File size: 14,930 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 | """
AAM Diffusion LLM — Trainer
Training loop for the AAM Diffusion Model.
Handles:
- Training loop with gradient accumulation
- Learning rate scheduling with warmup
- Mixed precision training (AMP)
- EMA model updates
- Checkpoint saving/loading
- Logging to console and Weights & Biases
- Evaluation on validation set
Analogi: Seperti latihan fisik Jin Soun — berulang-ulang,
bertahap meningkat intensitas, dengan instruktur yang
mengawasi dan memberi koreksi.
"""
from __future__ import annotations
import json
import logging
import math
import time
from pathlib import Path
from typing import Optional
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from diffusion_llm.config.model_config import AamDiffusionConfig
from diffusion_llm.model.aam_diffusion_model import AamDiffusionModel
from diffusion_llm.training.dataset import GraphNarrativeDataset, collate_fn
from diffusion_llm.tokenizer.aam_tokenizer import AamTokenizer
from diffusion_llm.training.losses import DiffusionLoss
logger = logging.getLogger(__name__)
class AamTrainer:
"""Trainer for the AAM Diffusion Model.
Args:
config: AamDiffusionConfig with training settings.
model: AamDiffusionModel instance.
tokenizer: AamTokenizer instance.
train_dataset: Training dataset.
val_dataset: Optional validation dataset.
"""
def __init__(
self,
config: AamDiffusionConfig,
model: AamDiffusionModel,
tokenizer: AamTokenizer,
train_dataset: GraphNarrativeDataset,
val_dataset: Optional[GraphNarrativeDataset] = None,
):
self.config = config
self.model = model
self.tokenizer = tokenizer
self.train_dataset = train_dataset
self.val_dataset = val_dataset
# Device
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu"
)
self.model.to(self.device)
logger.info("Training on device: %s", self.device)
# Optimizer
self.optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=config.training.learning_rate,
weight_decay=config.training.weight_decay,
betas=(config.training.adam_beta1, config.training.adam_beta2),
eps=config.training.adam_eps,
)
# Loss function
self.loss_fn = DiffusionLoss(config.diffusion)
# Data loaders
self.train_loader = DataLoader(
train_dataset,
batch_size=config.training.batch_size,
shuffle=True,
num_workers=config.training.num_workers,
collate_fn=collate_fn,
pin_memory=True,
)
if val_dataset:
self.val_loader = DataLoader(
val_dataset,
batch_size=config.training.batch_size,
shuffle=False,
num_workers=config.training.num_workers,
collate_fn=collate_fn,
pin_memory=True,
)
else:
self.val_loader = None
# LR scheduler
self.scheduler = self._create_lr_scheduler()
# AMP
self.scaler = None
if config.training.use_amp:
dtype = torch.bfloat16 if config.training.amp_dtype == "bf16" else torch.float16
self.scaler = torch.amp.GradScaler("cuda", enabled=(dtype == torch.float16))
# EMA
self.ema_model = None
if config.training.use_ema:
self.ema_model = self._create_ema_model()
# State tracking
self.global_step = 0
self.best_val_loss = float("inf")
self.train_losses: list[float] = []
# Output directory
self.output_dir = Path(config.output_dir)
self.output_dir.mkdir(parents=True, exist_ok=True)
# Seed
torch.manual_seed(config.seed)
def _create_lr_scheduler(self):
"""Create learning rate scheduler with warmup."""
total_steps = self.config.training.max_steps
warmup_steps = self.config.training.warmup_steps
def lr_lambda(step: int) -> float:
if step < warmup_steps:
return step / max(warmup_steps, 1)
if self.config.training.lr_schedule == "cosine":
progress = (step - warmup_steps) / max(total_steps - warmup_steps, 1)
return 0.5 * (1.0 + math.cos(math.pi * progress))
elif self.config.training.lr_schedule == "linear":
progress = (step - warmup_steps) / max(total_steps - warmup_steps, 1)
return 1.0 - progress
else:
return 1.0
return torch.optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda)
def _create_ema_model(self) -> AamDiffusionModel:
"""Create EMA copy of the model."""
import copy
ema = copy.deepcopy(self.model)
for param in ema.parameters():
param.requires_grad = False
return ema
@torch.no_grad()
def _update_ema(self) -> None:
"""Update EMA model weights."""
if self.ema_model is None:
return
decay = self.config.training.ema_decay
for ema_param, model_param in zip(
self.ema_model.parameters(), self.model.parameters()
):
ema_param.data.mul_(decay).add_(model_param.data, alpha=1 - decay)
def train(self) -> None:
"""Main training loop.
Runs for max_steps or max_epochs, whichever comes first.
Saves checkpoints and runs evaluation periodically.
"""
logger.info("Starting training...")
logger.info(" Max steps: %d", self.config.training.max_steps)
logger.info(" Batch size: %d", self.config.training.batch_size)
logger.info(" Gradient accumulation: %d", self.config.training.gradient_accumulation_steps)
logger.info(" Effective batch size: %d",
self.config.training.batch_size * self.config.training.gradient_accumulation_steps)
start_time = time.time()
epoch = 0
while self.global_step < self.config.training.max_steps:
epoch += 1
if epoch > self.config.training.max_epochs:
break
logger.info("=== Epoch %d ===", epoch)
epoch_loss = 0.0
n_batches = 0
for batch_idx, batch in enumerate(self.train_loader):
loss = self._train_step(batch)
epoch_loss += loss
n_batches += 1
# Logging
if self.global_step % self.config.training.log_every_steps == 0:
avg_loss = epoch_loss / max(n_batches, 1)
lr = self.optimizer.param_groups[0]["lr"]
elapsed = time.time() - start_time
steps_per_sec = self.global_step / max(elapsed, 1)
logger.info(
"Step %d | Loss: %.4f | LR: %.2e | Speed: %.1f steps/s",
self.global_step, loss, lr, steps_per_sec,
)
# Evaluation
if (self.global_step % self.config.training.eval_every_steps == 0
and self.val_loader is not None):
val_loss = self.evaluate()
logger.info("Validation loss: %.4f", val_loss)
if val_loss < self.best_val_loss:
self.best_val_loss = val_loss
self._save_checkpoint("best.pt")
# Checkpoint
if self.global_step % self.config.training.save_every_steps == 0:
self._save_checkpoint(f"step_{self.global_step}.pt")
# Stop condition
if self.global_step >= self.config.training.max_steps:
break
avg_epoch_loss = epoch_loss / max(n_batches, 1)
logger.info("Epoch %d complete. Average loss: %.4f", epoch, avg_epoch_loss)
# Final save
self._save_checkpoint("final.pt")
elapsed = time.time() - start_time
logger.info(
"Training complete! %d steps in %.1f hours",
self.global_step, elapsed / 3600,
)
def _train_step(self, batch: dict[str, torch.Tensor]) -> float:
"""Single training step.
Args:
batch: Batch of training data.
Returns:
Loss value for this step.
"""
self.model.train()
# Move batch to device
batch = {k: v.to(self.device) if isinstance(v, torch.Tensor) else v
for k, v in batch.items()}
# Sample random timesteps
batch_size = batch["token_ids"].shape[0]
t = torch.randint(
0, self.config.diffusion.n_timesteps,
(batch_size,), device=self.device,
)
# Forward pass
if self.scaler is not None:
with torch.amp.autocast("cuda", enabled=True):
predicted, target = self.model(
token_ids=batch["token_ids"],
timestep=t,
evidence_ids=batch.get("evidence_ids"),
evidence_confidence=batch.get("evidence_confidence"),
anomaly_ids=batch.get("anomaly_ids"),
anomaly_confidence=batch.get("anomaly_confidence"),
reasoning_ids=batch.get("reasoning_ids"),
reasoning_confidence=batch.get("reasoning_confidence"),
source_trust=batch.get("source_trust"),
)
loss = self.model.compute_loss(predicted, target, t)
loss = loss / self.config.training.gradient_accumulation_steps
else:
predicted, target = self.model(
token_ids=batch["token_ids"],
timestep=t,
evidence_ids=batch.get("evidence_ids"),
evidence_confidence=batch.get("evidence_confidence"),
anomaly_ids=batch.get("anomaly_ids"),
anomaly_confidence=batch.get("anomaly_confidence"),
reasoning_ids=batch.get("reasoning_ids"),
reasoning_confidence=batch.get("reasoning_confidence"),
source_trust=batch.get("source_trust"),
)
loss = self.model.compute_loss(predicted, target, t)
loss = loss / self.config.training.gradient_accumulation_steps
# Backward pass
if self.scaler is not None:
self.scaler.scale(loss).backward()
else:
loss.backward()
# Gradient accumulation
if (self.global_step + 1) % self.config.training.gradient_accumulation_steps == 0:
# Gradient clipping
if self.scaler is not None:
self.scaler.unscale_(self.optimizer)
torch.nn.utils.clip_grad_norm_(
self.model.parameters(),
self.config.training.grad_clip_norm,
)
# Optimizer step
if self.scaler is not None:
self.scaler.step(self.optimizer)
self.scaler.update()
else:
self.optimizer.step()
# LR schedule
self.scheduler.step()
# Zero gradients
self.optimizer.zero_grad()
# EMA update
self._update_ema()
self.global_step += 1
self.train_losses.append(loss.item())
return loss.item()
@torch.no_grad()
def evaluate(self) -> float:
"""Evaluate on validation set.
Returns:
Average validation loss.
"""
if self.val_loader is None:
return float("inf")
self.model.eval()
total_loss = 0.0
n_batches = 0
for batch in self.val_loader:
batch = {k: v.to(self.device) if isinstance(v, torch.Tensor) else v
for k, v in batch.items()}
batch_size = batch["token_ids"].shape[0]
t = torch.randint(
0, self.config.diffusion.n_timesteps,
(batch_size,), device=self.device,
)
predicted, target = self.model(
token_ids=batch["token_ids"],
timestep=t,
evidence_ids=batch.get("evidence_ids"),
evidence_confidence=batch.get("evidence_confidence"),
anomaly_ids=batch.get("anomaly_ids"),
anomaly_confidence=batch.get("anomaly_confidence"),
reasoning_ids=batch.get("reasoning_ids"),
reasoning_confidence=batch.get("reasoning_confidence"),
source_trust=batch.get("source_trust"),
)
loss = self.model.compute_loss(predicted, target, t)
total_loss += loss.item()
n_batches += 1
avg_loss = total_loss / max(n_batches, 1)
self.model.train()
return avg_loss
def _save_checkpoint(self, filename: str) -> None:
"""Save training checkpoint.
Args:
filename: Checkpoint filename.
"""
path = self.output_dir / filename
checkpoint = {
"model_state_dict": self.model.state_dict(),
"optimizer_state_dict": self.optimizer.state_dict(),
"scheduler_state_dict": self.scheduler.state_dict(),
"global_step": self.global_step,
"best_val_loss": self.best_val_loss,
"config": self.config.to_dict(),
}
if self.ema_model is not None:
checkpoint["ema_state_dict"] = self.ema_model.state_dict()
torch.save(checkpoint, path)
logger.info("Checkpoint saved: %s", path)
# Clean up old checkpoints
self._cleanup_checkpoints()
def _cleanup_checkpoints(self) -> None:
"""Remove old checkpoints, keeping only the last N."""
keep_n = self.config.training.keep_last_n_checkpoints
checkpoints = sorted(self.output_dir.glob("step_*.pt"))
while len(checkpoints) > keep_n:
oldest = checkpoints.pop(0)
oldest.unlink()
logger.info("Removed old checkpoint: %s", oldest)
def load_checkpoint(self, path: str) -> None:
"""Load from checkpoint.
Args:
path: Checkpoint file path.
"""
checkpoint = torch.load(path, map_location=self.device)
self.model.load_state_dict(checkpoint["model_state_dict"])
self.optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
self.scheduler.load_state_dict(checkpoint["scheduler_state_dict"])
self.global_step = checkpoint["global_step"]
self.best_val_loss = checkpoint.get("best_val_loss", float("inf"))
logger.info("Loaded checkpoint from step %d", self.global_step)
|