File size: 20,152 Bytes
f277b95 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 | """
Chess Transformer Model for the Chess Challenge.
Modifications:
- RoPE positional encoding controlled by config.use_rope (default: True)
- Optional one-hot embeddings controlled by config.one_hot_embeds (default: False)
- GPU/torch.compile-friendly attention via torch.nn.functional.scaled_dot_product_attention (SDPA)
Key components:
- ChessConfig: Configuration class for model hyperparameters
- ChessForCausalLM: The main model class for next-move prediction
"""
from __future__ import annotations
from typing import Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast
class ChessConfig(PretrainedConfig):
"""
Configuration class for the Chess Transformer model.
New attributes:
use_rope: If True, use Rotary Positional Embeddings (RoPE) instead of learned absolute positions.
rope_theta: Base (theta) for RoPE frequencies (default: 10000.0).
one_hot_embeds: If True, compute token embeddings via one-hot -> matmul with embedding matrix.
(More expensive; intended for experiments.)
"""
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 = True,
# NEW:
use_rope: bool = True,
rope_theta: float = 10000.0,
one_hot_embeds: bool = False,
# Tokens:
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.one_hot_embeds = bool(one_hot_embeds)
d_model = self.vocab_size if self.one_hot_embeds else n_embd
self.n_embd = d_model
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 # keep your budget choice
self.dropout = dropout
self.layer_norm_epsilon = layer_norm_epsilon
self.tie_weights = bool(tie_weights) and (not self.one_hot_embeds)
self.tie_word_embeddings = bool(self.tie_weights)
self.use_rope = bool(use_rope)
self.rope_theta = float(rope_theta)
self.one_hot_embeds = bool(one_hot_embeds)
# Inform HF base class about tying behavior
self.tie_word_embeddings = bool(tie_weights)
if self.n_embd % self.n_head != 0:
raise ValueError(
f"n_embd ({self.n_embd}) must be divisible by n_head ({self.n_head}). "
f"(If one_hot_embeds=True, n_embd=vocab_size={self.vocab_size})"
)
head_dim = self.n_embd // self.n_head
if self.use_rope and (head_dim % 2 != 0):
raise ValueError(
f"RoPE requires even head_dim, got head_dim={head_dim}. "
f"Choose n_head such that (n_embd/n_head) is even."
)
class OneHotEmbedding(nn.Module):
"""
True one-hot embedding:
token i -> e_i in R^V (V = vocab_size)
- No parameters
- No embedding matrix
- Returns a dense (B, L, V) tensor (this is inherently expensive)
"""
def __init__(self, vocab_size: int):
super().__init__()
self.vocab_size = int(vocab_size)
def forward(self, input_ids: torch.LongTensor) -> torch.Tensor:
# Pick a dtype that matches autocast (saves memory in bf16/fp16)
if torch.is_autocast_enabled():
if input_ids.is_cuda:
dtype = torch.get_autocast_gpu_dtype()
else:
dtype = torch.get_autocast_cpu_dtype()
else:
dtype = torch.float32
# Allocate the dense one-hot tensor directly in compute dtype
# Shape: (B, L, V)
out = torch.zeros(
(*input_ids.shape, self.vocab_size),
device=input_ids.device,
dtype=dtype,
)
out.scatter_(-1, input_ids.unsqueeze(-1), 1.0)
return out
class RMSNorm(nn.Module):
"""
Root Mean Square Layer Normalization.
Simpler and faster than LayerNorm - used in LLaMA, Mistral, etc.
Does not center (no mean subtraction), only scales by RMS.
"""
def __init__(self, hidden_size: int, eps: float = 1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.eps = eps
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x: (B, L, D)
# Compute RMS: sqrt(mean(x^2))
rms = torch.sqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
# Normalize and scale
return x / rms * self.weight
class RotaryEmbedding(nn.Module):
"""
Rotary Positional Embedding (RoPE) with precomputed sin/cos cache.
Cache is created up to max_position_embeddings and stored as buffers (not in state_dict).
Applies RoPE to Q and K in (B, H, L, D) format.
"""
def __init__(self, dim: int, max_position_embeddings: int, base: float = 10000.0):
super().__init__()
if dim % 2 != 0:
raise ValueError(f"RoPE requires an even dim, got dim={dim}")
self.dim = dim
self.max_position_embeddings = max_position_embeddings
# inv_freq: (dim/2,)
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
# Precompute cos/sin: (max_pos, dim/2)
t = torch.arange(max_position_embeddings, dtype=inv_freq.dtype)
freqs = torch.einsum("i,j->ij", t, inv_freq)
self.register_buffer("cos_cached", freqs.cos(), persistent=False)
self.register_buffer("sin_cached", freqs.sin(), persistent=False)
@staticmethod
def _apply_rotary(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
# x: (B, H, L, D)
# cos/sin: broadcastable to (B or 1, 1, L, D/2)
x1 = x[..., ::2] # (B,H,L,D/2)
x2 = x[..., 1::2] # (B,H,L,D/2)
# Apply rotation
# [x1; x2] -> [x1*cos - x2*sin ; x1*sin + x2*cos]
y1 = x1 * cos - x2 * sin
y2 = x1 * sin + x2 * cos
# Interleave back to (B,H,L,D)
return torch.stack((y1, y2), dim=-1).flatten(-2)
def forward(
self,
q: torch.Tensor,
k: torch.Tensor,
position_ids: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
# q,k: (B, H, L, D)
B, H, L, D = q.shape
if position_ids is None:
# Fast path: positions [0..L-1], same for whole batch
if L > self.cos_cached.size(0):
raise ValueError(
f"Sequence length {L} exceeds RoPE cache size {self.cos_cached.size(0)}. "
f"Increase n_ctx/max_position_embeddings."
)
cos = self.cos_cached[:L].to(dtype=q.dtype).unsqueeze(0).unsqueeze(0) # (1,1,L,D/2)
sin = self.sin_cached[:L].to(dtype=q.dtype).unsqueeze(0).unsqueeze(0)
else:
# position_ids: (B, L) (or (L,))
if position_ids.dim() == 1:
position_ids = position_ids.unsqueeze(0).expand(B, -1)
flat = position_ids.reshape(-1) # (B*L,)
cos = (
self.cos_cached.index_select(0, flat)
.reshape(B, L, -1)
.to(dtype=q.dtype)
.unsqueeze(1) # (B,1,L,D/2)
)
sin = (
self.sin_cached.index_select(0, flat)
.reshape(B, L, -1)
.to(dtype=q.dtype)
.unsqueeze(1) # (B,1,L,D/2)
)
q = self._apply_rotary(q, cos, sin)
k = self._apply_rotary(k, cos, sin)
return q, k
class MultiHeadAttention(nn.Module):
"""
Multi-head self-attention using SDPA, with correct padding masking.
- If attention_mask is provided, we build a boolean "keep mask" that combines:
* causal mask (lower triangular)
* key padding mask
and call SDPA with is_causal=False (mask already contains causal).
- If attention_mask is None, we call SDPA with is_causal=True (fast path).
"""
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
self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
self.c_proj = nn.Linear(config.n_embd, config.n_embd)
self.attn_dropout_p = float(config.dropout)
self.use_rope = bool(getattr(config, "use_rope", False))
if self.use_rope:
if self.head_dim % 2 != 0:
raise ValueError(f"RoPE requires even head_dim, got {self.head_dim}")
self.rope = RotaryEmbedding(
dim=self.head_dim,
max_position_embeddings=config.n_ctx,
base=float(getattr(config, "rope_theta", 10000.0)),
)
else:
self.rope = None
# Causal keep-mask buffer: True means "allowed"
self.register_buffer(
"causal_keep",
torch.tril(torch.ones(config.n_ctx, config.n_ctx, dtype=torch.bool)).view(
1, 1, config.n_ctx, config.n_ctx
),
persistent=False,
)
def forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
B, L, _ = x.size()
qkv = self.c_attn(x)
q, k, v = qkv.split(self.n_embd, dim=2)
q = q.reshape(B, L, self.n_head, self.head_dim).transpose(1, 2) # (B,H,L,D)
k = k.reshape(B, L, self.n_head, self.head_dim).transpose(1, 2)
v = v.reshape(B, L, self.n_head, self.head_dim).transpose(1, 2)
if self.use_rope:
q, k = self.rope(q, k, position_ids=position_ids)
dropout_p = self.attn_dropout_p if self.training else 0.0
# Correct masking (equivalent to the old code):
# - Old code: causal mask + key padding mask applied to attention scores.
# - Here: build a boolean keep-mask (True=keep, False=masked) for SDPA.
if attention_mask is None:
attn_mask = None
is_causal = True
else:
# key_keep: (B,1,1,L) True for real tokens, False for pads
key_keep = attention_mask[:, None, None, :].to(dtype=torch.bool)
# causal_keep: (1,1,L,L)
causal_keep = self.causal_keep[:, :, :L, :L]
# combined: (B,1,L,L) via broadcast
attn_mask = causal_keep & key_keep
is_causal = False # mask already contains causal
attn_output = F.scaled_dot_product_attention(
q,
k,
v,
attn_mask=attn_mask,
dropout_p=dropout_p,
is_causal=is_causal,
) # (B,H,L,D)
attn_output = attn_output.transpose(1, 2).reshape(B, L, self.n_embd)
return self.c_proj(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 = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.attn = MultiHeadAttention(config)
self.ln_2 = 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_ids: Optional[torch.Tensor] = None,
) -> torch.Tensor:
x = x + self.attn(self.ln_1(x), attention_mask=attention_mask, position_ids=position_ids)
x = x + self.mlp(self.ln_2(x))
return x
class ChessForCausalLM(PreTrainedModel):
"""
Chess Transformer for Causal Language Modeling (next-move prediction).
RoPE:
- If config.use_rope=True (default), no learned positional embeddings are used.
- RoPE is applied inside attention on Q and K.
One-hot embeddings:
- If config.one_hot_embeds=True, input embeddings are computed as:
one_hot(input_ids) @ wte.weight
This is heavier than nn.Embedding lookup, but matches the requested behavior.
"""
config_class = ChessConfig
base_model_prefix = "transformer"
supports_gradient_checkpointing = True
keys_to_ignore_on_load_missing = ["lm_head.weight"]
def __init__(self, config: ChessConfig):
super().__init__(config)
if config.one_hot_embeds:
self.wte = OneHotEmbedding(config.vocab_size)
else:
self.wte = nn.Embedding(config.vocab_size, config.n_embd)
# Positional embeddings only if not using RoPE
self.wpe = None if getattr(config, "use_rope", False) else 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 = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
if config.tie_weights:
self._tied_weights_keys = ["lm_head.weight"]
self.post_init()
if config.tie_weights and (not config.one_hot_embeds):
self.tie_weights()
def get_input_embeddings(self) -> nn.Module:
return self.wte
def set_input_embeddings(self, new_embeddings: nn.Module):
self.wte = new_embeddings
if getattr(self.config, "tie_weights", False):
self.tie_weights()
def get_output_embeddings(self) -> nn.Module:
return self.lm_head
def set_output_embeddings(self, new_embeddings: nn.Module):
self.lm_head = new_embeddings
def tie_weights(self):
if getattr(self.config, "one_hot_embeds", False):
return
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, 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]:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
B, L = input_ids.size()
device = input_ids.device
use_rope = bool(getattr(self.config, "use_rope", False))
one_hot_embeds = bool(getattr(self.config, "one_hot_embeds", False))
# Only build position_ids when needed for learned absolute positions.
# For RoPE, position_ids can be None (fast path), unless caller provides custom position_ids.
if (position_ids is None) and (not use_rope):
position_ids = torch.arange(L, device=device).unsqueeze(0).expand(B, -1)
# Token embeddings
if one_hot_embeds:
token_embeds = self.wte(input_ids)
hidden_states = token_embeds
else:
token_embeds = self.wte(input_ids)
hidden_states = token_embeds
# Absolute learned positions only if RoPE disabled
if not use_rope:
if self.wpe is None:
raise RuntimeError("wpe is None but use_rope is False (inconsistent init).")
pos_embeds = self.wpe(position_ids)
hidden_states = hidden_states + pos_embeds
# Optional: zero out padded positions early (cheap)
if attention_mask is not None:
hidden_states = hidden_states * attention_mask.unsqueeze(-1).to(dtype=hidden_states.dtype)
hidden_states = self.drop(hidden_states)
for block in self.h:
hidden_states = block(hidden_states, attention_mask=attention_mask, position_ids=position_ids)
hidden_states = self.ln_f(hidden_states)
logits = self.lm_head(hidden_states)
loss = None
if labels is not None:
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
loss = loss_fct(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
)
if not return_dict:
output = (logits,)
return ((loss,) + output) if loss is not None else output
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=None,
hidden_states=None,
attentions=None,
)
@torch.no_grad()
def generate_move(
self,
input_ids: torch.LongTensor,
temperature: float = 1.0,
top_k: Optional[int] = None,
top_p: Optional[float] = None,
) -> int:
self.eval()
outputs = self(input_ids)
logits = outputs.logits[:, -1, :] / temperature
if top_k is not None:
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits[indices_to_remove] = float("-inf")
if top_p is not None:
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
indices_to_remove = sorted_indices_to_remove.scatter(
dim=-1, index=sorted_indices, src=sorted_indices_to_remove
)
logits[indices_to_remove] = float("-inf")
probs = F.softmax(logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
return next_token.item()
# Register the model with Auto classes for easy loading
from transformers import AutoConfig, AutoModelForCausalLM
AutoConfig.register("chess_transformer", ChessConfig)
AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)
|