"""

This training script can be run both on a single gpu in debug mode,

and also in a larger training run with distributed data parallel (ddp).

To run on a single GPU small debug run, example:

$ python -m train.py --compile=False --eval_iters=10 --batch_size=8

To run with DDP on 4 gpus on 1 node, example:

$ torchrun --standalone --nproc_per_node=4 train.py

To run with DDP on 4 gpus across 2 nodes, example:

- Run on the first (master) node with example IP 123.456.123.456:

$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr=123.456.123.456 --master_port=1234 train.py

- Run on the worker node:

$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr=123.456.123.456 --master_port=1234 train.py

(If your cluster does not have Infiniband interconnect prepend NCCL_IB_DISABLE=1)

"""

import math

import os

import time

from contextlib import nullcontext

from datetime import datetime

from functools import partial

import torch

from model import Transformer, ModelArgs

from torch.distributed import destroy_process_group, init_process_group

from torch.nn.parallel import DistributedDataParallel as DDP

from tinystories import Task

# -----------------------------------------------------------------------------

# I/O

out_dir = "out"

eval_interval = 2000

log_interval = 1

eval_iters = 100

eval_only = False # if True, script exits right after the first eval

always_save_checkpoint = False # if True, always save a checkpoint after each eval

init_from = "scratch" # 'scratch' or 'resume'

# wandb logging

wandb_log = False # disabled by default

wandb_project = "llamac"

wandb_run_name = "run" + datetime.now().strftime("%Y_%m_%d_%H_%M_%S")

# data

batch_size = 128 # if gradient_accumulation_steps > 1, this is the micro-batch size

max_seq_len = 256

vocab_source = "llama2" # llama2|custom; use Lllama 2 vocab from Meta, or custom trained

vocab_size = 32000 # the Llama 2 tokenizer has 32K tokens

# model

dim = 288

n_layers = 6

n_heads = 6

multiple_of = 32

dropout = 0.0

# adamw optimizer

gradient_accumulation_steps = 4 # used to simulate larger batch sizes

learning_rate = 5e-4 # max learning rate

max_iters = 100000 # total number of training iterations

weight_decay = 1e-1

beta1 = 0.9

beta2 = 0.95

grad_clip = 1.0 # clip gradients at this value, or disable if == 0.0

# learning rate decay settings

decay_lr = True # whether to decay the learning rate

warmup_iters = 1000 # how many steps to warm up for

# system

device = "cuda" # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1' etc., or try 'mps' on macbooks

dtype = "bfloat16" # float32|bfloat16|float16

compile = True # use PyTorch 2.0 to compile the model to be faster

# -----------------------------------------------------------------------------

config_keys = [

]

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

# -----------------------------------------------------------------------------

# fixing some hyperparams to sensible defaults

lr_decay_iters = max_iters # should be ~= max_iters per Chinchilla

min_lr = 0.0 # minimum learning rate, should be ~= learning_rate/10 per Chinchilla

# validating checks

assert vocab_source in ["llama2", "custom"]

assert vocab_source == "custom" or vocab_size == 32000, "The vocab from Meta has 32K tokens"

# various inits, derived attributes, I/O setup

ddp = int(os.environ.get("RANK", -1)) != -1 # is this a ddp run?

if ddp:

init_process_group(backend="nccl")

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 # this process will do logging, checkpointing etc.

seed_offset = ddp_rank # each process gets a different seed

# world_size number of processes will be training simultaneously, so we can scale

# down the desired gradient accumulation iterations per process proportionally

assert gradient_accumulation_steps % ddp_world_size == 0

gradient_accumulation_steps //= ddp_world_size

else:

# if not ddp, we are running on a single gpu, and one process

master_process = True

seed_offset = 0

ddp_world_size = 1

tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * max_seq_len

if master_process:

print(f"tokens per iteration will be: {tokens_per_iter:,}")

print(f"breaks down as: {gradient_accumulation_steps} grad accum steps * {ddp_world_size} processes * {batch_size} batch size * {max_seq_len} max seq len")

if master_process:

os.makedirs(out_dir, exist_ok=True)

torch.manual_seed(1337 + seed_offset)

torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul

torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn

device_type = "cuda" if "cuda" in device else "cpu" # for later use in torch.autocast

# note: float16 data type will automatically use a GradScaler

ptdtype = {"float32": torch.float32, "bfloat16": torch.bfloat16, "float16": torch.float16}[dtype]

ctx = (

)

# task-specific setup

iter_batches = partial(

)

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

iter_num = 0

best_val_loss = 1e9

# model init

model_args = dict(

) # start with model_args from command line

if init_from == "scratch":

# init a new model from scratch

print("Initializing a new model from scratch")

gptconf = ModelArgs(**model_args)

model = Transformer(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.pt")

checkpoint = torch.load(ckpt_path, map_location=device)

checkpoint_model_args = checkpoint["model_args"]

# 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 ["dim", "n_layers", "n_heads", "n_kv_heads", "vocab_size", "multiple_of", "max_seq_len"]:

model_args[k] = checkpoint_model_args[k]

# create the model

gptconf = ModelArgs(**model_args)

model = Transformer(gptconf)

state_dict = checkpoint["model"]

# fix the keys of the state dictionary :(

# honestly no idea how checkpoints sometimes get this prefix, have to debug more

unwanted_prefix = "_orig_mod."

for k, v in list(state_dict.items()):

if k.startswith(unwanted_prefix):

state_dict[k[len(unwanted_prefix) :]] = state_dict.pop(k)

model.load_state_dict(state_dict)

iter_num = checkpoint["iter_num"]

best_val_loss = checkpoint["best_val_loss"]

model.to(device)

# initialize a GradScaler. If enabled=False scaler is a no-op

scaler = torch.cuda.amp.GradScaler(enabled=(dtype == "float16"))

# optimizer

optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)

if init_from == "resume" and "optimizer" in checkpoint:

optimizer.load_state_dict(checkpoint["optimizer"])

checkpoint = None # free up memory

# compile the model

if compile:

print("compiling the model... (takes a ~minute)")

unoptimized_model = model

model = torch.compile(model) # requires PyTorch 2.0

# wrap model into DDP container

if ddp:

# Ignore the `freqs_cis` buffer so that DDP does not broadcast it at

# construction time since NCCL does not support `ComplexFloat`

prefix = "_orig_mod." if compile else ""

model._ddp_params_and_buffers_to_ignore = {prefix + "freqs_cis"}

model = DDP(model, device_ids=[ddp_local_rank])

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

@torch.no_grad()

def estimate_loss():

return out

# learning rate decay scheduler (cosine with warmup)

def get_lr(it):

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

train_batch_iter = iter_batches(split="train")

X, Y = next(train_batch_iter) # fetch the very first batch

t0 = time.time()

local_iter_num = 0 # number of iterations in the lifetime of this process

raw_model = model.module if ddp else model # unwrap DDP container if needed

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

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:

try:

wandb.log(

{

"iter": iter_num,

"tokens": iter_num * tokens_per_iter,

"loss/train": losses["train"],

"loss/val": losses["val"],

"lr": lr,

"mfu": running_mfu * 100, # convert to percentage

}

)

except Exception as e:

print(f"logging to wandb failed: {e}")

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"))

raw_model.export(os.path.join(out_dir, "model.bin"))

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 = model(X, Y)

loss = raw_model.last_loss

loss = loss / gradient_accumulation_steps

# immediately async prefetch next batch while model is doing the forward pass on the GPU

X, Y = next(train_batch_iter)

# 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, scale up due to the divide above. note: this is a CPU-GPU sync point

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_num} | loss {lossf:.4f} | lr {lr:e} | {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()