train.py

import time
import os
import math
import pickle
from contextlib import nullcontext
from torch.nn import functional as F
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 torch.distributed as dist

from model import ModelArgs, Llama
from transformers import LlamaTokenizer

# -----------------------------------------------------------------------------
# default config values designed to train a model on Shakespeare dataset
# I/O
out_dir = 'out'
eval_interval = 10
log_interval = 10
eval_iters = 20
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 'path/to/checkpoint'
# wandb logging
wandb_log = False  # disabled by default
wandb_project = 'shakespeare'
wandb_run_name = 'transformer'
# data
dataset = 'shakespeare_char'
gradient_accumulation_steps = 1
batch_size = 64
block_size = 256
# model
n_layer = 6
n_head = 6
dim = 384
hidden_dim = 1536
# optimizer
learning_rate = 3e-4
max_iters = 40
weight_decay = 1e-2
beta1 = 0.9
beta2 = 0.95
grad_clip = 1.0
# learning rate decay settings
decay_lr = True
warmup_iters = 1000
lr_decay_iters = 4000
min_lr = 6e-5
# system
device = 'cuda'
dtype = 'float16'
compile = True
loss=0.0
backend = 'nccl' # 'nccl', 'gloo', etc.
# -----------------------------------------------------------------------------
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
ddp = int(os.environ.get('RANK', -1)) != -1
if ddp:
    dist.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 = f'cuda:{ddp_local_rank}'
    torch.cuda.set_device(device)
    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
tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size
print(f"tokens per iteration will be: {tokens_per_iter:,}")

if master_process:
    os.makedirs(out_dir, exist_ok=True)
torch.manual_seed(1337 + seed_offset)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
device_type = 'cuda' if 'cuda' in device else 'cpu'
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=torch.float16)

# Initialize the LLaMA tokenizer
tokenizer = LlamaTokenizer.from_pretrained("Doctor-Shotgun/TinyLlama-1.1B-32k-Instruct")

# poor man's data loader
data_dir = os.path.join('data', dataset)
train_data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
val_data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')

def get_batch(split):
    data = train_data if split == 'train' else val_data
    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])
    x, y = x.to(device), y.to(device)
    return x, y

# init these up here, can override if init_from='resume'
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 = ModelArgs(
    dim=dim, n_layers=n_layer, n_heads=n_head,
    vocab_size=meta_vocab_size, max_seq_len=block_size,
    hidden_dim=hidden_dim
)
model = Llama(model_args)

if init_from == 'scratch':
    print("Initializing a new model from scratch")
elif init_from == 'resume':
    print(f"Resuming training from {out_dir}")
    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
    checkpoint = torch.load(ckpt_path, map_location=device)
    model.load_state_dict(checkpoint['model'])
    iter_num = checkpoint['iter_num']
    best_val_loss = checkpoint['best_val_loss']
else:
    print(f"Initializing from {init_from}")
    checkpoint = torch.load(init_from, map_location=device)
    model.load_state_dict(checkpoint['model'])

model.to(device)

# optimizer
optimizer = torch.optim.AdamW(
    model.parameters(),
    lr=learning_rate,
    weight_decay=weight_decay,
    betas=(beta1, beta2)
)
if init_from == 'resume':
    optimizer.load_state_dict(checkpoint['optimizer'])

# compile the model
if compile:
    print("compiling the model... (takes a ~minute)")
    unoptimized_model = model
    model = torch.compile(model)
print("inshallah this will be AGI")
# wrap model into DDP container
if ddp:
    model = DDP(model, device_ids=[ddp_local_rank])

@torch.no_grad()
def estimate_loss():
    out = {}
    model.eval()
    for split in ['train', 'val']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            X, Y = get_batch(split)
            with ctx:
                logits, loss = model(X, Y)
            losses[k] = loss.item()
        out[split] = losses.mean()
    model.train()
    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
X, Y = get_batch('train')
t0 = time.time()
local_iter_num = 0
raw_model = model.module if ddp else model
running_mfu = -1.0
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float16'))
while True:

    # determine and set the learning rate for this iteration
    lr = get_lr(iter_num) if decay_lr else learning_rate
    for param_group in optimizer.param_groups:
        param_group['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': raw_model.state_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}")
                torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))

    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):
        if ddp:
            # in DDP training we only need to sync gradients at the last micro step.
            # the official way to do this is with model.no_sync() context manager, but
            # I really dislike that this bloats the code and forces us to repeat code
            # looking at the source of that context manager, it just toggles this variable
            model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
        with ctx:
            logits, loss = model(X, Y)
            loss = loss / gradient_accumulation_steps # scale the loss to account for gradient accumulation
        # 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
        scaler.scale(loss).backward()
    # clip the gradient
    if grad_clip != 0.0:
        scaler.unscale_(optimizer)
        torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
    # step the optimizer and scaler if training in fp16
    scaler.step(optimizer)
    scaler.update()
    # flush the gradients as soon as we can, no need for this memory anymore
    optimizer.zero_grad(set_to_none=True)

    # 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 = 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
        print(f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, mfu {running_mfu*100:.2f}%")
    iter_num += 1
    local_iter_num += 1

    # termination conditions
    if iter_num > max_iters:
        break

if ddp:
    destroy_process_group()