File size: 22,726 Bytes
2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 087c85d 2d7e335 | 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 | """
AAM Diffusion LLM β Diffusion Transformer (Denoiser)
The core denoising network. Takes noisy text embeddings and graph
conditioning, and predicts the noise (or clean data) at each
diffusion timestep.
Architecture:
Input: Noisy embeddings x_t + timestep t + graph conditioning
Output: Predicted noise epsilon (or x_0 or v)
The transformer uses:
- Self-attention over the text sequence
- Cross-attention to graph conditioning (evidence, anomalies, etc.)
- Timestep embedding (sinusoidal) injected via adaptive layer norm
- Optional flash attention for efficiency
- [v2.0] SwiGLU FFN (proven better in LLaMA/Mistral)
- [v2.0] RoPE via the dedicated rope.py module
- [v2.0] Evoformer integration points for layer recycling
This is the "brainstem" of the body β the core computation that
transforms noisy signals into coherent patterns.
Analogi: Seperti otot Jin Soun yang merespons sinyal dari otak β
model ini menerima "sinyal noise" dan "instruksi dari graph",
lalu mengubahnya menjadi gerakan yang koheren (kalimat).
"""
from __future__ import annotations
import math
from typing import Optional, List, Dict, Any
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusion_llm.config.model_config import ModelConfig, EvoformerConfig, MatryoshkaConfig
from diffusion_llm.model.rope import RotaryPositionEncoding
class SinusoidalTimestepEmbedding(nn.Module):
"""Sinusoidal embedding for diffusion timesteps.
Maps integer timesteps to d_model-dimensional vectors using
sinusoidal position encoding, similar to Transformers.
This allows the model to know "how noisy" the current input is,
which is essential for the denoising process.
"""
def __init__(self, d_model: int, max_period: int = 10000):
super().__init__()
self.d_model = d_model
self.max_period = max_period
# Two-layer MLP to project sinusoidal features
self.mlp = nn.Sequential(
nn.Linear(d_model, d_model * 4),
nn.GELU(),
nn.Linear(d_model * 4, d_model),
)
def forward(self, t: torch.Tensor) -> torch.Tensor:
"""Embed timesteps.
Args:
t: Timestep indices of shape (batch,).
Returns:
Timestep embeddings of shape (batch, d_model).
"""
device = t.device
half_dim = self.d_model // 2
emb = math.log(self.max_period) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, device=device, dtype=torch.float32) * -emb)
emb = t.float().unsqueeze(-1) * emb.unsqueeze(0)
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
if emb.shape[-1] < self.d_model:
emb = F.pad(emb, (0, self.d_model - emb.shape[-1]))
return self.mlp(emb)
class AdaptiveLayerNorm(nn.Module):
"""Adaptive Layer Normalization conditioned on timestep.
Instead of fixed scale/shift parameters, this layer norm
uses the timestep embedding to produce scale and shift:
y = (1 + scale(t)) * norm(x) + shift(t)
This allows the model to behave differently at different
noise levels β more "creative" at high noise, more
"precise" at low noise.
Analogi: Jin Soun menyesuaikan intensitas pikirannya
berdasarkan seberapa kabur situasinya β semakin kabur,
semakin "kreatif" pendekatannya.
"""
def __init__(self, d_model: int, eps: float = 1e-6):
super().__init__()
self.norm = nn.LayerNorm(d_model, elementwise_affine=False, eps=eps)
self.scale_proj = nn.Linear(d_model, d_model)
self.shift_proj = nn.Linear(d_model, d_model)
# Initialize shift to zero, scale to one
nn.init.zeros_(self.shift_proj.weight)
nn.init.zeros_(self.shift_proj.bias)
nn.init.ones_(self.scale_proj.weight)
nn.init.zeros_(self.scale_proj.bias)
def forward(
self,
x: torch.Tensor,
timestep_emb: torch.Tensor,
) -> torch.Tensor:
"""Apply adaptive layer norm.
Args:
x: Input tensor of shape (batch, seq_len, d_model).
timestep_emb: Timestep embedding of shape (batch, d_model).
Returns:
Normalized and modulated tensor.
"""
normalized = self.norm(x)
scale = (1 + self.scale_proj(timestep_emb)).unsqueeze(1)
shift = self.shift_proj(timestep_emb).unsqueeze(1)
return normalized * scale + shift
class TransformerBlock(nn.Module):
"""Single transformer block with self-attention, cross-attention, and FFN.
The block structure:
1. Adaptive Layer Norm + Self-Attention
2. Adaptive Layer Norm + Cross-Attention (to graph conditioning)
3. Adaptive Layer Norm + Feed-Forward Network
Each sub-layer has a residual connection.
v2.0 Changes:
- SwiGLU FFN replaces GELU FFN (proven better in LLaMA/Mistral)
- Optional Matryoshka elastic inference on the FFN
"""
def __init__(
self,
d_model: int,
n_heads: int,
d_ff: int,
dropout: float = 0.1,
norm_eps: float = 1e-6,
norm_type: str = "rmsnorm",
use_flash_attention: bool = True,
use_swiglu: bool = True,
matryoshka_config: Optional[MatryoshkaConfig] = None,
):
super().__init__()
self.d_model = d_model
self.n_heads = n_heads
self.use_swiglu = use_swiglu
self.matryoshka_config = matryoshka_config
# Norms
NormClass = nn.RMSNorm if norm_type == "rmsnorm" else nn.LayerNorm
# Self-attention
self.self_attn_norm = AdaptiveLayerNorm(d_model, eps=norm_eps)
self.self_attn = nn.MultiheadAttention(
embed_dim=d_model,
num_heads=n_heads,
dropout=dropout,
batch_first=True,
)
self.self_attn_dropout = nn.Dropout(dropout)
# Cross-attention (to graph conditioning)
self.cross_attn_norm = AdaptiveLayerNorm(d_model, eps=norm_eps)
self.cross_attn = nn.MultiheadAttention(
embed_dim=d_model,
num_heads=n_heads,
dropout=dropout,
batch_first=True,
kdim=d_model,
vdim=d_model,
)
self.cross_attn_dropout = nn.Dropout(dropout)
# Feed-forward β SwiGLU or legacy GELU
self.ff_norm = AdaptiveLayerNorm(d_model, eps=norm_eps)
if use_swiglu:
# SwiGLU FFN (proven better in LLaMA/Mistral)
self.gate_proj = nn.Linear(d_model, d_ff, bias=False)
self.up_proj = nn.Linear(d_model, d_ff, bias=False)
self.down_proj = nn.Linear(d_ff, d_model, bias=False)
self.ff_dropout = nn.Dropout(dropout)
else:
# Legacy GELU FFN (backward compatible)
self.ff = nn.Sequential(
nn.Linear(d_model, d_ff),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(d_ff, d_model),
nn.Dropout(dropout),
)
# Matryoshka elastic inference (optional)
if matryoshka_config is not None and use_swiglu:
self._matryoshka_d_ff = d_ff
self._matryoshka_factors = sorted(matryoshka_config.granularity_factors)
if matryoshka_config.use_adaptive:
self.size_selector = nn.Sequential(
nn.Linear(d_model, d_model // 8, bias=False),
nn.SiLU(),
nn.Linear(d_model // 8, 1, bias=False),
nn.Sigmoid(),
)
else:
self._matryoshka_d_ff = None
self._matryoshka_factors = None
# Layer scales (optional, helps with deep networks)
self.self_attn_scale = nn.Parameter(torch.ones(1) * 0.1)
self.cross_attn_scale = nn.Parameter(torch.ones(1) * 0.1)
self.ff_scale = nn.Parameter(torch.ones(1) * 0.1)
def _select_matryoshka_factor(self, x: torch.Tensor) -> float:
"""Adaptive factor selection for Matryoshka inference."""
if not hasattr(self, "size_selector"):
return 1.0
score = self.size_selector(x.mean(dim=1, keepdim=False))
score_val = score.mean().item()
min_dist = float("inf")
best_factor = self._matryoshka_factors[-1]
for f in self._matryoshka_factors:
dist = abs(score_val - f)
if dist < min_dist:
min_dist = dist
best_factor = f
return best_factor
def forward(
self,
x: torch.Tensor,
timestep_emb: torch.Tensor,
graph_keys: Optional[torch.Tensor] = None,
graph_values: Optional[torch.Tensor] = None,
causal_mask: Optional[torch.Tensor] = None,
granularity_factor: Optional[float] = None,
) -> torch.Tensor:
"""Forward pass.
Args:
x: Input sequence of shape (batch, seq_len, d_model).
timestep_emb: Timestep embedding of shape (batch, d_model).
graph_keys: Graph conditioning keys for cross-attention,
shape (batch, n_graph_nodes, d_model).
graph_values: Graph conditioning values for cross-attention,
shape (batch, n_graph_nodes, d_model).
causal_mask: Optional causal mask for self-attention.
granularity_factor: Optional Matryoshka granularity factor
for elastic inference (1.0 = full size).
Returns:
Output sequence of shape (batch, seq_len, d_model).
"""
# 1. Self-attention with adaptive layer norm
normed = self.self_attn_norm(x, timestep_emb)
attn_out, _ = self.self_attn(
normed, normed, normed,
attn_mask=causal_mask,
need_weights=False,
)
x = x + self.self_attn_scale * self.self_attn_dropout(attn_out)
# 2. Cross-attention to graph conditioning (if available)
if graph_keys is not None and graph_values is not None:
normed = self.cross_attn_norm(x, timestep_emb)
cross_out, _ = self.cross_attn(
normed, graph_keys, graph_values,
need_weights=False,
)
x = x + self.cross_attn_scale * self.cross_attn_dropout(cross_out)
# 3. Feed-forward with adaptive layer norm
normed = self.ff_norm(x, timestep_emb)
if self.use_swiglu:
# Determine Matryoshka factor
factor = granularity_factor
if factor is None and self._matryoshka_factors is not None:
factor = self._select_matryoshka_factor(normed)
elif factor is None:
factor = 1.0
# Clamp factor
if self._matryoshka_factors is not None:
factor = min(max(factor, min(self._matryoshka_factors)), 1.0)
else:
factor = 1.0
d_ff_active = max(1, int(self._matryoshka_d_ff * factor)) if self._matryoshka_d_ff else self.gate_proj.out_features
if factor >= 1.0 or self._matryoshka_d_ff is None:
# Full-size SwiGLU
gate = F.silu(self.gate_proj(normed))
up = self.up_proj(normed)
ff_out = self.down_proj(gate * up)
else:
# Matryoshka partial SwiGLU
d_ff_active = max(1, int(self._matryoshka_d_ff * factor))
gate = F.silu(F.linear(normed, self.gate_proj.weight[:d_ff_active, :]))
up = F.linear(normed, self.up_proj.weight[:d_ff_active, :])
ff_out = F.linear(gate * up, self.down_proj.weight[:, :d_ff_active])
ff_out = self.ff_dropout(ff_out)
else:
# Legacy GELU FFN
ff_out = self.ff(normed)
x = x + self.ff_scale * ff_out
return x
class DiffusionTransformer(nn.Module):
"""Diffusion Transformer β the core denoising network for AAM.
This transformer takes:
- Noisy text embeddings (x_t)
- Diffusion timestep (t)
- Graph conditioning (evidence, anomalies, reasoning chains)
And predicts the noise that was added (or the clean data,
depending on prediction_type).
Architecture Overview:
ββββββββββββββββββββββββββββββββββββββββββββββββββ
β Input Embedding: x_t (noisy) β embedding β
β + Positional Encoding (RoPE or learned) β
β β
β N x TransformerBlock: β
β ββ AdaLN + Self-Attention β
β ββ AdaLN + Cross-Attention (to graph) β
β ββ AdaLN + SwiGLU FFN (Matryoshka) β
β β
β [Evoformer Layer Recycling β optional] β
β β
β Output Projection: β predicted noise β
ββββββββββββββββββββββββββββββββββββββββββββββββββ
Key Features:
- Adaptive Layer Norm: timestep-conditioned normalization
- Cross-Attention: graph conditioning guides generation
- Layer Scales: helps training deep networks
- RoPE: better length generalization than learned positions
- [v2.0] SwiGLU FFN: proven better than GELU in LLaMA/Mistral
- [v2.0] Matryoshka: elastic inference at multiple sizes
- [v2.0] Evoformer: layer recycling for iterative refinement
Args:
config: ModelConfig with architecture hyperparameters.
evoformer_config: Optional EvoformerConfig for layer recycling.
matryoshka_config: Optional MatryoshkaConfig for elastic inference.
use_swiglu: Whether to use SwiGLU FFN (default True for v2.0).
"""
def __init__(
self,
config: ModelConfig,
evoformer_config: Optional[EvoformerConfig] = None,
matryoshka_config: Optional[MatryoshkaConfig] = None,
use_swiglu: bool = True,
):
super().__init__()
self.config = config
self.evoformer_config = evoformer_config
self.matryoshka_config = matryoshka_config
self.use_swiglu = use_swiglu
# Input embedding (from token IDs to d_model)
self.token_embedding = nn.Embedding(config.vocab_size, config.d_model)
# Timestep embedding
self.timestep_embedding = SinusoidalTimestepEmbedding(config.d_model)
# Positional encoding
if config.pos_encoding_type == "learned":
self.position_embedding = nn.Embedding(
config.max_seq_len, config.d_model
)
self.rope = None
else:
# RoPE is applied inside attention (no separate embedding)
self.position_embedding = None
# v2.0: Create RotaryPositionEncoding module for explicit RoPE
self.rope = RotaryPositionEncoding(
d_model=config.d_model,
max_seq_len=config.max_seq_len,
)
# Transformer blocks
self.blocks = nn.ModuleList([
TransformerBlock(
d_model=config.d_model,
n_heads=config.n_heads,
d_ff=config.d_ff,
dropout=config.dropout,
norm_eps=config.norm_eps,
norm_type=config.norm_type,
use_flash_attention=config.use_flash_attention,
use_swiglu=use_swiglu,
matryoshka_config=matryoshka_config,
)
for _ in range(config.n_layers)
])
# Final norm
NormClass = nn.RMSNorm if config.norm_type == "rmsnorm" else nn.LayerNorm
self.final_norm = NormClass(config.d_model, eps=config.norm_eps)
# Output projection (predict noise/x0/v)
self.output_proj = nn.Linear(config.d_model, config.d_model)
# Evoformer integration β lazy import to avoid circular deps
self._evoformer_manager = None
if evoformer_config is not None:
self._init_evoformer(evoformer_config)
# Initialize weights
self.apply(self._init_weights)
def _init_evoformer(self, evoformer_config: EvoformerConfig) -> None:
"""Initialize the Evoformer manager for layer recycling."""
from diffusion_llm.model.evoformer import EvoformerManager
self._evoformer_manager = EvoformerManager(evoformer_config)
def _init_weights(self, module: nn.Module) -> None:
"""Initialize weights with Xavier/GPT-2 style."""
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=self.config.init_std)
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=self.config.init_std)
def forward(
self,
x_t: torch.Tensor,
t: torch.Tensor,
token_ids: Optional[torch.Tensor] = None,
graph_keys: Optional[torch.Tensor] = None,
graph_values: Optional[torch.Tensor] = None,
granularity_factor: Optional[float] = None,
return_hidden_states: bool = False,
) -> torch.Tensor | tuple[torch.Tensor, List[torch.Tensor]]:
"""Forward pass: predict noise given noisy input and timestep.
Args:
x_t: Noisy text embeddings of shape (batch, seq_len, d_model).
If None, token_ids must be provided.
t: Timestep indices of shape (batch,).
token_ids: Token IDs of shape (batch, seq_len).
Used to create embeddings if x_t is not provided directly.
In training, x_t comes from the noise scheduler.
graph_keys: Graph conditioning keys for cross-attention,
shape (batch, n_graph_nodes, d_model).
graph_values: Graph conditioning values for cross-attention,
shape (batch, n_graph_nodes, d_model).
granularity_factor: Optional Matryoshka granularity factor
for elastic inference (1.0 = full size).
return_hidden_states: If True, also return per-layer hidden
states for Evoformer layer recycling.
Returns:
Predicted noise of shape (batch, seq_len, d_model).
If return_hidden_states is True, also returns a list of
per-layer hidden state tensors.
"""
# Get input embeddings
if x_t is None and token_ids is not None:
# Create embeddings from token IDs (used for initial x_0)
h = self.token_embedding(token_ids)
elif x_t is not None:
h = x_t
else:
raise ValueError("Either x_t or token_ids must be provided")
# Add positional encoding (learned positions only; RoPE is applied in attention)
if self.position_embedding is not None:
seq_len = h.shape[1]
positions = torch.arange(seq_len, device=h.device).unsqueeze(0)
h = h + self.position_embedding(positions)
# Embed timestep
t_emb = self.timestep_embedding(t)
# Pass through transformer blocks, collecting hidden states for Evoformer
hidden_states: List[torch.Tensor] = []
for block in self.blocks:
h = block(
h,
timestep_emb=t_emb,
graph_keys=graph_keys,
graph_values=graph_values,
granularity_factor=granularity_factor,
)
if return_hidden_states or self._evoformer_manager is not None:
hidden_states.append(h)
# Evoformer layer recycling (if enabled)
if self._evoformer_manager is not None and len(hidden_states) > 1:
hidden_states = self._evoformer_manager.recycle_layers(hidden_states)
# Use the last revised hidden state as the output
h = hidden_states[-1]
# Final norm and projection
h = self.final_norm(h)
output = self.output_proj(h)
if return_hidden_states:
return output, hidden_states
return output
def apply_evoformer_token_update(self, x: torch.Tensor) -> torch.Tensor:
"""Apply Evoformer bidirectional token update (Level 2).
Can be called externally as part of an Evoformer recycling loop.
Args:
x: Hidden state tensor of shape (batch, seq_len, d_model).
Returns:
Revised hidden state tensor.
"""
if self._evoformer_manager is not None:
return self._evoformer_manager.bidirectional_token_update(x)
return x
def apply_evoformer_decoder_feedback(
self,
hidden_state: torch.Tensor,
decoder_output: torch.Tensor,
) -> torch.Tensor:
"""Apply Evoformer decoder-predict feedback (Level 3).
Can be called externally during anchored decoder refinement.
Args:
hidden_state: Hidden state tensor of shape (batch, seq_len, d_model).
decoder_output: Decoder output tensor of shape (batch, seq_len, d_model).
Returns:
Revised hidden state tensor.
"""
if self._evoformer_manager is not None:
return self._evoformer_manager.apply_decoder_feedback(hidden_state, decoder_output)
return hidden_state
def apply_evoformer_prediction_recycling(
self,
hidden_states: torch.Tensor,
prediction_logits: torch.Tensor,
) -> torch.Tensor:
"""Apply Evoformer prediction-context recycling (Level 4).
Can be called externally to refine graph understanding from
predicted output.
Args:
hidden_states: Hidden states of shape (batch, seq_len, d_model).
prediction_logits: Prediction logits of shape (batch, seq_len, d_model).
Returns:
Revised hidden state tensor.
"""
if self._evoformer_manager is not None:
return self._evoformer_manager.apply_prediction_recycling(hidden_states, prediction_logits)
return hidden_states
def get_num_params(self) -> int:
"""Get total number of parameters."""
return sum(p.numel() for p in self.parameters())
def get_num_trainable_params(self) -> int:
"""Get number of trainable parameters."""
return sum(p.numel() for p in self.parameters() if p.requires_grad)
|