train.py

"""
This training script can be run on a single gpu mode).

To run on a single GPU, example:
$ python train.py --batch_size=32 --compile=False
"""

import os
import time
import math
import pickle

import utils
import numpy
import numpy as np
try:
    import cupy
    if cupy.cuda.is_available():
        print("CUDA available, run train on gpu...")
        np = cupy
    else:
        print("CUDA is NOT available, run train on cpu...")
except:
    print("CUDA is NOT available, run train on cpu...")
    pass

from model_gpu import GPTConfig, GPT
#from model import GPTConfig, GPT
# -----------------------------------------------------------------------------
# default config values designed to train a gpt2 (124M) on OpenWebText
# I/O

out_dir = 'out'
eval_interval = 50
log_interval = 1
eval_iters = 10 #200
eval_only = False # if True, script exits right after the first eval
always_save_checkpoint = True # if True, always save a checkpoint after each eval
init_from = 'scratch' # 'scratch' or 'resume' or 'gpt2*'
# wandb logging
wandb_log = False # disabled by default
wandb_project = 'owt'
wandb_run_name = 'gpt2' # 'run' + str(time.time())
# data
dataset = 'shakespeare'
gradient_accumulation_steps = 5 * 4 # used to simulate larger batch sizes
batch_size = 6 # if gradient_accumulation_steps > 1, this is the micro-batch size
block_size = 256 # 1024
# model
n_layer = 12
n_head = 12
n_embd = 768 #// 2
dropout = 0.0 # for pretraining 0 is good, for finetuning try 0.1+
bias = True # do we use bias inside LayerNorm and Linear layers?

# adamw optimizer
learning_rate = 6e-4 # max learning rate
max_iters = 300 # total number of training iterations
weight_decay = 1e-1
beta1 = 0.9
beta2 = 0.95
grad_clip = 0.0 # clip gradients at this value, or disable if == 0.0. Default - 0.0
# learning rate decay settings
decay_lr = True # whether to decay the learning rate
warmup_iters = 2000 # how many steps to warm up for
lr_decay_iters = 600000 # should be ~= max_iters per Chinchilla
min_lr = 6e-5 # minimum learning rate, should be ~= learning_rate/10 per Chinchilla
# DDP settings
backend = 'nccl' # 'nccl', 'gloo', etc.
# system
#dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32', 'bfloat16', or 'float16', the latter will auto implement a GradScaler
# -----------------------------------------------------------------------------
config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
exec(open('configurator.py').read()) # overrides from command line or config file
config = {k: globals()[k] for k in config_keys} # will be useful for logging
# -----------------------------------------------------------------------------

# various inits, derived attributes, I/O setup

master_process = True
seed_offset = 0
tokens_per_iter = gradient_accumulation_steps * batch_size * block_size
print(f"tokens per iteration will be: {tokens_per_iter:,}")

if master_process:
    os.makedirs(out_dir, exist_ok=True)

np.random.seed(42 + seed_offset)

# note: float16 data type will automatically use a GradScaler
# 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)

# poor man's data loader
data_dir = os.path.join('data', dataset)


def get_batch(split):
    # We recreate np.memmap every batch to avoid a memory leak, as per
    # https://stackoverflow.com/questions/45132940/numpy-memmap-memory-usage-want-to-iterate-once/61472122#61472122
    if split == 'train':
        data = numpy.memmap(os.path.join(data_dir, 'train.bin'), dtype=numpy.uint16, mode='r')
    else:
        data = numpy.memmap(os.path.join(data_dir, 'val.bin'), dtype=numpy.uint16, mode='r')

    ix = numpy.random.randint(len(data) - block_size, size=(batch_size,))
    x = np.asarray(numpy.stack([(data[i:i+block_size]).astype(numpy.int64) for i in ix]))
    y = np.asarray(numpy.stack([(data[i+1:i+1+block_size]).astype(numpy.int64) for i in ix]))
    return x, y


# init these up here, can override if init_from='resume' (i.e. from a checkpoint)
iter_num = 0
best_val_loss = 1e9

# attempt to derive vocab_size from the dataset
meta_path = os.path.join(data_dir, 'meta.pkl')
meta_vocab_size = None
if os.path.exists(meta_path):
    with open(meta_path, 'rb') as f:
        meta = pickle.load(f)
    meta_vocab_size = meta['vocab_size']
    print(f"found vocab_size = {meta_vocab_size} (inside {meta_path})")

# model init
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) # start with model_args from command line
if init_from == 'scratch':
    # init a new model from scratch
    print("Initializing a new model from scratch")
    # determine the vocab size we'll use for from-scratch training
    if meta_vocab_size is None:
        print("defaulting to vocab_size of GPT-2 to 50304 (50257 rounded up for efficiency)")
    model_args['vocab_size'] = meta_vocab_size if meta_vocab_size is not None else 50304
    gptconf = GPTConfig(**model_args)
    model = GPT(gptconf)

elif init_from == 'resume':
    print(f"Resuming training from {out_dir}")
    # resume training from a checkpoint.
    ckpt_path = os.path.join(out_dir, 'ckpt.npy')
    checkpoint = utils.load_params_dict(ckpt_path)
    # force these config attributes to be equal otherwise we can't even resume training
    # the rest of the attributes (e.g. dropout) can stay as desired from command line
    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias']:
        model_args[k] = checkpoint.any().get('config').get(k)
    model_args['vocab_size'] = checkpoint.any().get('config').get('meta_vocab_size')
    # create the model
    gptconf = GPTConfig(**model_args)
    model = GPT(gptconf)
    state_dict = checkpoint.any().get('model')
    # fix the keys of the state dictionary :(
    # honestly no idea how checkpoints sometimes get this prefix, have to debug more
    model.load_from_dict(state_dict)
    iter_num = checkpoint.any().get('iter_num')
    best_val_loss = checkpoint.any().get('best_val_loss')

elif init_from.startswith('gpt2'):
    raise Exception("Does not worked yet")
    print(f"Initializing from OpenAI GPT-2 weights: {init_from}")
    # initialize from OpenAI GPT-2 weights
    override_args = dict(dropout=dropout)
    model = GPT.from_pretrained(init_from, override_args)
    # read off the created config params, so we can store them into checkpoint correctly
    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
        model_args[k] = getattr(model.config, k)
# crop down the model block size if desired, using model surgery
# if block_size < model.config.block_size:
#     model.crop_block_size(block_size)
#     model_args['block_size'] = block_size # so that the checkpoint will have the right value

print(gptconf)
# optimizer
optimizer, optim_groups = model.configure_optimizers(weight_decay, learning_rate, [beta1, beta2])
if init_from == 'resume':
    optimizer.load_state_dict(checkpoint.any().get('optimizer'))
checkpoint = None # free up memory

# helps estimate an arbitrarily accurate loss over either split using many batches


def estimate_loss():
    out = {}
    for split in ['train', 'val']:
        losses = np.zeros(eval_iters)
        for k in range(eval_iters):
            X, Y = get_batch(split)
            logits, loss, _ = model.forward(X, Y)
            losses[k] = loss.item()
        out[split] = losses.mean()
    return out


# learning rate decay scheduler (cosine with warmup)
def get_lr(it):
    # 1) linear warmup for warmup_iters steps
    if it < warmup_iters:
        return learning_rate * it / warmup_iters
    # 2) if it > lr_decay_iters, return min learning rate
    if it > lr_decay_iters:
        return min_lr
    # 3) in between, use cosine decay down to min learning rate
    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
    assert 0 <= decay_ratio <= 1
    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
    return min_lr + coeff * (learning_rate - min_lr)


# logging
if wandb_log and master_process:
    import wandb
    wandb.init(project=wandb_project, name=wandb_run_name, config=config)

# training loop
total_time = time.time()
X, Y = get_batch('train')  # fetch the very first batch
t0 = time.time()
local_iter_num = 0  # number of iterations in the lifetime of this process
raw_model = model
running_mfu = -1.0
while True:

    # determine and set the learning rate for this iteration
    lr = get_lr(iter_num) if decay_lr else learning_rate
    optimizer.lr = lr

    # evaluate the loss on train/val sets and write checkpoints
    if iter_num % eval_interval == 0 and master_process:
        losses = estimate_loss()
        print(f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
        if wandb_log:
            wandb.log({
                "iter": iter_num,
                "train/loss": losses['train'],
                "val/loss": losses['val'],
                "lr": lr,
                "mfu": running_mfu*100,  # convert to percentage
            })
        if losses['val'] < best_val_loss or always_save_checkpoint:
            best_val_loss = losses['val']
            if iter_num > 0:
                checkpoint = {
                    'model': model.get_params_dict(),
                    'optimizer': optimizer.state_dict(),
                    'model_args': model_args,
                    'iter_num': iter_num,
                    'best_val_loss': best_val_loss,
                    'config': config,
                }
                print(f"saving checkpoint to {out_dir}")
                utils.save_params_dict(checkpoint, os.path.join(out_dir, 'ckpt'))

    if iter_num == 0 and eval_only:
        break

    # forward backward update, with optional gradient accumulation to simulate larger batch size
    # and using the GradScaler if data type is float16
    for micro_step in range(gradient_accumulation_steps):
        logits, loss, norm = model.forward(X, Y, gradient_accumulation_steps)
        # immediately async prefetch next batch while model is doing the forward pass on the GPU
        X, Y = get_batch('train')
        # backward pass, with gradient scaling if training in fp16
        model.backward()
    # clip the gradient
    if grad_clip != 0.0:
        model.clip_gradient(grad_clip, clip_norm=2)
    # step the optimizer and scaler if training in fp16
    optimizer.step()
    # flush the gradients as soon as we can, no need for this memory anymore
    optimizer.zero_grad()

    # timing and logging
    t1 = time.time()
    dt = t1 - t0
    t0 = t1
    if iter_num % log_interval == 0 and master_process:
        # get loss as float. note: this is a CPU-GPU sync point
        # scale up to undo the division above, approximating the true total loss (exact would have been a sum)
        lossf = loss.item() * gradient_accumulation_steps
        if local_iter_num >= 5:  # let the training loop settle a bit
            mfu = 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
        print(f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, mfu {running_mfu*100:.2f}%, grad norm: {norm:.4f}")
    iter_num += 1
    local_iter_num += 1
    # termination conditions
    if iter_num > max_iters:
        break
print(f"Total training time: {time.time() - total_time}")