File size: 9,159 Bytes
62caf8e | 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 | """
Train AR Control Model with identical hyperparameters to MDLM-BPE v3.
Model: 199M params, d_model=1024, 15 layers, 16 heads, seq_len=128
Data: SAME Ultra-FineWeb 1M docs, SAME tokenizer, SAME tokenization
Training: 3 epochs, bf16, gradient accumulation, identical scheduler
This produces a controlled comparison: same data, same compute budget,
same parameter count — only the architecture differs (AR vs MDLM).
"""
import json
import sys
import time
import math
import argparse
import numpy as np
from pathlib import Path
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
REPO = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(REPO / "src"))
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
CHECKPOINT_DIR = REPO / "checkpoints"
RESULTS_DIR = REPO / "results"
DATA_DIR = REPO / "data"
from ar_control import (
ARConfig, ARControlModel, BPETokenizer,
ar_loss, sample_ar, measure_perplexity,
)
def prepare_data(seq_len=128):
"""Tokenize 1M docs and pack into 128-token sequences.
Reuses the SAME tokenized data as MDLM v3 (train_tokens_v3_128.npy)
to guarantee identical training data.
"""
output_file = DATA_DIR / f"train_tokens_v3_{seq_len}.npy"
if output_file.exists():
arr = np.load(output_file, mmap_mode='r')
print(f" Cached: {output_file} ({len(arr):,} seqs, mmap)")
return np.array(arr), len(arr)
raise FileNotFoundError(
f"Training data not found: {output_file}\n"
f"Run scripts/train.py first to tokenize data for MDLM v3."
)
def train_ar(epochs=3, batch_size=32, lr=3e-4, seq_len=128,
warmup_ratio=0.05, eval_every=500, gradient_accumulation=4,
n_layers=15):
"""Train AR Control Model.
Hyperparameters intentionally identical to MDLM v3 train.py:
- Same lr, same scheduler (OneCycleLR), same batch size
- Same gradient accumulation, same warmup ratio
- Same seq_len, same data file
"""
print("=" * 70)
print("TRAINING AR CONTROL MODEL (199M PARAMS)")
print("=" * 70)
# Data — IDENTICAL to MDLM training
print("Loading data...")
tokens, n_seqs = prepare_data(seq_len=seq_len)
print(f" Total sequences: {n_seqs:,}")
tokens_int16 = np.array(tokens, dtype=np.int16)
del tokens
tokens_tensor = torch.from_numpy(tokens_int16).long()
dataset = TensorDataset(tokens_tensor)
loader = DataLoader(dataset, batch_size=batch_size, shuffle=True,
drop_last=True, num_workers=2, pin_memory=True)
# Model
tokenizer = BPETokenizer()
config = ARConfig(
vocab_size=tokenizer.vocab_size,
d_model=1024,
n_heads=16,
n_layers=n_layers,
max_seq_len=256,
)
model = ARControlModel(config, pad_id=tokenizer.pad_id).to(DEVICE)
n_params = sum(p.numel() for p in model.parameters())
print(f" Model: {n_params:,} ({n_params/1e6:.1f}M)")
print(f" Layers: {n_layers} (MDLM v3 has 10 + AdaLN)")
print(f" Data: {len(tokens_tensor):,} seqs × {seq_len} tokens = {len(tokens_tensor)*seq_len:,} tokens")
print(f" Epochs: {epochs}, Batch: {batch_size}, Accum: {gradient_accumulation}")
print(f" Effective batch: {batch_size * gradient_accumulation}")
print(f" Opt steps/epoch: {len(loader)//gradient_accumulation:,}")
print()
# Optimizer — identical to MDLM v3
optimizer = torch.optim.AdamW(
model.parameters(), lr=lr, weight_decay=0.01, betas=(0.9, 0.95),
)
optimizer_steps_per_epoch = len(loader) // gradient_accumulation
total_optimizer_steps = optimizer_steps_per_epoch * epochs
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer, max_lr=lr, total_steps=total_optimizer_steps,
pct_start=warmup_ratio,
)
# Training loop — mirrors MDLM v3 train.py structure
model.train()
micro_step = 0
opt_step = 0
best_eval = float('inf')
losses = []
start = time.time()
accum_loss = 0.0
# Hold out 1000 sequences for perplexity eval
n_holdout = 1000
holdout = tokens_tensor[:n_holdout].clone()
for epoch in range(epochs):
ep_loss = 0
ep_count = 0
for batch in loader:
micro_step += 1
batch_tokens = batch[0].to(DEVICE, non_blocking=True)
with torch.amp.autocast('cuda', dtype=torch.bfloat16):
loss = ar_loss(model, batch_tokens, pad_id=tokenizer.pad_id)
loss = loss / gradient_accumulation
loss.backward()
accum_loss += loss.item()
if micro_step % gradient_accumulation == 0:
opt_step += 1
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
if opt_step <= total_optimizer_steps:
scheduler.step()
optimizer.zero_grad(set_to_none=True)
ep_loss += accum_loss
ep_count += 1
losses.append(accum_loss)
accum_loss = 0.0
if opt_step % 100 == 0:
elapsed = time.time() - start
tps = micro_step * batch_size * seq_len / elapsed
lr_cur = optimizer.param_groups[0]['lr']
print(f" [E{epoch+1} O{opt_step:,}] loss={losses[-1]:.4f} "
f"avg={ep_loss/ep_count:.4f} lr={lr_cur:.2e} "
f"{opt_step/elapsed:.1f} opt/s {tps:,.0f} tok/s")
if opt_step % eval_every == 0:
model.eval()
eval_loss = quick_eval(model, loader, tokenizer.pad_id)
model.train()
ppl = math.exp(min(eval_loss, 15))
print(f" → eval_loss={eval_loss:.4f} PPL={ppl:.1f}")
if eval_loss < best_eval:
best_eval = eval_loss
torch.save({
"model_state": model.state_dict(),
"config": config.to_dict(),
"step": opt_step,
"eval_loss": eval_loss,
"ppl": ppl,
}, CHECKPOINT_DIR / "ar_control_best.pt")
samples = sample_ar(
model, tokenizer, max_new_tokens=64,
temperature=0.7, device=DEVICE,
)
print(f" → Best sample (AR):")
print(f" {samples.strip()[:150]}")
# Final eval with proper perplexity on holdout
model.eval()
final_loss, final_ppl = measure_perplexity(model, holdout, batch_size=32)
elapsed = time.time() - start
print(f"\n{'='*70}")
print(f"TRAINING COMPLETE — {elapsed:.1f}s ({elapsed/60:.1f} min)")
print(f" Optimizer steps: {opt_step}")
print(f" Best eval loss (training): {best_eval:.4f} (PPL={math.exp(min(best_eval,15)):.1f})")
print(f" Final holdout PPL: {final_ppl:.1f} (loss={final_loss:.4f})")
torch.save({
"model_state": model.state_dict(),
"config": config.to_dict(),
"step": opt_step,
"losses": losses[-1000:],
"holdout_ppl": final_ppl,
"holdout_loss": final_loss,
}, CHECKPOINT_DIR / "ar_control_final.pt")
results = {
"optimizer_steps": opt_step,
"time": elapsed,
"best_eval_loss": best_eval,
"best_ppl": math.exp(min(best_eval, 15)),
"final_holdout_ppl": final_ppl,
"final_holdout_loss": final_loss,
"final_loss": losses[-1],
"tokens_trained": micro_step * batch_size * seq_len,
"n_params": n_params,
"n_layers": n_layers,
"model_type": "autoregressive_control",
}
with open(RESULTS_DIR / "ar_control_training.json", "w") as f:
json.dump(results, f, indent=2)
return model
def quick_eval(model, loader, pad_id, n_batches=20):
"""Quick eval on training data batches (same as MDLM's quick_eval)."""
model.eval()
losses = []
with torch.no_grad():
for i, batch in enumerate(loader):
if i >= n_batches:
break
tokens = batch[0].to(DEVICE)
with torch.amp.autocast('cuda', dtype=torch.bfloat16):
loss = ar_loss(model, tokens, pad_id=pad_id)
losses.append(loss.item())
return np.mean(losses)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--epochs", type=int, default=3)
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument("--lr", type=float, default=3e-4)
parser.add_argument("--seq-len", type=int, default=128)
parser.add_argument("--accum", type=int, default=4)
parser.add_argument("--layers", type=int, default=15)
args = parser.parse_args()
train_ar(epochs=args.epochs, batch_size=args.batch_size,
lr=args.lr, seq_len=args.seq_len, gradient_accumulation=args.accum,
n_layers=args.layers)
|