File size: 11,824 Bytes
8174855 |
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 |
"""
Refactored training script for SupernovaModel
- AMP mixed precision training
- Resume from checkpoint (saves optimizer + scheduler state)
- TensorBoard logging
- Optional validation loop if --val-data-config provided
- DataLoader pin_memory and non_blocking transfers
- Save optimizer/scheduler/model/config/step
- CLI flags for common hyperparams
Usage:
python -m supernova.train_refactor --config path/to/config.json --data-config path/to/data.yaml
"""
import argparse
import math
import os
import time
from typing import Optional
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from transformers import get_cosine_schedule_with_warmup
from .config import ModelConfig
from .model import SupernovaModel
from .tokenizer import load_gpt2_tokenizer
from .data import load_sources_from_yaml, TokenChunkDataset
def compute_grad_norm(model: nn.Module) -> float:
total = 0.0
for p in model.parameters():
if p.grad is not None:
param_norm = p.grad.data.float().norm(2).item()
total += param_norm * param_norm
return math.sqrt(total)
class Trainer:
def __init__(
self,
cfg: ModelConfig,
tok,
train_sources,
device: torch.device,
seq_len: int = 1024,
batch_size: int = 16,
grad_accum: int = 8,
lr: float = 3e-4,
warmup_steps: int = 2000,
max_steps: int = 100_000,
out_dir: str = "checkpoints",
weight_decay: float = 0.1,
betas: tuple = (0.9, 0.95),
num_workers: int = 4,
pin_memory: bool = True,
seed: int = 42,
validate_every: Optional[int] = None,
val_sources: Optional[list] = None,
clip_grad_norm: Optional[float] = None,
):
torch.manual_seed(seed)
self.device = device
self.cfg = cfg
self.tok = tok
self.seq_len = seq_len
self.batch_size = batch_size
self.grad_accum = grad_accum
self.lr = lr
self.warmup_steps = warmup_steps
self.max_steps = max_steps
self.out_dir = out_dir
self.weight_decay = weight_decay
self.betas = betas
self.num_workers = num_workers
self.pin_memory = pin_memory
self.validate_every = validate_every
self.val_sources = val_sources
self.clip_grad_norm = clip_grad_norm
os.makedirs(out_dir, exist_ok=True)
self.model = SupernovaModel(cfg).to(device)
# optimizer + scheduler
self.optimizer = torch.optim.AdamW(
self.model.parameters(), lr=lr, betas=betas, weight_decay=weight_decay
)
self.scheduler = get_cosine_schedule_with_warmup(
self.optimizer, num_warmup_steps=warmup_steps, num_training_steps=max_steps
)
self.train_ds = TokenChunkDataset(tok, train_sources, seq_len=seq_len, eos_token_id=tok.eos_token_id)
self.train_dl = DataLoader(
self.train_ds,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=pin_memory,
drop_last=True,
)
if val_sources is not None:
self.val_ds = TokenChunkDataset(tok, val_sources, seq_len=seq_len, eos_token_id=tok.eos_token_id)
self.val_dl = DataLoader(self.val_ds, batch_size=batch_size, shuffle=False, num_workers=max(0, num_workers//2), pin_memory=pin_memory)
else:
self.val_dl = None
# AMP scaler
self.scaler = torch.cuda.amp.GradScaler() if device.type == "cuda" else None
# logging
self.writer = SummaryWriter(log_dir=os.path.join(out_dir, "logs"))
# training state
self.step = 0
self.micro = 0
self.running_loss = 0.0
# perf
torch.backends.cudnn.benchmark = True
def save_ckpt(self, path: str):
payload = {
"model_state_dict": self.model.state_dict(),
"optimizer_state_dict": self.optimizer.state_dict(),
"scheduler_state_dict": self.scheduler.state_dict(),
"config": self.cfg.__dict__,
"step": self.step,
}
torch.save(payload, path)
def load_ckpt(self, path: str):
ckpt = torch.load(path, map_location=self.device)
self.model.load_state_dict(ckpt["model_state_dict"])
if "optimizer_state_dict" in ckpt:
self.optimizer.load_state_dict(ckpt["optimizer_state_dict"])
if "scheduler_state_dict" in ckpt:
self.scheduler.load_state_dict(ckpt["scheduler_state_dict"])
self.step = ckpt.get("step", 0)
print(f"Resumed from {path}, step={self.step}")
@torch.no_grad()
def validate(self):
if self.val_dl is None:
return None
self.model.eval()
tot = 0.0
count = 0
for batch in self.val_dl:
x, y = batch
x = x.to(self.device, non_blocking=True)
y = y.to(self.device, non_blocking=True)
with torch.cuda.amp.autocast(enabled=(self.scaler is not None)):
_, loss = self.model(x, y)
tot += float(loss.detach().item())
count += 1
self.model.train()
return tot / max(1, count)
def train_loop(self, save_every: int = 10000, log_every: int = 50):
t0 = time.time()
for epoch in iter(int, 1): # infinite loop, break by max_steps
for batch in self.train_dl:
x, y = batch
x = x.to(self.device, non_blocking=True)
y = y.to(self.device, non_blocking=True)
# forward (AMP-capable)
if self.scaler is not None:
with torch.cuda.amp.autocast():
_, loss = self.model(x, y)
else:
_, loss = self.model(x, y)
loss = loss / self.grad_accum
if self.scaler is not None:
self.scaler.scale(loss).backward()
else:
loss.backward()
self.micro += 1
self.running_loss += float(loss.detach().item())
if self.micro % self.grad_accum == 0:
# optional clipping
if self.clip_grad_norm is not None:
if self.scaler is not None:
# unscale before clipping
self.scaler.unscale_(self.optimizer)
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.clip_grad_norm)
if self.scaler is not None:
self.scaler.step(self.optimizer)
self.scaler.update()
else:
self.optimizer.step()
self.optimizer.zero_grad(set_to_none=True)
self.scheduler.step()
self.step += 1
if self.step % log_every == 0:
grad_norm = compute_grad_norm(self.model)
avg_loss = self.running_loss * self.grad_accum / log_every
elapsed = time.time() - t0
lr_now = self.scheduler.get_last_lr()[0]
tokens_per_sec = (self.batch_size * self.seq_len * log_every) / max(1e-9, elapsed)
print(f"step={self.step} loss={avg_loss:.4f} grad_norm={grad_norm:.2f} lr={lr_now:.6f} elapsed={elapsed:.1f}s tokens/s={tokens_per_sec:.1f}")
# tensorboard
self.writer.add_scalar("train/loss", avg_loss, self.step)
self.writer.add_scalar("train/grad_norm", grad_norm, self.step)
self.writer.add_scalar("train/lr", lr_now, self.step)
self.writer.add_scalar("train/tokens_per_sec", tokens_per_sec, self.step)
self.running_loss = 0.0
t0 = time.time()
if save_every and self.step % save_every == 0:
ckpt_path = os.path.join(self.out_dir, f"supernova_step{self.step}.pt")
self.save_ckpt(ckpt_path)
print(f"Saved checkpoint {ckpt_path}")
if self.validate_every and self.step % self.validate_every == 0:
val_loss = self.validate()
if val_loss is not None:
print(f"Validation loss at step {self.step}: {val_loss:.4f}")
self.writer.add_scalar("val/loss", val_loss, self.step)
if self.step >= self.max_steps:
print("Reached max_steps; finishing training")
final_ckpt = os.path.join(self.out_dir, "supernova_final.pt")
self.save_ckpt(final_ckpt)
return
def parse_args():
ap = argparse.ArgumentParser()
ap.add_argument("--config", required=True)
ap.add_argument("--data-config", required=True)
ap.add_argument("--val-data-config", default=None)
ap.add_argument("--seq-len", type=int, default=1024)
ap.add_argument("--batch-size", type=int, default=16)
ap.add_argument("--grad-accum", type=int, default=8)
ap.add_argument("--lr", type=float, default=3e-4)
ap.add_argument("--warmup-steps", type=int, default=2000)
ap.add_argument("--max-steps", type=int, default=100000)
ap.add_argument("--save-every", type=int, default=10000)
ap.add_argument("--out-dir", type=str, default="checkpoints")
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--weight-decay", type=float, default=0.1)
ap.add_argument("--betas", type=float, nargs=2, default=(0.9, 0.95))
ap.add_argument("--num-workers", type=int, default=4)
ap.add_argument("--resume", type=str, default=None, help="path to checkpoint to resume from")
ap.add_argument("--validate-every", type=int, default=None)
ap.add_argument("--clip-grad-norm", type=float, default=None)
return ap.parse_args()
def main():
args = parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cfg = ModelConfig.from_json_file(args.config)
cfg.assert_exact_params(expected=25_000_000)
tok = load_gpt2_tokenizer()
assert tok.vocab_size == cfg.vocab_size, (
f"Tokenizer vocab size ({tok.vocab_size}) != config ({cfg.vocab_size})"
)
train_sources = load_sources_from_yaml(args.data_config)
val_sources = load_sources_from_yaml(args.val_data_config) if args.val_data_config else None
trainer = Trainer(
cfg=cfg,
tok=tok,
train_sources=train_sources,
device=device,
seq_len=args.seq_len,
batch_size=args.batch_size,
grad_accum=args.grad_accum,
lr=args.lr,
warmup_steps=args.warmup_steps,
max_steps=args.max_steps,
out_dir=args.out_dir,
weight_decay=args.weight_decay,
betas=tuple(args.betas),
num_workers=args.num_workers,
seed=args.seed,
validate_every=args.validate_every,
val_sources=val_sources,
clip_grad_norm=args.clip_grad_norm,
)
if args.resume:
trainer.load_ckpt(args.resume)
trainer.train_loop(save_every=args.save_every)
if __name__ == "__main__":
main() |