File size: 36,748 Bytes
b9eef49 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 |
# ====================================================================
# modeling_unified.py
# ====================================================================
"""
Unified Language Model with GPAS + LNS Integration + xIELU Activation + CoLA (Linear Only) + LaX + Weight Tying + Canon Layers (A+C Only)
MIGRATED TO HUGGINGFACE TRANSFORMERS - FINAL VERSION WITH ALL FIXES + CORRECTED LaX IMPLEMENTATION
UPDATED: Standard Transformer with advanced variance control, parameter efficiency, Canon horizontal information flow, and WORKING LaX Inter-Layer
Combines advanced Transformer architecture with CORRECTED variance control mechanisms,
advanced variance control via GPAS and LNS, xIELU activation function, FIXED LaX integration, and Canon Layers (A+C only)
Based on LLaMA 3 architecture with 30M parameters
MIGRATION TO HUGGINGFACE - FINAL FIXED VERSION + LaX CORRECTION:
==============================================================
1. **HUGGINGFACE INTEGRATION**: Migrado de PyTorch Lightning a Transformers v4.53.3
2. **UPDATED API**: processing_class en lugar de tokenizer (deprecated)
3. **UPDATED COMPUTE_LOSS**: Método actualizado con num_items_in_batch parameter
4. **FIXED LOGGING**: Corregido self.log() syntax según documentación oficial HF
5. **RESTORED PAD HANDLING**: pad_token_id → -100 conversion for CrossEntropyLoss (from original code)
6. **NATIVE TORCH COMPILE**: Moved to TrainingArguments (torch_compile=True)
7. **FIXED WEIGHT TYING**: Corrected _tied_weights_keys as class attribute (HF standard)
8. **VALIDATION DIAGNOSTIC**: Added simple method to diagnose validation loss issues
9. **CUSTOM CONFIGURATION**: PretrainedConfig personalizada con todos los parámetros
10. **PRETRAINED MODEL**: Hereda de PreTrainedModel para compatibilidad completa
11. **MAINTAINED OPTIMIZERS**: Muon + AdamW híbrido preservado
12. **MAINTAINED PRECISION**: bf16-true preservado
13. **MAINTAINED TRAINING**: Custom Trainer con todas las métricas y logging
14. **MAINTAINED ARCHITECTURE**: Toda la arquitectura personalizada preservada
15. **AUTO TOKENIZER**: Integración completa con AutoTokenizer dinámico
16. **AUTOCLASS SUPPORT**: Registro completo para AutoConfig y AutoModel
17. **✅ FIXED LaX**: Implementación correcta Inter-Layer con Linear Gate funcional
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
from transformers import (
AutoTokenizer,
AutoConfig,
AutoModel,
AutoModelForCausalLM,
PreTrainedModel,
)
import math
import os
from typing import Optional, Tuple, Dict, Any, cast, List
from flash_attn import flash_attn_func
import numpy as np
# ✅ ABSOLUTE IMPORT - No relative imports for Hub compatibility
from configuration_unified import UnifiedModelConfig
# Fix tokenizer parallelism warnings
os.environ["TOKENIZERS_PARALLELISM"] = "false"
torch.set_float32_matmul_precision('high')
def init_cola_components(A: nn.Linear, B: nn.Linear):
nn.init.kaiming_normal_(A.weight, mode='fan_in', nonlinearity='relu')
nn.init.xavier_normal_(B.weight, gain=0.8)
if B.bias is not None:
nn.init.zeros_(B.bias)
def init_embedding(embedding: nn.Embedding):
nn.init.normal_(embedding.weight, mean=0.0, std=0.02)
class CanonLayer(nn.Module):
def __init__(self, hidden_dim: int, kernel_size: int = 4):
"""
Canon layer using a 1D causal convolution with residual connection.
"""
super().__init__()
self.hidden_dim = hidden_dim
self.kernel_size = kernel_size
# Use causal convolution with explicit initialization
self.causal_conv1d = nn.Conv1d(
in_channels=hidden_dim,
out_channels=hidden_dim,
kernel_size=kernel_size,
groups=hidden_dim, # Depthwise convolution
padding=0, # No automatic padding
bias=True
)
# Initialize weights more conservatively (as per paper)
nn.init.zeros_(self.causal_conv1d.weight)
nn.init.zeros_(self.causal_conv1d.bias)
def forward(self, h: torch.Tensor) -> torch.Tensor:
"""
Applies the Canon layer transformation with causal masking.
"""
# Conv1d expects input shape (batch_size, channels, sequence_length)
h_permuted = h.permute(0, 2, 1) # (batch, hidden_dim, seq_len)
# Add padding of (kernel_size - 1) only to the left side
padding = self.kernel_size - 1
h_padded = F.pad(h_permuted, (padding, 0))
# Apply causal convolution
conv_out = self.causal_conv1d(h_padded)
# Permute back to the original shape
conv_out_permuted = conv_out.permute(0, 2, 1)
# Add the residual connection
output = h + conv_out_permuted
return output
class CoLA_Linear(nn.Module):
def __init__(self, in_features: int, out_features: int, rank: Optional[int] = None, activation=F.gelu, bias: bool = True):
super().__init__()
if rank is None:
rank = in_features // 4
self.rank = rank
self.activation = activation
self.A = nn.Linear(in_features, rank, bias=False)
self.B = nn.Linear(rank, out_features, bias=bias)
init_cola_components(self.A, self.B)
def forward(self, x: torch.Tensor, prev_latent: Optional[torch.Tensor] = None, lax_beta: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Forward pass with optional LaX Inter-Layer integration.
Args:
x: Input tensor
prev_latent: Previous latent from same module type in previous layer (for LaX)
lax_beta: Linear gate parameter (scalar) for LaX
Returns:
Tuple of (output, current_latent) where current_latent can be used for next layer
"""
# Standard CoLA forward: A -> activation
latent = self.A(x)
latent_activated = self.activation(latent)
# Apply LaX Inter-Layer if previous latent exists
if prev_latent is not None and lax_beta is not None and prev_latent.shape == latent_activated.shape:
# Linear Gate: h_i = h_i + β * h_{i-1}
latent_activated = latent_activated + lax_beta * prev_latent
# B projection
output = self.B(latent_activated)
return output, latent_activated
class LayerNormScaling(nn.Module):
def __init__(self, layer_depth: int):
super().__init__()
if layer_depth < 1:
raise ValueError(f"layer_depth debe ser ≥ 1, got {layer_depth}")
self.layer_depth = layer_depth
self.scaling_factor = 1.0 / math.sqrt(float(layer_depth))
def forward(self, normalized_input: torch.Tensor) -> torch.Tensor:
return normalized_input * self.scaling_factor
class GPAS(nn.Module):
def __init__(self, d_model: int):
super().__init__()
self.d_model = d_model
self.alpha = nn.Parameter(torch.zeros(1))
def forward(self, x: torch.Tensor) -> torch.Tensor:
x_detached = x.detach()
scaled_component = F.silu(self.alpha) * x_detached
x_scaled = x - scaled_component
return x_scaled
class RotaryEmbedding(nn.Module):
def __init__(self, dim: int, max_position_embeddings: int = 2048, base: float = 10000):
super().__init__()
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float() / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
def forward(self, x, seq_len=None):
if seq_len is None:
seq_len = x.shape[-2]
t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
freqs = torch.outer(t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos().to(x.dtype), emb.sin().to(x.dtype)
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None):
def rotate_half(x):
x1 = x[..., : x.shape[-1] // 2]
x2 = 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 XIELU(nn.Module):
def __init__(self, alpha_p_init: float = 0.8, alpha_n_init: float = 0.8, beta: float = 0.5):
super().__init__()
self.beta = beta
self.alpha_p = nn.Parameter(torch.log(torch.exp(torch.tensor(alpha_p_init)) - 1))
self.alpha_n = nn.Parameter(torch.log(torch.exp(torch.tensor(alpha_n_init - self.beta)) - 1))
self.register_buffer('eps', torch.tensor(-1e-6))
def forward(self, x: torch.Tensor) -> torch.Tensor:
alpha_p = F.softplus(self.alpha_p)
alpha_n = self.beta + F.softplus(self.alpha_n)
return torch.where(
x > 0,
alpha_p * x * x + self.beta * x,
alpha_n * torch.expm1(torch.clamp(x, min=self.eps)) - alpha_n * x + self.beta * x
)
class StandardMLP(nn.Module):
def __init__(self, hidden_size: int, intermediate_size: int, dropout: float = 0.0, config=None, layer_idx: int = 0):
super().__init__()
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.config = config
self.layer_idx = layer_idx
self.up_proj = CoLA_Linear(hidden_size, intermediate_size, bias=False)
self.down_proj = CoLA_Linear(intermediate_size, hidden_size, bias=False)
if config is not None:
self.activation = XIELU(
alpha_p_init=config.xielu_alpha_p_init,
alpha_n_init=config.xielu_alpha_n_init,
beta=config.xielu_beta
)
else:
self.activation = XIELU(alpha_p_init=0.8, alpha_n_init=0.8, beta=0.5)
self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
# LaX Linear Gate parameters (β scalars)
if config is not None and config.lax_enabled:
self.lax_beta_up = nn.Parameter(torch.full((1,), 0.2)) # 0.0 → 0.2
self.lax_beta_down = nn.Parameter(torch.full((1,), 0.2)) # 0.0 → 0.2
else:
self.lax_beta_up = None
self.lax_beta_down = None
def forward(self, x: torch.Tensor, lax_buffer: Optional[Dict] = None) -> torch.Tensor:
# LaX: Get previous latents from buffer
prev_up_latent = None
prev_down_latent = None
if lax_buffer is not None and self.lax_beta_up is not None:
prev_up_latent = lax_buffer.get(('mlp_up', self.layer_idx - 1))
prev_down_latent = lax_buffer.get(('mlp_down', self.layer_idx - 1))
# Up projection with LaX
intermediate, up_latent = self.up_proj(x, prev_up_latent, self.lax_beta_up)
# Store current up latent for next layer
if lax_buffer is not None:
lax_buffer[('mlp_up', self.layer_idx)] = up_latent.clone()
# Activation and dropout
activated = self.activation(intermediate)
activated = self.dropout(activated)
# Down projection with LaX
output, down_latent = self.down_proj(activated, prev_down_latent, self.lax_beta_down)
# Store current down latent for next layer
if lax_buffer is not None:
lax_buffer[('mlp_down', self.layer_idx)] = down_latent.clone()
return output
class GroupedQueryAttention(nn.Module):
def __init__(self, config, layer_idx: int = 0):
super().__init__()
self.config = config
self.layer_idx = layer_idx
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.num_key_value_heads = config.num_key_value_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
# FANFormer components
self.fanformer_p = getattr(config, 'fanformer_p', 0.15)
self.d_periodic = int(self.hidden_size * self.fanformer_p)
self.d_standard = self.hidden_size - 2 * self.d_periodic
assert self.d_standard > 0, \
f"fanformer_p={self.fanformer_p} is too high. d_standard={self.d_standard} must be > 0"
self.fan_w_p = CoLA_Linear(self.hidden_size, self.d_periodic, bias=False)
self.fan_w_p_bar = CoLA_Linear(self.hidden_size, self.d_standard, bias=False)
self.q_proj = CoLA_Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = CoLA_Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
self.v_proj = CoLA_Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
self.o_proj = CoLA_Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
self.q_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.k_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.v_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.rotary_emb = RotaryEmbedding(
self.head_dim,
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta
)
# LaX Linear Gate parameters (β scalars) - NO o_proj según plan
if config.lax_enabled:
self.lax_beta_q = nn.Parameter(torch.full((1,), 0.2)) # 0.0 → 0.2
self.lax_beta_k = nn.Parameter(torch.full((1,), 0.2)) # 0.0 → 0.2
self.lax_beta_v = nn.Parameter(torch.full((1,), 0.2)) # 0.0 → 0.2
else:
self.lax_beta_q = None
self.lax_beta_k = None
self.lax_beta_v = None
def _fan_layer_prime(self, x: torch.Tensor) -> torch.Tensor:
periodic_proj, _ = self.fan_w_p(x)
standard_proj, _ = self.fan_w_p_bar(x)
cos_component = torch.cos(periodic_proj)
sin_component = torch.sin(periodic_proj)
x_f = torch.cat([cos_component, sin_component, standard_proj], dim=-1)
return x_f
def _compute_flash_attention(
self,
query_states: torch.Tensor,
key_states: torch.Tensor,
value_states: torch.Tensor,
seq_len: int,
position_ids: Optional[torch.Tensor] = None
) -> torch.Tensor:
batch_size = query_states.shape[0]
q_rope = query_states.transpose(1, 2)
k_rope = key_states.transpose(1, 2)
cos, sin = self.rotary_emb(value_states, seq_len=seq_len)
q_rope, k_rope = apply_rotary_pos_emb(q_rope, k_rope, cos, sin, position_ids)
query_states = q_rope.transpose(1, 2)
key_states = k_rope.transpose(1, 2)
from flash_attn import flash_attn_func
attn_output = flash_attn_func(
query_states,
key_states,
value_states,
dropout_p=self.config.attention_dropout if self.training else 0.0,
causal=True,
)
return attn_output
def forward(self, hidden_states, position_ids=None, attention_mask=None, lax_buffer: Optional[Dict] = None):
batch_size, seq_len, _ = hidden_states.shape
enhanced_input = self._fan_layer_prime(hidden_states)
# LaX: Get previous latents from buffer
prev_q_latent = None
prev_k_latent = None
prev_v_latent = None
if lax_buffer is not None and self.lax_beta_q is not None:
prev_q_latent = lax_buffer.get(('attn_q', self.layer_idx - 1))
prev_k_latent = lax_buffer.get(('attn_k', self.layer_idx - 1))
prev_v_latent = lax_buffer.get(('attn_v', self.layer_idx - 1))
# Q/K/V projections with LaX
query_states, q_latent = self.q_proj(enhanced_input, prev_q_latent, self.lax_beta_q)
key_states, k_latent = self.k_proj(enhanced_input, prev_k_latent, self.lax_beta_k)
value_states, v_latent = self.v_proj(enhanced_input, prev_v_latent, self.lax_beta_v)
# Store current latents for next layer
if lax_buffer is not None:
lax_buffer[('attn_q', self.layer_idx)] = q_latent.clone()
lax_buffer[('attn_k', self.layer_idx)] = k_latent.clone()
lax_buffer[('attn_v', self.layer_idx)] = v_latent.clone()
query_states = query_states.view(batch_size, seq_len, self.num_heads, self.head_dim)
key_states = key_states.view(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
value_states = value_states.view(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
q_flat = query_states.reshape(-1, self.head_dim)
k_flat = key_states.reshape(-1, self.head_dim)
v_flat = value_states.reshape(-1, self.head_dim)
q_normalized = self.q_norm(q_flat)
k_normalized = self.k_norm(k_flat)
v_normalized = self.v_norm(v_flat)
query_states = q_normalized.view(batch_size, seq_len, self.num_heads, self.head_dim)
key_states = k_normalized.view(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
value_states = v_normalized.view(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
attn_output = self._compute_flash_attention(
query_states=query_states,
key_states=key_states,
value_states=value_states,
seq_len=seq_len,
position_ids=position_ids
)
attn_output = attn_output.reshape(batch_size, seq_len, self.hidden_size)
# O projection WITHOUT LaX (según plan)
output, _ = self.o_proj(attn_output)
return output
class DecoderLayer(nn.Module):
def __init__(self, config, layer_idx: int):
super().__init__()
self.config = config
self.layer_idx = layer_idx
if layer_idx < 0:
raise ValueError(f"layer_idx debe ser >= 0, got {layer_idx}")
self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.self_attn = GroupedQueryAttention(config, layer_idx)
self.post_attention_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.mlp = StandardMLP(
config.hidden_size,
config.intermediate_size,
config.mlp_dropout,
config,
layer_idx
)
self.dropout_output = nn.Dropout(0.01)
self.lns_attention = LayerNormScaling(layer_depth=layer_idx + 1)
self.lns_mlp = LayerNormScaling(layer_depth=layer_idx + 1)
self.gpas_attention = GPAS(config.hidden_size)
self.gpas_mlp = GPAS(config.hidden_size)
# Canon layers (A+C only)
# Canon-A: Before attention block
if config.canon_enabled and config.canon_a_enabled:
self.canon_a = CanonLayer(config.hidden_size, config.canon_kernel_size)
else:
self.canon_a = None
# Canon-C: Before MLP block
if config.canon_enabled and config.canon_c_enabled:
self.canon_c = CanonLayer(config.hidden_size, config.canon_kernel_size)
else:
self.canon_c = None
def forward(self, hidden_states: torch.Tensor, position_ids: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, lax_buffer: Optional[Dict] = None) -> torch.Tensor:
residual = hidden_states
# Apply Canon-A before attention
if self.canon_a is not None:
hidden_states = self.canon_a(hidden_states)
attention_input = self.input_layernorm(hidden_states)
attention_input = self.lns_attention(attention_input)
attention_output = self.self_attn(attention_input, position_ids, attention_mask, lax_buffer)
hidden_states = residual + attention_output
hidden_states = self.gpas_attention(hidden_states)
hidden_states = self.dropout_output(hidden_states)
residual = hidden_states
# Apply Canon-C before MLP
if self.canon_c is not None:
hidden_states = self.canon_c(hidden_states)
mlp_input = self.post_attention_layernorm(hidden_states)
mlp_input = self.lns_mlp(mlp_input)
mlp_output = self.mlp(mlp_input, lax_buffer)
hidden_states = residual + mlp_output
hidden_states = self.gpas_mlp(hidden_states)
hidden_states = self.dropout_output(hidden_states)
return hidden_states
class UnifiedModel(PreTrainedModel):
"""
UnifiedModel that inherits from PreTrainedModel for full HuggingFace compatibility.
With AutoClass support for seamless Hub integration.
"""
config_class = UnifiedModelConfig
# ✅ FIXED: _tied_weights_keys as class attribute (HuggingFace standard)
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config: UnifiedModelConfig):
super().__init__(config)
self.config = config
if config.vocab_size is None:
raise ValueError("config.vocab_size must be set from tokenizer before model initialization")
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
self.embedding_dropout = nn.Dropout(config.embedding_dropout)
self.output_dropout = nn.Dropout(0.05)
# Create lm_head for output projections
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
self.layers = nn.ModuleList()
for i in range(config.num_hidden_layers):
self.layers.append(DecoderLayer(config, i))
self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
# Initialize weights
self.post_init()
self._print_configuration()
def tie_weights(self):
"""
✅ FIXED: Simplified tie_weights method following HuggingFace standard.
Tie the word embeddings and the output layer.
This is called automatically if config.tie_word_embeddings is True.
"""
if self.config.tie_word_embeddings:
print("🔗 Applying weight tying: lm_head.weight = embed_tokens.weight")
self.lm_head.weight = self.embed_tokens.weight
print("✅ Weight tying successful: Parameters are properly shared")
def _init_weights(self, module):
"""Initialize weights following the custom initialization scheme."""
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.trunc_normal_(module.weight, mean=0.0, std=0.02, a=-0.04, b=0.04)
elif isinstance(module, CoLA_Linear):
pass # CoLA_Linear has its own initialization
def _print_configuration(self):
# Conteo ingenuo de todos los parámetros registrados
total_params_naive = sum(p.numel() for p in self.parameters())
# Conteo inteligente considerando weight tying
total_params_actual = total_params_naive
vocab_params = self.config.vocab_size * self.config.hidden_size
tied_savings = 0
# ✅ CORRECCIÓN: Detectar y ajustar por weight tying real
if self.config.tie_word_embeddings:
# Verificar si los tensors están realmente atados en memoria
embed_weight = self.embed_tokens.weight
lm_head_weight = self.lm_head.weight
if embed_weight is lm_head_weight:
# Los tensors son idénticos - restar la duplicación
tied_savings = vocab_params
total_params_actual = total_params_naive - tied_savings
else:
# Weight tying configurado pero no aplicado aún
tied_savings = 0
# Cálculos de optimización existentes
total_linear_params = 0
total_cola_params = 0
canon_params = 0
lax_params = 0
for name, module in self.named_modules():
if isinstance(module, CoLA_Linear):
in_features = module.A.in_features
out_features = module.B.out_features
rank = module.rank
standard_params = in_features * out_features
cola_params = (in_features * rank) + (rank * out_features)
total_linear_params += standard_params
total_cola_params += cola_params
elif isinstance(module, CanonLayer):
# Canon layer parameters: depthwise conv1d + bias
canon_layer_params = module.hidden_dim * module.kernel_size + module.hidden_dim
canon_params += canon_layer_params
elif hasattr(module, 'lax_beta_q') and module.lax_beta_q is not None:
# Count LaX β parameters
lax_params += 3 # q, k, v
elif hasattr(module, 'lax_beta_up') and module.lax_beta_up is not None:
# Count LaX β parameters
lax_params += 2 # up, down
cola_reduction = ((total_linear_params - total_cola_params) / total_linear_params) * 100 if total_linear_params > 0 else 0
canon_overhead = (canon_params / total_params_actual) * 100 if total_params_actual > 0 else 0
lax_overhead = (lax_params / total_params_actual) * 100 if total_params_actual > 0 else 0
print(f"\n📊 UNIFIED Model + GPAS + LNS + xIELU + CoLA (Linear Only) + LaX + Canon (A+C) + Weight Tying:")
# ✅ MEJORADO: Mostrar conteo real vs ingenuo para transparencia
if self.config.tie_word_embeddings and tied_savings > 0:
print(f"🎯 Total Parameters: {total_params_actual/1e6:.2f}M (effective)")
print(f"📊 Parameter Breakdown:")
print(f" • Naive count: {total_params_naive/1e6:.2f}M (all registered params)")
print(f" • Actual count: {total_params_actual/1e6:.2f}M (after weight tying)")
print(f" • Weight tying savings: {tied_savings/1e6:.2f}M ({tied_savings/total_params_naive*100:.1f}%)")
else:
print(f"🎯 Total Parameters: {total_params_actual/1e6:.2f}M")
print(f"📚 DYNAMIC Vocabulary Size: {self.config.vocab_size} (from tokenizer)")
print(f"🔗 ✅ PROPER Weight Tying: {'ENABLED' if self.config.tie_word_embeddings else 'DISABLED'}")
# ✅ CORRECCIÓN: Mostrar estado real del weight tying
if self.config.tie_word_embeddings:
if tied_savings > 0:
print(f"💾 Weight Tying Status: ✅ ACTIVE (tensors are shared in memory)")
else:
print(f"💾 Weight Tying Status: ⏳ CONFIGURED (will be applied during post_init)")
print(f"🚀 ACTIVATION: xIELU (αp_init={self.config.xielu_alpha_p_init}, αn_init={self.config.xielu_alpha_n_init}, β={self.config.xielu_beta})")
print(f"🔄 UPGRADE: SwiGLU → StandardMLP + xIELU (better efficiency & adaptability)")
print(f"🗜️ CoLA Integration: {cola_reduction:.1f}% parameter reduction in internal projections")
print(f"🔀 LaX Enabled: {'YES' if self.config.lax_enabled else 'NO'} ✅ WORKING Inter-Layer (Linear Gate)")
if self.config.lax_enabled:
print(f" • LaX Method: Inter-Layer with Linear Gate (β scalars)")
print(f" • LaX Applied to: q_proj, k_proj, v_proj, up_proj, down_proj (NOT o_proj)")
print(f" • LaX Parameters: {lax_params} β scalars ({lax_overhead:.6f}% overhead)")
print(f" • LaX Initialization: β=0.0 (conservative start)")
print(f"🎼 Canon Layers Enabled: {'YES' if self.config.canon_enabled else 'NO'} (A+C ONLY)")
if self.config.canon_enabled:
print(f" • Canon-A (Before Attention): {'✅' if self.config.canon_a_enabled else '❌'}")
print(f" • Canon-B (Inside Attention): ❌ PERMANENTLY DISABLED")
print(f" • Canon-C (Before MLP): {'✅' if self.config.canon_c_enabled else '❌'}")
print(f" • Canon-D (Inside MLP): ❌ PERMANENTLY DISABLED")
print(f" • Canon Kernel Size: {self.config.canon_kernel_size}")
print(f" • Canon Parameters Overhead: {canon_overhead:.3f}% ({canon_params/1e3:.1f}K params)")
print(f"⚡ GPAS Enabled: ALWAYS (Dynamic variance control)")
print(f"📏 LNS Enabled: ALWAYS (Static depth scaling)")
print(f"🔧 Variance Control: Triple-level (LNS + GPAS + Canon A+C) ALWAYS")
print(f"🔗 Residual Connections: STANDARD + HORIZONTAL (Canon A+C only)")
print(f"🧹 CLEAN: Standard transformer architecture - CrossEntropyLoss manages PAD naturally")
print(f"⚡ FlashAttention: Scaled Dot-Product Attention with GQA + automatic causal masking")
print(f"🎯 TOKENIZER AGNOSTIC: Dynamic vocab_size and pad_token_id")
print(f"🎯 SIMPLIFIED: CoLA Linear Only + Canon A+C Only = Better performance & less overhead")
print(f"🔗 ✅ FIXED Weight Tying: _tied_weights_keys as class attribute (HF standard)")
print(f"🎼 Canon A+C BENEFITS: Strategic horizontal information flow with minimal parameters")
print(f"🔀 ✅ FIXED LaX: Functional Inter-Layer with ephemeral buffer (no broken reset)")
print(f"🤗 HUGGINGFACE COMPATIBLE: Full PreTrainedModel integration v4.53.3")
print(f"⚡ ✅ NATIVE TORCH COMPILE: Will be handled by TrainingArguments")
print(f"🚀 ✅ AUTOCLASS SUPPORT: Compatible with AutoConfig.from_pretrained() and AutoModel.from_pretrained()")
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
**kwargs
):
batch_size, seq_len = input_ids.shape
# ✅ LaX: Create ephemeral buffer for this forward pass
lax_buffer = {} if self.config.lax_enabled else None
hidden_states = self.embed_tokens(input_ids)
hidden_states = self.embedding_dropout(hidden_states)
for layer in self.layers:
hidden_states = layer(hidden_states, position_ids=position_ids, attention_mask=attention_mask, lax_buffer=lax_buffer)
hidden_states = self.norm(hidden_states)
hidden_states = self.output_dropout(hidden_states)
logits = self.lm_head(hidden_states)
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = nn.CrossEntropyLoss()
shift_logits = shift_logits.view(-1, self.config.vocab_size)
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
# ✅ RESTORED: Change pad tokens to -100 so CrossEntropyLoss ignores them (from original code)
if self.config.pad_token_id is not None:
shift_labels[shift_labels == self.config.pad_token_id] = -100
loss = loss_fct(shift_logits, shift_labels)
# ✅ LaX buffer is automatically cleaned up (ephemeral, goes out of scope)
# Return in HuggingFace format
from transformers.modeling_outputs import CausalLMOutputWithPast
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=None,
hidden_states=None,
attentions=None,
)
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
@torch.no_grad()
def generate(
self,
input_ids: torch.Tensor,
max_new_tokens: int = 50,
temperature: float = 1.0,
top_p: float = 0.9,
top_k: Optional[int] = None,
do_sample: bool = True,
pad_token_id: Optional[int] = None,
eos_token_id: Optional[int] = None,
**kwargs
) -> torch.Tensor:
"""
Generate sequences using the model.
Compatible with AutoModelForCausalLM interface.
"""
# Set default token IDs
if pad_token_id is None:
pad_token_id = self.config.pad_token_id
if eos_token_id is None:
eos_token_id = self.config.eos_token_id
batch_size = input_ids.shape[0]
device = input_ids.device
generated = input_ids.clone()
for _ in range(max_new_tokens):
# Forward pass (LaX buffer is created fresh each time)
outputs = self.forward(generated)
logits = outputs.logits
# Get the logits for the last token
next_token_logits = logits[:, -1, :]
if do_sample:
# Apply temperature
if temperature != 1.0:
next_token_logits = next_token_logits / temperature
# Apply top-k filtering
if top_k is not None:
values, indices = torch.topk(next_token_logits, top_k)
next_token_logits[next_token_logits < values[:, [-1]]] = -float('inf')
# Apply top-p (nucleus) filtering
if top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(next_token_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
# Shift the indices to the right to keep also the first token above the threshold
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
# Scatter sorted tensors to original indexing
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
next_token_logits[indices_to_remove] = -float('inf')
# Sample from the filtered distribution
probs = F.softmax(next_token_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
else:
# Greedy decoding
next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
# Append the new token
generated = torch.cat([generated, next_token], dim=1)
# Check for EOS token
if eos_token_id is not None and (next_token == eos_token_id).all():
break
return generated
# ✅ AUTOCLASS REGISTRATION - Required for Hub compatibility
# Register the configuration and model for AutoClass support
AutoConfig.register("unified_model", UnifiedModelConfig)
AutoModel.register(UnifiedModelConfig, UnifiedModel)
AutoModelForCausalLM.register(UnifiedModelConfig, UnifiedModel)
print("🚀 ✅ AUTOCLASS REGISTRATION COMPLETE:")
print(" • AutoConfig.register('unified_model', UnifiedModelConfig)")
print(" • AutoModel.register(UnifiedModelConfig, UnifiedModel)")
print(" • AutoModelForCausalLM.register(UnifiedModelConfig, UnifiedModel)")
print(" • Users can now load with: AutoModel.from_pretrained('your-repo', trust_remote_code=True)") |