# torchrun --standalone --nproc_per_node=2 train.py --batch_size=96
# train.py
import os
import time
import math
from contextlib import nullcontext
import json
import numpy as np
import torch
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group
import pandas as pd
import tiktoken
from model import GPTConfig, GPT
# Import wandb and tqdm
import wandb
from tqdm.auto import tqdm
# -----------------------------------------------------------------------------
# Default configuration with added positional encoding options
# I/O
out_dir = 'out'
eval_interval = 100 # Evaluate every 100 iterations
log_interval = 1 # Log every iteration
eval_iters = 100
eval_only = False
always_save_checkpoint = True
init_from = 'scratch' # 'scratch' | 'resume' | 'checkpoint'
checkpoint_path = '' # Path to a specific checkpoint to load
# wandb logging
wandb_log = True
wandb_project = 'gpt2_positional_encodings_100B'
wandb_run_name = 'experiment'
# data
dataset = 'fineweb'
gradient_accumulation_steps = 40
batch_size = 12
block_size = 512
# model
n_layer = 4
n_head = 4
n_embd = 256
dropout = 0.0
bias = False
# adamw optimizer
learning_rate = 6e-4
max_iters = 10000
weight_decay = 1e-1
beta1 = 0.9
beta2 = 0.95
grad_clip = 1.0
# learning rate decay settings
decay_lr = True
warmup_iters = 100
lr_decay_iters = 10000
min_lr = 6e-5
# DDP settings
backend = 'nccl'
# system
device = 'cuda'
dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16'
compile = True
# Positional Encodings
embedding_types = ['wavelet']
attention_types = ['default']
# Data collection options
collect_attention_patterns = False # Set to True to collect attention patterns
collect_activations = False # Set to True to collect activations
# Evaluation datasets
eval_datasets = ['wikitext-103-v1', 'ptb', 'lambada'] # WikiText-103 and Penn Treebank
seed = 1337
# -----------------------------------------------------------------------------
config_keys = [k for k, v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str, list, tuple))]
exec(open('configurator.py').read())
config = {k: globals()[k] for k in config_keys}
# -----------------------------------------------------------------------------
def is_compatible(embedding_type, attention_type):
# Incompatible combinations can be specified here
incompatible_combinations = [
# If specific combinations are incompatible
]
# If embedding_type or attention_type is 'none', some attention methods may not function properly
if embedding_type == 'none' and attention_type in ['relative', 'rope']:
return False
# 'rope' attention requires even dimension per head
if attention_type == 'rope' and ((n_embd // n_head) % 2 != 0):
return False
return (embedding_type, attention_type) not in incompatible_combinations
def main():
# Initialize DDP if needed
global gradient_accumulation_steps
ddp = int(os.environ.get('RANK', -1)) != -1
if ddp:
init_process_group(backend=backend)
ddp_rank = int(os.environ['RANK'])
ddp_local_rank = int(os.environ['LOCAL_RANK'])
ddp_world_size = int(os.environ['WORLD_SIZE'])
device_local = f'cuda:{ddp_local_rank}'
torch.cuda.set_device(device_local)
master_process = ddp_rank == 0
seed_offset = ddp_rank
assert gradient_accumulation_steps % ddp_world_size == 0
gradient_accumulation_steps //= ddp_world_size
else:
master_process = True
seed_offset = 0
ddp_world_size = 1
device_local = device # Use the default device
tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size
if master_process:
print(f"Tokens per iteration will be: {tokens_per_iter:,}")
if master_process:
os.makedirs(out_dir, exist_ok=True)
# Set random seed
global seed
seed += seed_offset
torch.manual_seed(seed)
np.random.seed(seed)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
device_type = 'cuda' if 'cuda' in device_local else 'cpu'
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
# Load tokenizer using tiktoken
tokenizer = tiktoken.get_encoding("gpt2")
# Prepare evaluation datasets
eval_data = {}
for eval_dataset in eval_datasets:
eval_data_path = os.path.join('data', eval_dataset)
if not os.path.exists(eval_data_path):
raise FileNotFoundError(f"Dataset {eval_dataset} not found. Please run prepare_evaluation_data.py first.")
if eval_dataset in ['wikitext-2-v1', 'wikitext-103-v1']:
train_file = [f for f in os.listdir(eval_data_path) if f.startswith('train')][0]
val_file = [f for f in os.listdir(eval_data_path) if f.startswith('validation')][0]
train_df = pd.read_parquet(os.path.join(eval_data_path, train_file))
val_df = pd.read_parquet(os.path.join(eval_data_path, val_file))
train_text = '\n'.join(train_df['text'])
val_text = '\n'.join(val_df['text'])
elif eval_dataset == 'ptb':
with open(os.path.join(eval_data_path, 'train.txt'), 'r') as f:
train_text = f.read()
with open(os.path.join(eval_data_path, 'valid.txt'), 'r') as f:
val_text = f.read()
elif eval_dataset == 'lambada':
with open(os.path.join(eval_data_path, 'lambada_test.jsonl'), 'r') as f:
data = [json.loads(line) for line in f]
test_text = '\n'.join([item['text'] for item in data])
train_text = test_text[:len(test_text)//2] # Use first half as pseudo-train
val_text = test_text[len(test_text)//2:] # Use second half as pseudo-val
else:
raise ValueError(f"Unknown dataset: {eval_dataset}")
# Tokenize
train_ids = tokenizer.encode_ordinary(train_text)
val_ids = tokenizer.encode_ordinary(val_text)
# Convert to numpy arrays
train_ids = np.array(train_ids, dtype=np.uint16)
val_ids = np.array(val_ids, dtype=np.uint16)
eval_data[eval_dataset] = {'train': train_ids, 'val': val_ids}
# Data loading
data_dir = os.path.join('data', dataset)
# Update the get_batch function to handle evaluation datasets
def get_batch(split, dataset='main'):
if dataset == 'main':
if split == 'train':
data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
else:
data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')
else:
data = eval_data[dataset][split]
ix = torch.randint(len(data) - block_size, (batch_size,))
x = torch.stack([torch.from_numpy((data[i:i+block_size]).astype(np.int64)) for i in ix])
y = torch.stack([torch.from_numpy((data[i+1:i+1+block_size]).astype(np.int64)) for i in ix])
if device_type == 'cuda':
x, y = x.pin_memory().to(device_local, non_blocking=True), y.pin_memory().to(device_local, non_blocking=True)
else:
x, y = x.to(device_local), y.to(device_local)
return x, y
# Attempt to derive vocab_size from the dataset
meta_path = os.path.join(data_dir, 'meta.json')
meta_vocab_size = None
if os.path.exists(meta_path):
with open(meta_path, 'r') as f:
meta = json.load(f)
meta_vocab_size = meta['vocab_size']
if master_process:
print(f"Found vocab_size = {meta_vocab_size} (inside {meta_path})")
# Helps estimate loss and collect attention patterns and activations
@torch.no_grad()
def estimate_loss(model, collect_attention_patterns=False, collect_activations=False, save_dir=None, max_batches_to_save=None):
out = {}
model.eval()
# Access the underlying model if wrapped with DDP
raw_model = model.module if hasattr(model, 'module') else model
# Set tracking flags on the underlying model
raw_model.config.track_attention_patterns = collect_attention_patterns
raw_model.config.track_activations = collect_activations
if collect_attention_patterns or collect_activations:
if save_dir is None:
raise ValueError("save_dir must be specified when collecting attention patterns or activations.")
if master_process:
os.makedirs(save_dir, exist_ok=True)
for split in ['train', 'val']:
losses = torch.zeros(eval_iters)
save_count = 0 # Counter for saved batches
for k in range(eval_iters):
X, Y = get_batch(split)
with ctx:
logits, loss = model(X, Y)
losses[k] = loss.item()
# Collect and save attention patterns and activations
if (collect_attention_patterns or collect_activations) and save_count < (max_batches_to_save or eval_iters):
if collect_attention_patterns or collect_activations:
if master_process:
batch_dir = os.path.join(save_dir, f"{split}_batch_{k}")
os.makedirs(batch_dir, exist_ok=True)
# Save activations
if collect_activations and hasattr(raw_model, 'activations'):
for idx, activation in enumerate(raw_model.activations):
activation_path = os.path.join(batch_dir, f"activation_layer_{idx}.pt")
torch.save(activation, activation_path)
# Save attention patterns
if collect_attention_patterns and hasattr(raw_model, 'attention_patterns'):
for idx, attention in enumerate(raw_model.attention_patterns):
attention_path = os.path.join(batch_dir, f"attention_layer_{idx}.pt")
torch.save(attention, attention_path)
# Clear activations and attention patterns from the model
raw_model.activations = []
raw_model.attention_patterns = []
save_count += 1
out[split] = losses.mean().item()
# Evaluate on additional datasets
for eval_dataset in eval_datasets:
split_losses = {}
for split in ['train', 'val']:
losses = torch.zeros(eval_iters)
save_count = 0 # Counter for saved batches
for k in range(eval_iters):
X, Y = get_batch(split, dataset=eval_dataset)
with ctx:
logits, loss = model(X, Y)
losses[k] = loss.item()
# Collect and save attention patterns and activations
if (collect_attention_patterns or collect_activations) and save_count < (max_batches_to_save or eval_iters):
if collect_attention_patterns or collect_activations:
if master_process:
batch_dir = os.path.join(save_dir, f"{eval_dataset}_{split}_batch_{k}")
os.makedirs(batch_dir, exist_ok=True)
# Save activations
if collect_activations and hasattr(raw_model, 'activations'):
for idx, activation in enumerate(raw_model.activations):
activation_path = os.path.join(batch_dir, f"activation_layer_{idx}.pt")
torch.save(activation, activation_path)
# Save attention patterns
if collect_attention_patterns and hasattr(raw_model, 'attention_patterns'):
for idx, attention in enumerate(raw_model.attention_patterns):
attention_path = os.path.join(batch_dir, f"attention_layer_{idx}.pt")
torch.save(attention, attention_path)
# Clear activations and attention patterns from the model
raw_model.activations = []
raw_model.attention_patterns = []
save_count += 1
split_losses[split] = losses.mean().item()
out[eval_dataset] = split_losses
model.train()
# Reset tracking flags
raw_model.config.track_attention_patterns = False
raw_model.config.track_activations = False
return out
# Learning rate decay scheduler
def get_lr(it):
if it < warmup_iters:
return learning_rate * it / warmup_iters
if it > lr_decay_iters:
return min_lr
decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
return min_lr + coeff * (learning_rate - min_lr)
# Training loop over positional encoding combinations
for embedding_type in embedding_types:
for attention_type in attention_types:
if not is_compatible(embedding_type, attention_type):
if master_process:
print(f"Skipping incompatible combination: Embedding={embedding_type}, Attention={attention_type}")
continue
# Configure model arguments
model_args = dict(
n_layer=n_layer,
n_head=n_head,
n_embd=n_embd,
block_size=block_size,
bias=bias,
vocab_size=None,
dropout=dropout,
embedding_type=embedding_type,
attention_type=attention_type,
track_activations=False,
track_attention_patterns=False,
)
# Initialize or resume model
iter_num = 0
best_val_loss = 1e9 # initialize best val loss to a high value
checkpoint = None
run_id = None # Initialize run_id to None
if init_from == 'scratch':
if master_process:
print(f"\nInitializing new model with embedding_type={embedding_type}, attention_type={attention_type}")
if meta_vocab_size is None:
if master_process:
print("Defaulting to vocab_size of GPT-2 to 50257")
model_args['vocab_size'] = meta_vocab_size if meta_vocab_size is not None else 50257
gptconf = GPTConfig(**model_args)
model = GPT(gptconf)
elif init_from == 'resume':
# Resume from the latest checkpoint
ckpt_path = os.path.join(out_dir, f"ckpt_{embedding_type}_{attention_type}.pt")
if not os.path.exists(ckpt_path):
raise FileNotFoundError(f"Checkpoint not found at {ckpt_path}")
if master_process:
print(f"\nResuming training from checkpoint {ckpt_path}")
checkpoint = torch.load(ckpt_path, map_location=device_local)
gptconf = GPTConfig(**checkpoint['model_args'])
model = GPT(gptconf)
model.load_state_dict(checkpoint['model'])
iter_num = checkpoint['iter_num']
best_val_loss = checkpoint['best_val_loss']
seed = checkpoint.get('seed', seed)
run_id = checkpoint.get('wandb_run_id', None)
elif init_from == 'checkpoint':
# Resume from a specific checkpoint
if not checkpoint_path or not os.path.exists(checkpoint_path):
raise FileNotFoundError(f"Checkpoint not found at {checkpoint_path}")
if master_process:
print(f"\nLoading model from checkpoint {checkpoint_path}")
checkpoint = torch.load(checkpoint_path, map_location=device_local)
gptconf = GPTConfig(**checkpoint['model_args'])
model = GPT(gptconf)
model.load_state_dict(checkpoint['model'])
iter_num = checkpoint['iter_num']
best_val_loss = checkpoint['best_val_loss']
seed = checkpoint.get('seed', seed)
run_id = checkpoint.get('wandb_run_id', None)
else:
raise ValueError(f"Unknown init_from '{init_from}'")
# Set random seed
seed += seed_offset
torch.manual_seed(seed)
np.random.seed(seed)
model.to(device_local)
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float16'))
optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)
# Load optimizer state if resuming
if checkpoint is not None:
optimizer.load_state_dict(checkpoint['optimizer'])
if compile:
if master_process:
print("Compiling the model... (takes a ~minute)")
unoptimized_model = model
model = torch.compile(model)
if ddp:
model = DDP(model, device_ids=[ddp_local_rank])
# Logging with WandB
if wandb_log and master_process:
run_name = f"{embedding_type}_{attention_type}_{wandb_run_name}"
# Initialize WandB
wandb.init(project=wandb_project, name=run_name, config=config, resume='allow', id=run_id)
# Save the run ID for resuming later
run_id = wandb.run.id
else:
run_id = None
# Training loop
X, Y = get_batch('train')
t0 = time.time()
local_iter_num = 0
raw_model = model.module if hasattr(model, 'module') else model
running_mfu = -1.0
progress_bar = tqdm(total=max_iters, initial=iter_num, desc=f"Training {embedding_type} + {attention_type}", disable=not master_process)
progress_bar_update_freq = 1 # Update progress bar every iteration
while True:
# Determine learning rate
lr = get_lr(iter_num) if decay_lr else learning_rate
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# Evaluate and checkpoint
if iter_num % eval_interval == 0 and iter_num > 0:
# Define save_dir for collected data
eval_data_dir = os.path.join('data', 'eval_data', f"{embedding_type}_{attention_type}", f"step_{iter_num}")
# Set a limit on the number of batches to save during evaluation
max_batches_to_save = 10 # Adjust this number as needed to control storage usage
losses = estimate_loss(model,
collect_attention_patterns=collect_attention_patterns,
collect_activations=collect_activations,
save_dir=eval_data_dir,
max_batches_to_save=max_batches_to_save)
if master_process:
print(f"\nStep {iter_num}:")
print(f"Train loss: {losses['train']:.4f}, Val loss: {losses['val']:.4f}")
for eval_dataset in eval_datasets:
print(f"{eval_dataset} - Train loss: {losses[eval_dataset]['train']:.4f}, Val loss: {losses[eval_dataset]['val']:.4f}")
# Log to wandb
if wandb_log:
wandb_metrics = {
"iter": iter_num,
"train/loss": losses['train'],
"val/loss": losses['val'],
"lr": lr,
"mfu": running_mfu * 100,
}
for eval_dataset in eval_datasets:
wandb_metrics[f"{eval_dataset}/train_loss"] = losses[eval_dataset]['train']
wandb_metrics[f"{eval_dataset}/val_loss"] = losses[eval_dataset]['val']
wandb.log(wandb_metrics, step=iter_num)
if losses['val'] < best_val_loss or always_save_checkpoint:
best_val_loss = losses['val']
if iter_num > 0:
checkpoint = {
'model': raw_model.state_dict(),
'optimizer': optimizer.state_dict(),
'model_args': model_args,
'iter_num': iter_num,
'best_val_loss': best_val_loss,
'config': config,
'seed': seed,
'wandb_run_id': run_id
}
ckpt_path = os.path.join(out_dir, f"ckpt_{embedding_type}_{attention_type}.pt")
if master_process:
print(f"Saving checkpoint to {ckpt_path}")
torch.save(checkpoint, ckpt_path)
# Update progress bar postfix
if master_process:
postfix_dict = {
'train_loss': f"{losses['train']:.4f}",
'val_loss': f"{losses['val']:.4f}"
}
for eval_dataset in eval_datasets:
postfix_dict[f"{eval_dataset}_val_loss"] = f"{losses[eval_dataset]['val']:.4f}"
progress_bar.set_postfix(postfix_dict)
if eval_only:
break
# Forward backward update
for micro_step in range(gradient_accumulation_steps):
if ddp:
model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
with ctx:
logits, loss = model(X, Y)
loss = loss / gradient_accumulation_steps
X, Y = get_batch('train')
scaler.scale(loss).backward()
if grad_clip != 0.0:
scaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad(set_to_none=True)
# Logging
t1 = time.time()
dt = t1 - t0
t0 = t1
if iter_num % log_interval == 0:
lossf = loss.item() * gradient_accumulation_steps
if local_iter_num >= 5:
mfu = raw_model.estimate_mfu(batch_size * gradient_accumulation_steps, dt)
running_mfu = mfu if running_mfu == -1.0 else 0.9 * running_mfu + 0.1 * mfu
if master_process:
progress_bar.set_postfix({
'loss': f"{lossf:.4f}",
'lr': f"{lr:.2e}",
'mfu': f"{running_mfu*100:.2f}%",
'time_per_iter_ms': f"{dt * 1000:.2f}ms",
})
if wandb_log:
wandb.log({
"iter": iter_num,
"train/loss": lossf,
"lr": lr,
"mfu": running_mfu * 100,
"time_per_iter_ms": dt * 1000,
}, step=iter_num)
iter_num += 1
local_iter_num += 1
if master_process:
progress_bar.update(progress_bar_update_freq)
# Termination conditions
if iter_num > max_iters:
break
if master_process:
progress_bar.close()
if wandb_log and master_process:
wandb.finish()
# Destroy the process group after all models have been trained
if ddp:
destroy_process_group()
if __name__ == '__main__':
main()