finetune.py

"""
To resume a learning or fine-tune(addtitional training) a model.
"""

import os
import time
import math
import pickle
from contextlib import nullcontext

import numpy as np
import torch
from torch.nn.parallel import DistributedDataParallel
from torch.distributed import init_process_group, destroy_process_group
import pynvml

from model import GPTConfig, GPT

# -----------------------------------------------------------------------------
# I/O

# learning data directory, train.bin and valid .bin are expected. You should prepare them using tokenize.py
data_dir = "/mnt/nfs-mnj-hot-02/tmp/i23_eric/i23_eric/Data/pretrain/sequences_30M_len512" 
ckpt_path = "" # checkpoint path. the model for additional training
out_dir = '/mnt/nfs-mnj-hot-02/tmp/i23_eric/i23_eric/Data/out/30M_len512_2'  # output directory
log_dir = os.path.join(out_dir, 'logs') # logs will be written in to out_dir/logs

# -----------------------------------------------------------------------------
# model parameters
meta_vocab_size = 2048 
block_size = 1024
n_layer = 12
n_head = 12
n_embd = 768
bias = False # do we use bias inside LayerNorm and Linear layers?

# -----------------------------------------------------------------------------
# learning parameters
max_iters = 1000000 # total number of training iterations
eval_interval = 3000
log_interval = 1
eval_iters = 50
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*'
gradient_accumulation_steps = 8 # used to simulate larger batch sizes, should be multiple of GPU number
batch_size = 8 # A100 * 16

# adamw optimizer
learning_rate = 1e-3 # max learning rate
dropout = 0.0 # for pretraining 0 is good, for finetuning try 0.1+
weight_decay = 0
beta1 = 0.9
beta2 = 0.999
grad_clip = 1.0 # clip gradients at this value, or disable if == 0.0

# learning rate decay settings
decay_lr = False # whether to decay the learning rate
warmup_iters = 2000 # how many steps to warm up for
lr_decay_iters = 800000 # should be ~= max_iters per Chinchilla
min_lr = 2e-4 # minimum learning rate, should be ~= learning_rate/10 per Chinchilla
# DDP settings
backend = 'nccl' # 'nccl', 'gloo', etc.
# system
device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1' etc., or try 'mps' on macbooks
dtype = 'float32' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32', 'bfloat16', or 'float16', the latter will auto implement a GradScaler
compile = True # use PyTorch 2.0 to compile the model to be faster
# -----------------------------------------------------------------------------
config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
exec(open(config_path).read()) # overrides from command line or config file
config = {k: globals()[k] for k in config_keys} # will be useful for logging
# -----------------------------------------------------------------------------
# print and log all parameters
def log_and_write(filename, message):
    with open(filename, 'a') as f:
        f.write(message + "\n")
    print(message)
    
log_and_write(log_dir,f'gradient_accumulation_steps: {gradient_accumulation_steps}, batch_size: {batch_size}, \nblock_size: {block_size}, \nn_layer: {n_layer}, n_head: {n_head}, n_embd: {n_embd}, dropout: {dropout}, bias: {bias}, \nlearning_rate: {learning_rate}, max_iters: {max_iters}, \nweight_decay: {weight_decay}, beta1: {beta1}, beta2: {beta2}, grad_clip: {grad_clip}, decay_lr: {decay_lr}, \nwarmup_iters: {warmup_iters}, lr_decay_iters: {lr_decay_iters}, \nmin_lr: {min_lr}, backend: {backend}, device: {device},\n dtype: {dtype}, compile: {compile}')
log_and_write(log_dir, f'meta_vocab_size: {meta_vocab_size}')


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


# various inits, derived attributes, I/O setup
ddp = int(os.environ.get('RANK', -1)) != -1 # is this a ddp run?

if "OMPI_COMM_WORLD_SIZE" in os.environ:
    ddp = True
    os.environ['WORLD_SIZE'] = os.environ['OMPI_COMM_WORLD_SIZE']
    os.environ['RANK'] = os.environ['OMPI_COMM_WORLD_RANK']
    # os.environ['LOCAL_RANK'] = os.environ['OMPI_COMM_WORLD_LOCAL_RANK'] #搞不清WORLD LOCAL RANK是个什么
else:
    ddp = False

import pytorch_pfn_extras as ppe

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'])

    # ddp_world_size = int(os.environ['OMPI_COMM_WORLD_SIZE'])
    (
            ddp_world_size,
            ddp_rank,
            ddp_local_rank,
        ) = ppe.distributed.initialize_ompi_environment(
            backend="nccl", init_method="env"
        )
    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
    if gradient_accumulation_steps != 0 and gradient_accumulation_steps % ddp_world_size != 0:
        gradient_accumulation_steps = ddp_world_size
    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 * block_size
log_and_write(log_dir, f"tokens per iteration will be: {tokens_per_iter:,}")

pynvml.nvmlInit()
def print_gpu_memory_usage():
    handle = pynvml.nvmlDeviceGetHandleByIndex(0)
    info = pynvml.nvmlDeviceGetMemoryInfo(handle)
    print(f"Used: {info.used / 1024**2:.2f}MB/{info.total / 1024**2:.2f}MB ({info.used / info.total * 100:.2f}%)")

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 = 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)
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])
    if device_type == 'cuda':
        # pin arrays x,y, which allows us to move them to GPU asynchronously (non_blocking=True)
        x, y = x.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)
    else:
        x, y = x.to(device), y.to(device)
    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')
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}")

# 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.
    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 ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
        model_args[k] = checkpoint_model_args[k]
    # create the model
    gptconf = GPTConfig(**model_args)
    model = GPT(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']
    # print("checkpoint_args")
    # print(checkpoint)
    # print("model_args")
    # print(model_args)
elif init_from.startswith('gpt2'):
    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
model.to(device)

# initialize a GradScaler. If enabled=False scaler is a no-op
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float32'))

# optimizer
optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)
if init_from == 'resume' and inheritance_optimizer == True:
    optimizer.load_state_dict(checkpoint['optimizer'])
    print("optimizer_resumed")
    print(optimizer.state)
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:
    model = DistributedDataParallel(model, device_ids=[ddp_local_rank])

# helps estimate an arbitrarily accurate loss over either split using many batches
@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') # 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()
        log_and_write(log_dir, f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
        if iter_num % 200 == 0:
            print_gpu_memory_usage()
        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,
                }
                log_and_write(log_dir, f"saving checkpoint to {out_dir}")
                torch.save(checkpoint, os.path.join(out_dir, f'ckpt_{iter_num}.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
        log_and_write(log_dir, f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, lr {lr}, 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()

pynvml.nvmlShutdown()