
+    """
+
+    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_heads: Optional[int] = None,
+        n_ctx: int = 256,
+        n_inner: Optional[int] = None,
+        dropout: float = 0.1,
+        layer_norm_epsilon: float = 1e-5,
+        tie_weights: bool = True,
+        # New modular options
+        attention_type: AttentionType = "mha",
+        pos_encoding: PositionEncoding = "learned",
+        ffn_type: FFNType = "gelu",
+        rope_theta: float = 10000.0,
+        legal_loss_weight: float = 0.0,
+        # Token IDs
+        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
+        # Inform HF base class about tying behavior
+        self.tie_word_embeddings = bool(tie_weights)
+
+        # Modular architecture options
+        self.attention_type = attention_type
+        self.pos_encoding = pos_encoding
+        self.ffn_type = ffn_type
+        self.rope_theta = rope_theta
+        self.legal_loss_weight = legal_loss_weight
+
+        # Handle n_kv_heads based on attention type
+        if n_kv_heads is None:
+            if attention_type == "mqa":
+                self.n_kv_heads = 1
+            elif attention_type == "gqa":
+                # Default to n_head // 2 for GQA, but at least 1
+                self.n_kv_heads = max(1, n_head // 2)
+            else:  # mha
+                self.n_kv_heads = n_head
+        else:
+            self.n_kv_heads = n_kv_heads
+
+        # Validation
+        assert n_embd % n_head == 0, f"n_embd ({n_embd}) must be divisible by n_head ({n_head})"
+        assert n_head % self.n_kv_heads == 0, f"n_head ({n_head}) must be divisible by n_kv_heads ({self.n_kv_heads})"
+        assert attention_type in ("mha", "gqa", "mqa"), f"Invalid attention_type: {attention_type}"
+        assert pos_encoding in ("learned", "rope", "alibi"), f"Invalid pos_encoding: {pos_encoding}"
+        assert ffn_type in ("gelu", "swiglu"), f"Invalid ffn_type: {ffn_type}"
+
+
+# ==============================================================================
+# Position Encoding Modules
+# ==============================================================================
+
+
+class RotaryEmbedding(nn.Module):
+    """
+    Rotary Position Embedding (RoPE).
+
+    Applies rotary embeddings to queries and keys, encoding position
+    information through rotation in the complex plane. This allows
+    relative position information without explicit position embeddings.
+
+    Reference: https://arxiv.org/abs/2104.09864
+    """
+
+    def __init__(self, dim: int, max_seq_len: int = 256, theta: float = 10000.0):
+        super().__init__()
+        self.dim = dim
+        self.max_seq_len = max_seq_len
+        self.theta = theta
+
+        # Precompute frequency bands
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Precompute sin/cos for all positions
+        self._build_cache(max_seq_len)
+
+    def _build_cache(self, seq_len: int):
+        """Build sin/cos cache for given sequence length."""
+        positions = torch.arange(seq_len, dtype=torch.float32)
+        freqs = torch.outer(positions, self.inv_freq)
+        # Create [cos, sin] interleaved for rotation
+        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, x: torch.Tensor, seq_len: int) -> 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(x.dtype),
+            self.sin_cached[:seq_len].to(x.dtype),
+        )
+
+
+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 embedding to queries and keys.
+
+    Args:
+        q: Query tensor of shape (batch, n_heads, seq_len, head_dim)
+        k: Key tensor of shape (batch, n_kv_heads, seq_len, head_dim)
+        cos: Cosine of rotation angles
+        sin: Sine of rotation angles
+
+    Returns:
+        Rotated q and k tensors
+    """
+    # cos/sin shape: (seq_len, head_dim) -> (1, 1, seq_len, head_dim)
+    cos = cos.unsqueeze(0).unsqueeze(0)
+    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
+
+
+def build_alibi_slopes(n_heads: int) -> torch.Tensor:
+    """
+    Build ALiBi slopes for attention bias.
+
+    ALiBi adds a linear bias to attention scores based on position distance.
+    The slope decreases geometrically for each head.
+
+    Reference: https://arxiv.org/abs/2108.12409
+    """
+    def get_slopes_power_of_2(n: int) -> list:
+        start = 2 ** (-(2 ** -(math.log2(n) - 3)))
+        ratio = start
+        return [start * (ratio ** i) for i in range(n)]
+
+    if math.log2(n_heads).is_integer():
+        slopes = get_slopes_power_of_2(n_heads)
+    else:
+        # For non-power-of-2, use closest power of 2 and interpolate
+        closest_power_of_2 = 2 ** math.floor(math.log2(n_heads))
+        slopes = get_slopes_power_of_2(closest_power_of_2)
+        extra_slopes = get_slopes_power_of_2(2 * closest_power_of_2)
+        slopes = slopes + extra_slopes[0::2][: n_heads - closest_power_of_2]
+
+    return torch.tensor(slopes, dtype=torch.float32)
+
+
+def build_alibi_bias(seq_len: int, slopes: torch.Tensor) -> torch.Tensor:
+    """
+    Build the ALiBi attention bias matrix.
+
+    Args:
+        seq_len: Sequence length
+        slopes: ALiBi slopes tensor of shape (n_heads,)
+
+    Returns:
+        Bias tensor of shape (1, n_heads, seq_len, seq_len)
+    """
+    # Create distance matrix: distance[i, j] = j - i (negative for causal)
+    positions = torch.arange(seq_len)
+    distance = positions.unsqueeze(0) - positions.unsqueeze(1)  # (seq_len, seq_len)
+
+    # Apply slopes: (n_heads, 1, 1) * (seq_len, seq_len) -> (n_heads, seq_len, seq_len)
+    alibi = slopes.unsqueeze(1).unsqueeze(1) * distance.unsqueeze(0)
+
+    return alibi.unsqueeze(0)  # (1, n_heads, seq_len, seq_len)
+
+
+# ==============================================================================
+# Attention Modules
+# ==============================================================================
+
+
+class Attention(nn.Module):
+    """
+    Unified attention module supporting MHA, GQA, and MQA.
+
+    Supports multiple position encoding methods:
+    - learned: Standard learned position embeddings (handled externally)
+    - rope: Rotary Position Embeddings (applied to Q and K)
+    - alibi: Attention with Linear Biases (added to attention scores)
+
+    Architecture variants:
+    - MHA (Multi-Head Attention): n_kv_heads == n_head
+    - GQA (Grouped Query Attention): n_kv_heads < n_head, n_head % n_kv_heads == 0
+    - MQA (Multi-Query Attention): n_kv_heads == 1
+    """
+
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+
+        self.n_head = config.n_head
+        self.n_kv_heads = config.n_kv_heads
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        self.n_rep = config.n_head // config.n_kv_heads  # Repetition factor for GQA/MQA
+        self.pos_encoding = config.pos_encoding
+
+        # Compute projection sizes
+        # Q: n_head * head_dim = n_embd
+        # K, V: n_kv_heads * head_dim (smaller for GQA/MQA)
+        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_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.n_embd, config.n_kv_heads * self.head_dim, bias=False)
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=False)
+
+        self.dropout = nn.Dropout(config.dropout)
+
+        # Position encoding components
+        if config.pos_encoding == "rope":
+            self.rotary_emb = RotaryEmbedding(
+                dim=self.head_dim,
+                max_seq_len=config.n_ctx,
+                theta=config.rope_theta,
+            )
+        elif config.pos_encoding == "alibi":
+            # Precompute ALiBi slopes
+            slopes = build_alibi_slopes(config.n_head)
+            self.register_buffer("alibi_slopes", slopes, persistent=False)
+
+        # 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 to match the number of query heads.
+
+        For GQA/MQA, we need to expand K and V to match Q's head count.
+        Input shape: (batch, n_kv_heads, seq_len, head_dim)
+        Output shape: (batch, n_head, seq_len, head_dim)
+        """
+        if self.n_rep == 1:
+            return x
+        batch, n_kv_heads, seq_len, head_dim = x.shape
+        x = x.unsqueeze(2).expand(batch, n_kv_heads, self.n_rep, seq_len, head_dim)
+        return x.reshape(batch, n_kv_heads * 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()
+
+        # Compute Q, K, V projections
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        # 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_kv_heads, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.n_kv_heads, self.head_dim).transpose(1, 2)
+
+        # Apply rotary embeddings if using RoPE
+        if self.pos_encoding == "rope":
+            cos, sin = self.rotary_emb(q, seq_len)
+            q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        # Repeat K and V for GQA/MQA
+        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 ALiBi bias if using ALiBi
+        if self.pos_encoding == "alibi":
+            alibi_bias = build_alibi_bias(seq_len, self.alibi_slopes.to(x.device))
+            attn_weights = attn_weights + alibi_bias.to(attn_weights.dtype)
+
+        # Apply causal mask
+        causal_mask = self.causal_mask[:, :, :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 = 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
+
+
+# Alias for backward compatibility
+MultiHeadAttention = Attention
+
+
+# ==============================================================================
+# Feed-Forward Modules
+# ==============================================================================
+
+
+class FeedForward(nn.Module):
+    """
+    Feed-forward network (MLP) module with configurable activation.
+
+    Supports:
+    - gelu: Standard GELU activation (2 weight matrices)
+    - swiglu: SwiGLU activation (3 weight matrices, better performance)
+
+    For SwiGLU, the hidden dimension is adjusted to keep parameter count similar:
+    - GELU: 2 * n_embd * n_inner parameters
+    - SwiGLU: 3 * n_embd * n_inner_swiglu parameters
+    To match, n_inner_swiglu = 2/3 * n_inner
+    """
+
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+
+        self.ffn_type = config.ffn_type
+
+        if config.ffn_type == "swiglu":
+            # SwiGLU uses 3 projections, so reduce hidden dim to compensate
+            # Adjust n_inner for SwiGLU to maintain similar parameter count
+            hidden_dim = int(2 * config.n_inner / 3)
+            # Round to nearest multiple of 8 for efficiency
+            hidden_dim = ((hidden_dim + 7) // 8) * 8
+
+            self.w1 = nn.Linear(config.n_embd, hidden_dim, bias=False)  # Gate
+            self.w2 = nn.Linear(config.n_embd, hidden_dim, bias=False)  # Up
+            self.w3 = nn.Linear(hidden_dim, config.n_embd, bias=False)  # Down
+        else:  # gelu
+            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:
+        if self.ffn_type == "swiglu":
+            # SwiGLU: Swish(W1*x) * W2*x, then W3
+            gate = F.silu(self.w1(x))  # Swish activation
+            up = self.w2(x)
+            x = gate * up
+            x = self.w3(x)
+            x = self.dropout(x)
+        else:  # gelu
+            x = self.c_fc(x)
+            x = F.gelu(x)
+            x = self.c_proj(x)
+            x = self.dropout(x)
+        return x
+
+
+# ==============================================================================
+# Transformer Block
+# ==============================================================================
+
+
+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.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = Attention(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:
+        # 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
+
+
+# ==============================================================================
+# Main Model
+# ==============================================================================
+
+
+class ChessForCausalLM(PreTrainedModel, GenerationMixin):
+    """
+    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 modular GPT-style architecture with:
+    - Token embeddings for chess moves
+    - Configurable positional embeddings (learned/RoPE/ALiBi)
+    - Stacked transformer blocks with configurable attention (MHA/GQA/MQA)
+    - Configurable FFN activation (GELU/SwiGLU)
+    - Linear head for next-token prediction
+
+    The model supports weight tying between the embedding layer and the
+    output projection to save parameters.
+
+    Example:
+        >>> # Baseline configuration
+        >>> config = ChessConfig(vocab_size=1200, n_embd=128, n_layer=6)
+        >>> model = ChessForCausalLM(config)
+
+        >>> # GQA with RoPE (saves parameters, allows more layers)
+        >>> config = ChessConfig(
+        ...     vocab_size=1200, n_embd=128, n_layer=8,
+        ...     attention_type="gqa", n_kv_heads=2,
+        ...     pos_encoding="rope"
+        ... )
+        >>> model = ChessForCausalLM(config)
+    """
+
+    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)
+
+        self.pos_encoding = config.pos_encoding
+
+        # Token embeddings (always needed)
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+
+        # Position embeddings (only for learned position encoding)
+        if config.pos_encoding == "learned":
+            self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
+        else:
+            # RoPE and ALiBi don't need position embeddings
+            self.wpe = None
+
+        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 prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Tuple] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> dict:
+        # No KV-cache support; fall back to full forward each step.
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "past_key_values": past_key_values,
+        }
+    
+    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,
+        legal_token_ids: Optional[List[List[int]]] = 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
+
+        # Get token embeddings
+        hidden_states = self.wte(input_ids)
+
+        # Add position embeddings only for learned encoding
+        if self.pos_encoding == "learned":
+            if position_ids is None:
+                position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
+            position_embeds = self.wpe(position_ids)
+            hidden_states = hidden_states + position_embeds
+
+        # Apply dropout
+        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 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 self.config.legal_loss_weight > 0 and legal_token_ids:
+                aux_loss = self._legal_move_loss(logits, labels, legal_token_ids)
+                if aux_loss is not None:
+                    loss = loss + self.config.legal_loss_weight * aux_loss
+        
+        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 _legal_move_loss(
+        self,
+        logits: torch.Tensor,
+        labels: torch.Tensor,
+        legal_token_ids: List[List[int]],
+    ) -> Optional[torch.Tensor]:
+        batch_size = logits.size(0)
+        total_loss = logits.new_tensor(0.0)
+        count = 0
+
+        for batch_idx in range(batch_size):
+            if batch_idx >= len(legal_token_ids):
+                continue
+            legal_ids = legal_token_ids[batch_idx]
+            if not legal_ids:
+                continue
+
+            label_row = labels[batch_idx]
+            valid_mask = label_row != -100
+            for special_id in (
+                getattr(self.config, "pad_token_id", None),
+                getattr(self.config, "bos_token_id", None),
+                getattr(self.config, "eos_token_id", None),
+            ):
+                if special_id is not None:
+                    valid_mask = valid_mask & (label_row != int(special_id))
+
+            valid_positions = valid_mask.nonzero(as_tuple=False)
+            if valid_positions.numel() == 0:
+                continue
+
+            last_pos = int(valid_positions[-1].item())
+            pred_pos = last_pos - 1
+            if pred_pos < 0:
+                continue
+
+            logits_slice = logits[batch_idx, pred_pos]
+            legal_logits = logits_slice.index_select(
+                0,
+                torch.tensor(legal_ids, device=logits_slice.device, dtype=torch.long),
+            )
+
+            loss = torch.logsumexp(logits_slice, dim=-1) - torch.logsumexp(legal_logits, dim=-1)
+            total_loss = total_loss + loss
+            count += 1
+
+        if count == 0:
+            return None
+        return total_loss / count
+    
+    @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)