full_sft.py

import os
import platform
import argparse
import time
import math
import warnings

import pandas as pd
import torch
import torch.nn.functional as F
import torch.distributed as dist
from contextlib import nullcontext

from torch import optim
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data import DataLoader, DistributedSampler
from transformers import AutoTokenizer, AutoModel
from models.model_llama import Transformer
from models.LMConfig import LMConfig
from models.dataset import SFTDataset

warnings.filterwarnings('ignore')


def Logger(content):
    if not ddp or dist.get_rank() == 0:
        print(content)


def get_lr(it, all):
    warmup_iters = args.warmup_iters
    lr_decay_iters = all
    min_lr = args.learning_rate / 10

    if it < warmup_iters:
        return args.learning_rate * it / warmup_iters
    if it > lr_decay_iters:
        return min_lr
    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))
    return min_lr + coeff * (args.learning_rate - min_lr)


def train_epoch(epoch, wandb, start_step=0):
    start_time = time.time()
    for step, (X, Y, loss_mask) in enumerate(train_loader, start=start_step):  # 从 start_step 开始
        X = X.to(args.device)
        Y = Y.to(args.device)
        loss_mask = loss_mask.to(args.device)
        lr = get_lr(epoch * iter_per_epoch + step, args.epochs * iter_per_epoch)
        for param_group in optimizer.param_groups:
            param_group['lr'] = lr

        with ctx:
            logits = model(X, Y).logits
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), Y.view(-1), ignore_index=0, reduction='none')
            loss_mask = loss_mask.view(-1)
            loss = torch.sum(loss * loss_mask) / loss_mask.sum()

        scaler.scale(loss).backward()

        if (step + 1) % args.accumulation_steps == 0:
            scaler.unscale_(optimizer)
            torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)

            scaler.step(optimizer)
            scaler.update()

            optimizer.zero_grad(set_to_none=True)

        if step % args.log_interval == 0:
            spend_time = time.time() - start_time
            Logger(
                'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.7f} epoch_Time:{}min:'.format(
                    epoch,
                    args.epochs,
                    step,
                    iter_per_epoch,
                    loss.item(),
                    optimizer.param_groups[-1]['lr'],
                    spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))

            if (wandb is not None) and (not ddp or dist.get_rank() == 0):
                wandb.log({"loss": loss,
                           "lr": optimizer.param_groups[-1]['lr'],
                           "epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})

        if (step + 1) % args.save_interval == 0 and (not ddp or dist.get_rank() == 0):
            model.eval()
            moe_path = '_moe' if lm_config.use_moe else ''
            ckp = f'{args.save_dir}/full_sft_{lm_config.dim}{moe_path}.pth'

            # 保存模型、优化器和其他状态
            checkpoint = {
                'model_state': model.module.state_dict() if isinstance(model, torch.nn.parallel.DistributedDataParallel) else model.state_dict(),
                'optimizer_state': optimizer.state_dict(),
                'scaler_state': scaler.state_dict(),
                'epoch': epoch,
                'step': step,
                'args': vars(args)  # 保存训练的参数
            }

            torch.save(checkpoint, ckp)
            Logger(f"Checkpoint saved at {ckp}")
            model.train()

            
def load_checkpoint(checkpoint_path, model, optimizer, scaler):
    if os.path.exists(checkpoint_path):
        Logger(f"Loading checkpoint from {checkpoint_path}")
        checkpoint = torch.load(checkpoint_path, map_location=args.device)
        
        model.load_state_dict(checkpoint['model_state'])
        optimizer.load_state_dict(checkpoint['optimizer_state'])
        scaler.load_state_dict(checkpoint['scaler_state'])
        start_epoch = checkpoint['epoch'] + 1  # 从上次的epoch继续
        start_step = checkpoint['step'] + 1    # 从上次的step继续

        return start_epoch, start_step
    else:
        Logger(f"No checkpoint found at {checkpoint_path}, starting from scratch.")
        return 0, 0  # 如果没有找到检查点，从头开始


def init_model():
    tokenizer = AutoTokenizer.from_pretrained('./models/tokenizer_model')
    model_from = 1  # 1从权重，2用transformers

    def count_parameters(model):
        return sum(p.numel() for p in model.parameters() if p.requires_grad)

    if model_from == 1:
        model = Transformer(lm_config)
        moe_path = '_moe' if lm_config.use_moe else ''
        ckp = f'./out/pretrain_{lm_config.dim}{moe_path}.pth'
        state_dict = torch.load(ckp, map_location=args.device)
        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, strict=False)
    else:
        model = AutoModel.from_pretrained('./', trust_remote_code=True)

    Logger(f'LLM总参数量：{count_parameters(model) / 1e6:.3f} 百万')
    model = model.to(args.device)

    return model, tokenizer


def init_distributed_mode():
    if not ddp: return
    global ddp_local_rank, DEVICE

    dist.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)


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Dylan Full SFT")
    parser.add_argument("--out_dir", type=str, default="out", help="Output directory")
    parser.add_argument("--epochs", type=int, default=19, help="Number of epochs")
    parser.add_argument("--batch_size", type=int, default=32, help="Batch size")
    parser.add_argument("--learning_rate", type=float, default=1e-4, help="Learning rate")
    parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu", help="Device to use")
    parser.add_argument("--dtype", type=str, default="bfloat16", help="Data type")
    parser.add_argument("--use_wandb", action="store_true", help="Use Weights & Biases")
    parser.add_argument("--wandb_project", type=str, default="Dylan-Full-SFT", help="Weights & Biases project name")
    parser.add_argument("--num_workers", type=int, default=8, help="Number of workers for data loading")
    parser.add_argument("--ddp", action="store_true", help="Use DistributedDataParallel")
    parser.add_argument("--accumulation_steps", type=int, default=1, help="Gradient accumulation steps")
    parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping threshold")
    parser.add_argument("--warmup_iters", type=int, default=0, help="Number of warmup iterations")
    parser.add_argument("--log_interval", type=int, default=100, help="Logging interval")
    parser.add_argument("--save_interval", type=int, default=1000, help="Model saving interval")
    parser.add_argument('--local_rank', type=int, default=-1, help='local rank for distributed training')

    args = parser.parse_args()

    lm_config = LMConfig()
    max_seq_len = lm_config.max_seq_len
    args.save_dir = os.path.join(args.out_dir)
    os.makedirs(args.save_dir, exist_ok=True)
    os.makedirs(args.out_dir, exist_ok=True)
    tokens_per_iter = args.batch_size * max_seq_len
    torch.manual_seed(1337)
    device_type = "cuda" if "cuda" in args.device else "cpu"

    args.wandb_run_name = f"Dylan-Full-SFT-Epoch-{args.epochs}-BatchSize-{args.batch_size}-LearningRate-{args.learning_rate}"

    ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast()
    ddp = int(os.environ.get("RANK", -1)) != -1  # is this a ddp run?
    ddp_local_rank, DEVICE = 0, "cuda:0"
    if ddp:
        init_distributed_mode()
        args.device = torch.device(DEVICE)

    if args.use_wandb and (not ddp or ddp_local_rank == 0):
        import wandb
        wandb.init(project=args.wandb_project, name=args.wandb_run_name)
    else:
        wandb = None

    model, tokenizer = init_model()

    # 加载数据集
    df = pd.read_csv('./data/processed/sft_data_single_large.csv')
    df = df.sample(frac=1.0)
    train_ds = SFTDataset(df, tokenizer, max_length=max_seq_len)
    train_sampler = DistributedSampler(train_ds) if ddp else None
    train_loader = DataLoader(
        train_ds,
        batch_size=args.batch_size,
        pin_memory=True,
        drop_last=False,
        shuffle=False,
        num_workers=args.num_workers,
        sampler=train_sampler
    )

    # 初始化优化器和scaler
    scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16']))
    optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)

    # 检查点路径
    moe_path = '_moe' if lm_config.use_moe else ''
    checkpoint_path = f'{args.save_dir}/full_sft_{lm_config.dim}{moe_path}.pth'

    # 加载检查点，恢复训练
    start_epoch, start_step = load_checkpoint(checkpoint_path, model, optimizer, scaler)

    # 编译模型（如果适用）
    if False and not lm_config.use_moe and platform.system() != 'Windows' and float(torch.__version__.split('.')[0]) >= 2:
        Logger("compiling the model... (takes a ~minute)")
        unoptimized_model = model
        model = torch.compile(model)

    # 分布式训练设置
    if ddp:
        model._ddp_params_and_buffers_to_ignore = {"pos_cis"}
        model = DistributedDataParallel(model, device_ids=[ddp_local_rank])

    # 每个epoch的迭代次数
    iter_per_epoch = len(train_loader)

    # 开始训练
    for epoch in range(start_epoch, args.epochs):
        train_epoch(epoch, wandb, start_step)
        start_step = 0  # 从第二个 epoch 开始，step 重新从 0 开始