Instructions to use LLM-course/amine-chess-optimal-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LLM-course/amine-chess-optimal-v2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="LLM-course/amine-chess-optimal-v2", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("LLM-course/amine-chess-optimal-v2", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LLM-course/amine-chess-optimal-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LLM-course/amine-chess-optimal-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LLM-course/amine-chess-optimal-v2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/LLM-course/amine-chess-optimal-v2

SGLang

How to use LLM-course/amine-chess-optimal-v2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LLM-course/amine-chess-optimal-v2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LLM-course/amine-chess-optimal-v2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LLM-course/amine-chess-optimal-v2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LLM-course/amine-chess-optimal-v2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use LLM-course/amine-chess-optimal-v2 with Docker Model Runner:
```
docker model run hf.co/LLM-course/amine-chess-optimal-v2
```

amine-chess-optimal-v2 / model.py

Bichrai

Chess Challenge submission by Bichrai

ead5630 verified 10 months ago

Raw History Blame Contribute Delete

19.5 kB

"""
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

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 RMSNorm(nn.Module):
    """
    Root Mean Square Layer Normalization.
    RMSNorm is more efficient than LayerNorm as it:
    - Does not subtract mean (re-centering)
    - Does not have a bias parameter
    - Only has scale parameter (weight)
    This saves computation and parameters while maintaining performance.
    Paper: https://arxiv.org/abs/1910.07467
    """

    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x: torch.Tensor) -> torch.Tensor:
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        output = self._norm(x.float()).type_as(x)
        return output * self.weight


class RotaryPositionalEmbedding(nn.Module):
    """
    Rotary Position Embedding (RoPE).
    RoPE encodes absolute positional information with rotation matrices
    and incorporates relative positional information naturally.
    Advantages over learned embeddings:
    - No additional parameters (saves n_ctx * n_embd parameters)
    - Better extrapolation to longer sequences
    - Encodes relative positions naturally
    Paper: https://arxiv.org/abs/2104.09864
    """

    def __init__(self, dim: int, max_seq_len: int = 2048, base: int = 10000):
        super().__init__()
        self.dim = dim
        self.max_seq_len = max_seq_len
        self.base = base

        # Compute inverse frequencies
        inv_freq = 1.0 / (self.base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        # Pre-compute frequencies for max_seq_len
        self._set_cos_sin_cache(max_seq_len)

    def _set_cos_sin_cache(self, seq_len: int):
        self.max_cached_len = seq_len
        t = torch.arange(seq_len, device=self.inv_freq.device).type_as(self.inv_freq)
        freqs = torch.outer(t, self.inv_freq)
        # Different from paper, but uses a different permutation to get same result
        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]:
        # x: [batch_size, seq_len, n_head, head_dim]
        if seq_len > self.max_cached_len:
            self._set_cos_sin_cache(seq_len)

        return (
            self.cos_cached[:seq_len, ...].to(x.device),
            self.sin_cached[:seq_len, ...].to(x.device),
        )


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.
    Args:
        q: Query tensor of shape [batch_size, seq_len, n_head, head_dim]
        k: Key tensor of shape [batch_size, seq_len, n_head, head_dim]
        cos: Cosine values of shape [seq_len, head_dim]
        sin: Sine values of shape [seq_len, head_dim]
    Returns:
        Tuple of rotated (q, k) tensors
    """
    # Reshape cos and sin for broadcasting
    cos = cos.unsqueeze(0).unsqueeze(2)  # [1, seq_len, 1, head_dim]
    sin = sin.unsqueeze(0).unsqueeze(2)  # [1, seq_len, 1, head_dim]

    # Rotate half
    def rotate_half(x):
        x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
        return torch.cat((-x2, x1), dim=-1)

    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)

    return q_embed, k_embed


class ChessConfig(PretrainedConfig):
    """
    Configuration class for the Chess Transformer model.
    This configuration is designed for a ~1M parameter model.
    Optimizations applied:
    - RMSNorm instead of LayerNorm (fewer params, same performance)
    - RoPE instead of learned positional embeddings (saves n_ctx * n_embd params)
    - Weight tying between embeddings and output (saves vocab_size * n_embd params)
    Parameter budget breakdown (with default values):
    - Token Embeddings: 1200 x 128 = 153,600
    - Position Embeddings: 0 (using RoPE - no parameters)
    - Transformer Layers: 6 x ~115,000 = ~690,000 (RMSNorm saves params)
    - LM Head (with weight tying): 0 (shared with embeddings)
    - Total: ~843,000 parameters (157k saved vs LayerNorm + learned pos embeddings)
    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.
        rms_norm_eps: Epsilon for RMS normalization.
        tie_weights: Whether to tie embedding and output weights.
        rope_base: Base for RoPE frequency calculation.
    """
    
    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,
        rms_norm_eps: float = 1e-6,
        tie_weights: bool = True,
        rope_base: int = 10000,
        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.rms_norm_eps = rms_norm_eps
        self.tie_weights = tie_weights
        self.rope_base = rope_base
        # Inform HF base class about tying behavior
        self.tie_word_embeddings = bool(tie_weights)


class MultiHeadAttention(nn.Module):
    """
    Multi-head self-attention module with RoPE.
    This implementation uses:
    - Rotary Position Embeddings (RoPE) for position encoding
    - Causal masking for autoregressive generation
    - Scaled dot-product attention
    """

    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

        # 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)

        # RoPE for positional encoding
        self.rotary_emb = RotaryPositionalEmbedding(
            dim=self.head_dim,
            max_seq_len=config.n_ctx,
            base=config.rope_base,
        )

        # Causal mask
        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()

        # Compute Q, K, V
        qkv = self.c_attn(x)
        q, k, v = qkv.split(self.n_embd, dim=2)

        # Reshape for multi-head attention: [batch, seq_len, n_head, head_dim]
        q = q.view(batch_size, seq_len, self.n_head, self.head_dim)
        k = k.view(batch_size, seq_len, self.n_head, self.head_dim)
        v = v.view(batch_size, seq_len, self.n_head, self.head_dim)

        # Apply RoPE to Q and K
        cos, sin = self.rotary_emb(q, seq_len)
        q, k = apply_rotary_pos_emb(q, k, cos, sin)

        # Transpose for attention: [batch, n_head, seq_len, head_dim]
        q = q.transpose(1, 2)
        k = k.transpose(1, 2)
        v = v.transpose(1, 2)

        # 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 (RMSNorm before attention/FFN) for better
    training stability and parameter efficiency.
    """

    def __init__(self, config: ChessConfig):
        super().__init__()

        self.rms_1 = RMSNorm(config.n_embd, eps=config.rms_norm_eps)
        self.attn = MultiHeadAttention(config)
        self.rms_2 = RMSNorm(config.n_embd, eps=config.rms_norm_eps)
        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.rms_1(x), attention_mask=attention_mask)
        # Pre-norm FFN
        x = x + self.mlp(self.rms_2(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 an optimized GPT-style architecture with:
    - Token embeddings for chess moves
    - RoPE (Rotary Position Embeddings) - no learned positional embeddings
    - RMSNorm for efficient normalization
    - 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.
    Optimizations:
    - RoPE saves n_ctx * n_embd parameters vs learned embeddings
    - RMSNorm saves parameters vs LayerNorm (no bias, no mean centering)
    - Weight tying saves vocab_size * n_embd 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 embeddings (no positional embeddings - using RoPE instead)
        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 RMS norm
        self.rms_f = RMSNorm(config.n_embd, eps=config.rms_norm_eps)

        # 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 with adaptations for RMSNorm."""
        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, RMSNorm):
            torch.nn.init.ones_(module.weight)
    
    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: Not used (kept for compatibility). Position encoding via RoPE.
            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

        # Get token embeddings (no positional embeddings - using RoPE in attention)
        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 RMS norm
        hidden_states = self.rms_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 = 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)

	"""
	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

	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 RMSNorm(nn.Module):
	"""
	Root Mean Square Layer Normalization.

	RMSNorm is more efficient than LayerNorm as it:
	- Does not subtract mean (re-centering)
	- Does not have a bias parameter
	- Only has scale parameter (weight)

	This saves computation and parameters while maintaining performance.
	Paper: https://arxiv.org/abs/1910.07467
	"""

	def __init__(self, dim: int, eps: float = 1e-6):
	super().__init__()
	self.eps = eps
	self.weight = nn.Parameter(torch.ones(dim))

	def _norm(self, x: torch.Tensor) -> torch.Tensor:
	return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

	def forward(self, x: torch.Tensor) -> torch.Tensor:
	output = self._norm(x.float()).type_as(x)
	return output * self.weight


	class RotaryPositionalEmbedding(nn.Module):
	"""
	Rotary Position Embedding (RoPE).

	RoPE encodes absolute positional information with rotation matrices
	and incorporates relative positional information naturally.

	Advantages over learned embeddings:
	- No additional parameters (saves n_ctx * n_embd parameters)
	- Better extrapolation to longer sequences
	- Encodes relative positions naturally

	Paper: https://arxiv.org/abs/2104.09864
	"""

	def __init__(self, dim: int, max_seq_len: int = 2048, base: int = 10000):
	super().__init__()
	self.dim = dim
	self.max_seq_len = max_seq_len
	self.base = base

	# Compute inverse frequencies
	inv_freq = 1.0 / (self.base ** (torch.arange(0, dim, 2).float() / dim))
	self.register_buffer("inv_freq", inv_freq, persistent=False)

	# Pre-compute frequencies for max_seq_len
	self._set_cos_sin_cache(max_seq_len)

	def _set_cos_sin_cache(self, seq_len: int):
	self.max_cached_len = seq_len
	t = torch.arange(seq_len, device=self.inv_freq.device).type_as(self.inv_freq)
	freqs = torch.outer(t, self.inv_freq)
	# Different from paper, but uses a different permutation to get same result
	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]:
	# x: [batch_size, seq_len, n_head, head_dim]
	if seq_len > self.max_cached_len:
	self._set_cos_sin_cache(seq_len)

	return (
	self.cos_cached[:seq_len, ...].to(x.device),
	self.sin_cached[:seq_len, ...].to(x.device),
	)


	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.

	Args:
	q: Query tensor of shape [batch_size, seq_len, n_head, head_dim]
	k: Key tensor of shape [batch_size, seq_len, n_head, head_dim]
	cos: Cosine values of shape [seq_len, head_dim]
	sin: Sine values of shape [seq_len, head_dim]

	Returns:
	Tuple of rotated (q, k) tensors
	"""
	# Reshape cos and sin for broadcasting
	cos = cos.unsqueeze(0).unsqueeze(2) # [1, seq_len, 1, head_dim]
	sin = sin.unsqueeze(0).unsqueeze(2) # [1, seq_len, 1, head_dim]

	# Rotate half
	def rotate_half(x):
	x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
	return torch.cat((-x2, x1), dim=-1)

	q_embed = (q * cos) + (rotate_half(q) * sin)
	k_embed = (k * cos) + (rotate_half(k) * sin)

	return q_embed, k_embed


	class ChessConfig(PretrainedConfig):
	"""
	Configuration class for the Chess Transformer model.

	This configuration is designed for a ~1M parameter model.
	Optimizations applied:
	- RMSNorm instead of LayerNorm (fewer params, same performance)
	- RoPE instead of learned positional embeddings (saves n_ctx * n_embd params)
	- Weight tying between embeddings and output (saves vocab_size * n_embd params)

	Parameter budget breakdown (with default values):
	- Token Embeddings: 1200 x 128 = 153,600
	- Position Embeddings: 0 (using RoPE - no parameters)
	- Transformer Layers: 6 x ~115,000 = ~690,000 (RMSNorm saves params)
	- LM Head (with weight tying): 0 (shared with embeddings)
	- Total: ~843,000 parameters (157k saved vs LayerNorm + learned pos embeddings)

	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.
	rms_norm_eps: Epsilon for RMS normalization.
	tie_weights: Whether to tie embedding and output weights.
	rope_base: Base for RoPE frequency calculation.
	"""

	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,
	rms_norm_eps: float = 1e-6,
	tie_weights: bool = True,
	rope_base: int = 10000,
	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.rms_norm_eps = rms_norm_eps
	self.tie_weights = tie_weights
	self.rope_base = rope_base
	# Inform HF base class about tying behavior
	self.tie_word_embeddings = bool(tie_weights)


	class MultiHeadAttention(nn.Module):
	"""
	Multi-head self-attention module with RoPE.

	This implementation uses:
	- Rotary Position Embeddings (RoPE) for position encoding
	- Causal masking for autoregressive generation
	- Scaled dot-product attention
	"""

	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

	# 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)

	# RoPE for positional encoding
	self.rotary_emb = RotaryPositionalEmbedding(
	dim=self.head_dim,
	max_seq_len=config.n_ctx,
	base=config.rope_base,
	)

	# Causal mask
	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()

	# Compute Q, K, V
	qkv = self.c_attn(x)
	q, k, v = qkv.split(self.n_embd, dim=2)

	# Reshape for multi-head attention: [batch, seq_len, n_head, head_dim]
	q = q.view(batch_size, seq_len, self.n_head, self.head_dim)
	k = k.view(batch_size, seq_len, self.n_head, self.head_dim)
	v = v.view(batch_size, seq_len, self.n_head, self.head_dim)

	# Apply RoPE to Q and K
	cos, sin = self.rotary_emb(q, seq_len)
	q, k = apply_rotary_pos_emb(q, k, cos, sin)

	# Transpose for attention: [batch, n_head, seq_len, head_dim]
	q = q.transpose(1, 2)
	k = k.transpose(1, 2)
	v = v.transpose(1, 2)

	# 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 (RMSNorm before attention/FFN) for better
	training stability and parameter efficiency.
	"""

	def __init__(self, config: ChessConfig):
	super().__init__()

	self.rms_1 = RMSNorm(config.n_embd, eps=config.rms_norm_eps)
	self.attn = MultiHeadAttention(config)
	self.rms_2 = RMSNorm(config.n_embd, eps=config.rms_norm_eps)
	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.rms_1(x), attention_mask=attention_mask)
	# Pre-norm FFN
	x = x + self.mlp(self.rms_2(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 an optimized GPT-style architecture with:
	- Token embeddings for chess moves
	- RoPE (Rotary Position Embeddings) - no learned positional embeddings
	- RMSNorm for efficient normalization
	- 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.

	Optimizations:
	- RoPE saves n_ctx * n_embd parameters vs learned embeddings
	- RMSNorm saves parameters vs LayerNorm (no bias, no mean centering)
	- Weight tying saves vocab_size * n_embd 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 embeddings (no positional embeddings - using RoPE instead)
	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 RMS norm
	self.rms_f = RMSNorm(config.n_embd, eps=config.rms_norm_eps)

	# 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 with adaptations for RMSNorm."""
	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, RMSNorm):
	torch.nn.init.ones_(module.weight)

	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: Not used (kept for compatibility). Position encoding via RoPE.
	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

	# Get token embeddings (no positional embeddings - using RoPE in attention)
	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 RMS norm
	hidden_states = self.rms_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 = 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)