File size: 18,419 Bytes
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 | """
AAM Diffusion LLM β Complete Model
Combines the Diffusion Transformer, Graph Encoder, and Noise Scheduler
into a single, unified model for training and inference.
This is the "body" of AAM β the specialized sentence composer that
takes graph conditioning as input and produces coherent narratives
through iterative denoising.
Architecture:
ββββββββββββββββββββββββββββββββββββββββββββββββββββ
β AAM Diffusion Model (The Body) β
β β
β Input: β
β - Token IDs (text) β
β - Graph conditioning (evidence, compositions, β
β confidence, anomalies, reasoning chains) β
β β
β Training Process: β
β 1. Tokenize text β embeddings β
β 2. Sample random timestep t β
β 3. Add noise: x_t = schedule.add_noise(x_0, t) β
β 4. Encode graph conditioning β
β 5. Predict noise: eps = transformer(x_t, t, c) β
β 6. Compute loss: L = MSE(eps, eps_target) β
β β
β Inference Process: β
β 1. Start from pure noise x_T β
β 2. Encode graph conditioning β
β 3. For t = T, T-1, ..., 1: β
β a. Predict noise: eps = transformer(x_t, t) β
β b. Denoise: x_{t-1} = schedule.step(eps) β
β 4. Decode final x_0 β text tokens β
β 5. Detokenize β natural language narrative β
β β
β Key Constraint: β
β The model CANNOT generate information not β
β present in the graph conditioning. It can only β
β ARRANGE what the graph knows into sentences. β
β β
β Analogi: Jin Soun (mind/graph) + tubuhnya β
β (this model). Tubuhnya hanya bisa mengucapkan β
β apa yang dipikirkannya β tidak bisa mengarang. β
ββββββββββββββββββββββββββββββββββββββββββββββββββββ
Analogi: Ini adalah seluruh "tubuh" Jin Soun β bukan hanya
ototnya (transformer), tapi juga sistem saraf (graph encoder)
dan kemampuan untuk memperbaiki diri (diffusion denoising).
"""
from __future__ import annotations
import logging
from typing import Optional
import torch
import torch.nn as nn
from diffusion_llm.config.model_config import AamDiffusionConfig
from diffusion_llm.model.noise_scheduler import NoiseScheduler
from diffusion_llm.model.graph_encoder import GraphConditioningEncoder
from diffusion_llm.model.diffusion_transformer import DiffusionTransformer
logger = logging.getLogger(__name__)
class AamDiffusionModel(nn.Module):
"""Complete AAM Diffusion LLM model.
Combines:
- DiffusionTransformer: Core denoising network
- GraphConditioningEncoder: Encodes graph structure for conditioning
- NoiseScheduler: Manages the diffusion process
This model is designed to be trained on GraphβNarrative pairs,
where the graph data comes from the RSVS Knowledge Graph and
the narrative is the target natural language output.
Args:
config: AamDiffusionConfig with all hyperparameters.
"""
def __init__(self, config: AamDiffusionConfig):
super().__init__()
self.config = config
# Core components
self.noise_scheduler = NoiseScheduler(
n_timesteps=config.diffusion.n_timesteps,
schedule_type=config.diffusion.schedule_type,
beta_start=config.diffusion.beta_start,
beta_end=config.diffusion.beta_end,
prediction_type=config.diffusion.prediction_type,
)
self.graph_encoder = GraphConditioningEncoder(
config=config.graph_encoder,
vocab_size=config.model.vocab_size,
)
# Align graph encoder output dim with transformer's d_model
self.graph_encoder.set_output_dim(config.model.d_model)
self.transformer = DiffusionTransformer(config.model)
# Token-to-embedding projection (shared with transformer)
# The transformer's token_embedding is used for both
# encoding input text and decoding output text
# Output head: project from d_model to vocab_size
self.lm_head = nn.Linear(
config.model.d_model, config.model.vocab_size, bias=False
)
# Tie weights between token embedding and LM head
# This is standard practice and reduces parameter count
self.lm_head.weight = self.transformer.token_embedding.weight
# EMA model (for inference, updated during training)
self._ema_model: Optional[AamDiffusionModel] = None
self._ema_decay = config.training.ema_decay
logger.info(
"AamDiffusionModel initialized: %s params, %s",
self._format_params(self.get_num_params()),
config.model_name,
)
def forward(
self,
token_ids: torch.Tensor,
timestep: torch.Tensor,
evidence_ids: Optional[torch.Tensor] = None,
evidence_confidence: Optional[torch.Tensor] = None,
evidence_timestamps: Optional[torch.Tensor] = None,
composition_ids: Optional[torch.Tensor] = None,
composition_confidence: Optional[torch.Tensor] = None,
anomaly_ids: Optional[torch.Tensor] = None,
anomaly_confidence: Optional[torch.Tensor] = None,
anomaly_timestamps: Optional[torch.Tensor] = None,
reasoning_ids: Optional[torch.Tensor] = None,
reasoning_confidence: Optional[torch.Tensor] = None,
source_trust: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Forward pass for training.
1. Get clean embeddings from token IDs
2. Add noise at the given timestep
3. Encode graph conditioning
4. Predict noise via transformer
5. Return predicted noise (loss computed externally)
Args:
token_ids: Clean text token IDs, shape (batch, seq_len).
timestep: Random timestep indices, shape (batch,).
evidence_ids: Evidence node token IDs.
evidence_confidence: Evidence confidence scores.
evidence_timestamps: Evidence timestamps.
composition_ids: Composition token IDs.
composition_confidence: Composition confidence.
anomaly_ids: Anomaly token IDs.
anomaly_confidence: Anomaly confidence.
anomaly_timestamps: Anomaly timestamps.
reasoning_ids: Reasoning step token IDs.
reasoning_confidence: Reasoning confidence.
source_trust: Source trust score.
Returns:
Predicted noise tensor of shape (batch, seq_len, d_model).
"""
# Step 1: Get clean embeddings (x_0)
x_0 = self.transformer.token_embedding(token_ids)
# Step 2: Add noise
noise = torch.randn_like(x_0)
x_t = self.noise_scheduler.add_noise(x_0, noise, timestep)
# Step 3: Encode graph conditioning
batch_size = token_ids.shape[0]
graph_cond = self.graph_encoder(
evidence_ids=evidence_ids,
evidence_confidence=evidence_confidence,
evidence_timestamps=evidence_timestamps,
composition_ids=composition_ids,
composition_confidence=composition_confidence,
anomaly_ids=anomaly_ids,
anomaly_confidence=anomaly_confidence,
anomaly_timestamps=anomaly_timestamps,
reasoning_ids=reasoning_ids,
reasoning_confidence=reasoning_confidence,
source_trust=source_trust,
batch_size=batch_size,
)
# Extract cross-attention keys/values from graph conditioning
graph_keys = graph_cond.get("keys")
graph_values = graph_cond.get("values")
# Step 4: Predict noise via transformer
predicted = self.transformer(
x_t=x_t,
t=timestep,
graph_keys=graph_keys,
graph_values=graph_values,
)
return predicted, noise
def compute_loss(
self,
predicted: torch.Tensor,
target: torch.Tensor,
timestep: torch.Tensor,
) -> torch.Tensor:
"""Compute diffusion training loss.
Supports different loss types and weighting strategies.
Args:
predicted: Model output (predicted noise/x0/v).
target: Target (actual noise/x0/v).
timestep: Timestep indices for loss weighting.
Returns:
Scalar loss value.
"""
# Base loss
if self.config.diffusion.loss_type == "mse":
loss = nn.functional.mse_loss(predicted, target, reduction="none")
elif self.config.diffusion.loss_type == "mae":
loss = nn.functional.l1_loss(predicted, target, reduction="none")
elif self.config.diffusion.loss_type == "huber":
loss = nn.functional.smooth_l1_loss(predicted, target, reduction="none")
else:
raise ValueError(f"Unknown loss_type: {self.config.diffusion.loss_type}")
# Average over feature dimension
loss = loss.mean(dim=-1) # (batch, seq_len)
# Apply loss weighting
if self.config.diffusion.loss_weighting == "min_snr":
loss = self._apply_min_snr_weighting(loss, timestep)
elif self.config.diffusion.loss_weighting == "p2":
loss = self._apply_p2_weighting(loss, timestep)
# Average over sequence and batch
return loss.mean()
def _apply_min_snr_weighting(
self,
loss: torch.Tensor,
timestep: torch.Tensor,
gamma: float = 5.0,
) -> torch.Tensor:
"""Apply Min-SNR weighting strategy.
Weights the loss by min(SNR, gamma) / SNR, where
SNR = alpha_bar / (1 - alpha_bar).
This helps balance the loss across timesteps, preventing
high-noise steps from dominating.
Args:
loss: Unweighted loss.
timestep: Timestep indices.
gamma: SNR clipping value.
Returns:
Weighted loss.
"""
alpha_bar = self.noise_scheduler.alphas_cumprod.to(loss.device)
snr = alpha_bar[timestep] / (1 - alpha_bar[timestep] + 1e-8)
weight = torch.clamp(snr, max=gamma) / (snr + 1e-8)
# Expand weight to match loss shape
weight = weight.unsqueeze(-1).expand_as(loss)
return loss * weight
def _apply_p2_weighting(
self,
loss: torch.Tensor,
timestep: torch.Tensor,
) -> torch.Tensor:
"""Apply P2 weighting strategy.
weight = 1 / (SNR^gamma + k)
Args:
loss: Unweighted loss.
timestep: Timestep indices.
Returns:
Weighted loss.
"""
alpha_bar = self.noise_scheduler.alphas_cumprod.to(loss.device)
snr = alpha_bar[timestep] / (1 - alpha_bar[timestep] + 1e-8)
gamma = self.config.diffusion.p2_gamma
k = self.config.diffusion.p2_k
weight = 1.0 / (snr ** gamma + k)
weight = weight.unsqueeze(-1).expand_as(loss)
return loss * weight
@torch.no_grad()
def sample(
self,
graph_cond: dict[str, torch.Tensor],
n_steps: Optional[int] = None,
method: str = "ddim",
shape: Optional[tuple[int, ...]] = None,
device: Optional[torch.device] = None,
) -> torch.Tensor:
"""Generate samples via iterative denoising.
This is the INFERENCE method β start from pure noise and
iteratively denoise to produce coherent text embeddings.
Args:
graph_cond: Graph conditioning dict from GraphConditioningEncoder.
n_steps: Number of denoising steps. Uses config if None.
method: Sampling method ('ddpm' or 'ddim').
shape: Shape of the output (batch, seq_len, d_model).
device: Device to generate on.
Returns:
Denoised embeddings of shape (batch, seq_len, d_model).
"""
if n_steps is None:
n_steps = self.config.diffusion.n_inference_steps
if device is None:
device = next(self.parameters()).device
if shape is None:
shape = (1, self.config.model.max_seq_len, self.config.model.d_model)
# Start from pure noise
x = torch.randn(shape, device=device)
# Get graph conditioning
graph_keys = graph_cond.get("keys")
graph_values = graph_cond.get("values")
if method == "ddpm":
# Full DDPM sampling
for t in reversed(range(self.config.diffusion.n_timesteps)):
t_tensor = torch.full((shape[0],), t, device=device, dtype=torch.long)
predicted = self.transformer(
x_t=x, t=t_tensor,
graph_keys=graph_keys,
graph_values=graph_values,
)
x = self.noise_scheduler.step_ddpm(predicted, x, t_tensor)
elif method == "ddim":
# Fast DDIM sampling
timesteps = self.noise_scheduler.get_timestep_schedule(n_steps)
for i in range(len(timesteps) - 1):
t = timesteps[i]
t_prev = timesteps[i + 1] if i + 1 < len(timesteps) else 0
t_tensor = torch.full((shape[0],), t, device=device, dtype=torch.long)
predicted = self.transformer(
x_t=x, t=t_tensor,
graph_keys=graph_keys,
graph_values=graph_values,
)
x = self.noise_scheduler.step_ddim(
predicted, x, t, t_prev,
eta=self.config.diffusion.eta_ddim,
)
return x
def embeddings_to_tokens(
self,
embeddings: torch.Tensor,
temperature: float = 1.0,
top_k: int = 50,
) -> torch.Tensor:
"""Convert continuous embeddings to discrete token IDs.
This is the final step of generation β project embeddings
to vocabulary logits and sample tokens.
Args:
embeddings: Denoised embeddings of shape (batch, seq_len, d_model).
temperature: Sampling temperature.
top_k: Top-k sampling cutoff.
Returns:
Token IDs of shape (batch, seq_len).
"""
logits = self.lm_head(embeddings) / temperature
# Top-k sampling
if top_k > 0:
top_k_values, top_k_indices = torch.topk(logits, top_k, dim=-1)
probs = torch.softmax(top_k_values, dim=-1)
sampled_indices = torch.multinomial(
probs.view(-1, top_k), 1
).view(logits.shape[0], logits.shape[1])
token_ids = top_k_indices.gather(
-1, sampled_indices.unsqueeze(-1)
).squeeze(-1)
else:
probs = torch.softmax(logits, dim=-1)
token_ids = torch.argmax(logits, dim=-1)
return token_ids
def get_num_params(self) -> int:
"""Get total number of parameters."""
return sum(p.numel() for p in self.parameters())
@staticmethod
def _format_params(n: int) -> str:
"""Format parameter count for display."""
if n >= 1e9:
return f"{n / 1e9:.1f}B"
elif n >= 1e6:
return f"{n / 1e6:.1f}M"
elif n >= 1e3:
return f"{n / 1e3:.1f}K"
return str(n)
def save(self, path: str) -> None:
"""Save model checkpoint.
Args:
path: Output file path.
"""
torch.save({
"model_state_dict": self.state_dict(),
"config": self.config.to_dict(),
}, path)
logger.info("Model saved to %s", path)
@classmethod
def load(cls, path: str, device: str = "cpu") -> AamDiffusionModel:
"""Load model from checkpoint.
Args:
path: Checkpoint file path.
device: Device to load to.
Returns:
Loaded AamDiffusionModel.
"""
checkpoint = torch.load(path, map_location=device, weights_only=False)
config_dict = checkpoint.get("config", {})
if isinstance(config_dict, dict):
config = AamDiffusionConfig()
# Try to reconstruct config from dict
try:
from diffusion_llm.config.model_config import (
ModelConfig, DiffusionConfig, GraphEncoderConfig,
TokenizerConfig, TrainingConfig, InferenceConfig,
)
config = AamDiffusionConfig(
model=ModelConfig(**config_dict.get("model", {})),
diffusion=DiffusionConfig(**config_dict.get("diffusion", {})),
graph_encoder=GraphEncoderConfig(**config_dict.get("graph_encoder", {})),
tokenizer=TokenizerConfig(**config_dict.get("tokenizer", {})),
training=TrainingConfig(**config_dict.get("training", {})),
inference=InferenceConfig(**config_dict.get("inference", {})),
model_name=config_dict.get("model_name", "aam-diffusion-v0.1"),
output_dir=config_dict.get("output_dir", "./output"),
seed=config_dict.get("seed", 42),
)
except Exception:
logger.warning("Could not reconstruct config from checkpoint, using defaults")
else:
config = config_dict
model = cls(config)
model.load_state_dict(checkpoint["model_state_dict"])
model.to(device)
logger.info("Model loaded from %s", path)
return model
|