""" 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, List 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. """ model_type = "chess_transformer" def __init__( self, vocab_size: int = 200, # Approx size for component vocab n_embd: int = 120, # Reduced to be divisible by heads and fit budget n_layer: int = 6, n_head: int = 4, n_ctx: int = 250, # Max moves (not tokens) 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) class MultiHeadAttention(nn.Module): def __init__(self, config: ChessConfig): super().__init__() assert config.n_embd % config.n_head == 0 self.n_head = config.n_head self.n_embd = config.n_embd self.head_dim = config.n_embd // config.n_head 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) 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, attention_mask=None): batch_size, seq_len, _ = x.size() qkv = self.c_attn(x) q, k, v = qkv.split(self.n_embd, dim=2) 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) attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim) causal_mask = self.bias[:, :, :seq_len, :seq_len] attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf")) if attention_mask is not None: # Mask should be broadcastable attn_weights = attn_weights + attention_mask attn_weights = F.softmax(attn_weights, dim=-1) attn_weights = self.dropout(attn_weights) attn_output = torch.matmul(attn_weights, v) attn_output = attn_output.transpose(1, 2).contiguous().view( batch_size, seq_len, self.n_embd ) return self.c_proj(attn_output) class FeedForward(nn.Module): 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): x = self.c_fc(x) x = F.gelu(x) x = self.c_proj(x) x = self.dropout(x) return x class TransformerBlock(nn.Module): 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, attention_mask=None): x = x + self.attn(self.ln_1(x), attention_mask=attention_mask) x = x + self.mlp(self.ln_2(x)) return x class ChessForCausalLM(PreTrainedModel): config_class = ChessConfig base_model_prefix = "transformer" supports_gradient_checkpointing = True def __init__(self, config: ChessConfig): super().__init__(config) # Component embeddings (Color, Piece, Src, Dst, Suffix) self.wte_color = nn.Embedding(config.vocab_size, config.n_embd) self.wte_piece = nn.Embedding(config.vocab_size, config.n_embd) self.wte_src = nn.Embedding(config.vocab_size, config.n_embd) self.wte_dst = nn.Embedding(config.vocab_size, config.n_embd) self.wte_suf = nn.Embedding(config.vocab_size, config.n_embd) self.wpe = nn.Embedding(config.n_ctx, config.n_embd) self.drop = nn.Dropout(config.dropout) self.h = nn.ModuleList([ TransformerBlock(config) for _ in range(config.n_layer) ]) self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon) # 5 Heads for predicting next components # We model p(NextMove | History). # Components of NextMove are predicted conditionally independent given History (simplification) # or we could make them autoregressive within the move. # For "product encoding", parallel prediction is natural. self.head_color = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.head_piece = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.head_src = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.head_dst = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.head_suf = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.post_init() def _init_weights(self, module): 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 get_input_embeddings(self): # Return first embedding as proxy, though we have multiple return self.wte_color 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]: return_dict = return_dict if return_dict is not None else self.config.use_return_dict batch_size, seq_len = input_ids.size() # Ensure sequence length is multiple of 5 if seq_len % 5 != 0: # Pad or truncate? For training we expect aligned batches # Truncate to nearest multiple of 5 new_len = (seq_len // 5) * 5 input_ids = input_ids[:, :new_len] if labels is not None: labels = labels[:, :new_len] if attention_mask is not None: attention_mask = attention_mask[:, :new_len] seq_len = new_len num_moves = seq_len // 5 # Reshape to (B, L, 5) # Components: 0=Color, 1=Piece, 2=Src, 3=Dst, 4=Suf reshaped_ids = input_ids.view(batch_size, num_moves, 5) # Product Embedding emb_c = self.wte_color(reshaped_ids[:, :, 0]) emb_p = self.wte_piece(reshaped_ids[:, :, 1]) emb_s = self.wte_src(reshaped_ids[:, :, 2]) emb_d = self.wte_dst(reshaped_ids[:, :, 3]) emb_f = self.wte_suf(reshaped_ids[:, :, 4]) # Element-wise product token_embeds = emb_c * emb_p * emb_s * emb_d * emb_f # Position Embeddings device = input_ids.device if position_ids is None: position_ids = torch.arange(num_moves, device=device).unsqueeze(0) position_embeds = self.wpe(position_ids) hidden_states = self.drop(token_embeds + position_embeds) # Attention mask adaptation # input mask is (B, 5L). We need (B, L). # We consider a move valid if ALL components are valid? Or ANY? # Typically padding is consistent. if attention_mask is not None: # Take every 5th element or min/max reshaped_mask = attention_mask.view(batch_size, num_moves, 5) # If any part is unmasked (1), keep it? # Usually PAD=0. If all are PAD, then 0. chess_mask = reshaped_mask.all(dim=-1).float() # (B, L) # Standard broadcast for attention: (B, 1, 1, L) extended_attention_mask = (1.0 - chess_mask) * -10000.0 extended_attention_mask = extended_attention_mask.unsqueeze(1).unsqueeze(2) else: extended_attention_mask = None # Transformer for block in self.h: hidden_states = block(hidden_states, attention_mask=extended_attention_mask) hidden_states = self.ln_f(hidden_states) # Output Heads (Predicting Next Move Components) logits_c = self.head_color(hidden_states) logits_p = self.head_piece(hidden_states) logits_s = self.head_src(hidden_states) logits_d = self.head_dst(hidden_states) logits_f = self.head_suf(hidden_states) # Stack logits: (B, L, 5, V) logits_stacked = torch.stack([logits_c, logits_p, logits_s, logits_d, logits_f], dim=2) # Compute Loss loss = None if labels is not None: # Reshape labels: (B, L, 5) labels_reshaped = labels.view(batch_size, num_moves, 5) # Shift: Hidden[t] predicts Labels[t+1] shift_logits = logits_stacked[:, :-1, :, :].contiguous() shift_labels = labels_reshaped[:, 1:, :].contiguous() # Flatten loss_fct = nn.CrossEntropyLoss(ignore_index=-100) loss = loss_fct( shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1) ) # Return structured output # To satisfy Trainer, we might need to return (B, 5L, V) logits? # But we produced (B, L, 5, V). Flattening gives (B, 5L, V). # Trainer expects logits matching input length usually, or labels length. flat_logits = logits_stacked.view(batch_size, -1, self.config.vocab_size) if not return_dict: output = (flat_logits,) return ((loss,) + output) if loss is not None else output return CausalLMOutputWithPast( loss=loss, logits=flat_logits, ) @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, ) -> List[int]: """ Generate the next move (5 tokens). """ self.eval() # Forward pass # input_ids (1, 5L) outputs = self(input_ids) # Logits: (1, 5L, V) # We want the last move prediction. # The logits for the NEXT move are at the very end. # Specifically, the last block of 5 logits corresponds to predictions from the last hidden state. # Check dimensions next_move_logits = outputs.logits[:, -5:, :] # (1, 5, V) generated = [] for i in range(5): logits = next_move_logits[:, i, :] / temperature # Apply filtering if top_k is not None: v, _ = torch.topk(logits, top_k) logits[logits < v[:, [-1]]] = -float('Inf') probs = F.softmax(logits, dim=-1) next_token = torch.multinomial(probs, num_samples=1) generated.append(next_token.item()) return generated # Register from transformers import AutoConfig, AutoModelForCausalLM AutoConfig.register("chess_transformer", ChessConfig) AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)