train_distill_reason.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, nn
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data import DataLoader, DistributedSampler
from transformers import AutoTokenizer, AutoModelForCausalLM
from model.model import MiniMindLM
from model.LMConfig import LMConfig
from model.dataset import SFTDataset

warnings.filterwarnings('ignore')


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


def get_lr(current_step, total_steps, lr):
    return lr / 10 + 0.5 * lr * (1 + math.cos(math.pi * current_step / total_steps))


def train_epoch(epoch, wandb):
    # 思考标签占位符
    start_of_think_ids = tokenizer('<think>').input_ids
    end_of_think_ids = tokenizer('</think>').input_ids
    start_of_answer_ids = tokenizer('<answer>').input_ids
    end_of_answer_ids = tokenizer('</answer>').input_ids
    loss_fct = nn.CrossEntropyLoss(reduction='none')
    start_time = time.time()
    for step, (X, Y, loss_mask) in enumerate(train_loader):
        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, args.learning_rate)
        for param_group in optimizer.param_groups:
            param_group['lr'] = lr

        with ctx:
            res = model(X)
            loss = loss_fct(
                res.logits.view(-1, res.logits.size(-1)),
                Y.view(-1)
            ).view(Y.size())
            sp_ids = torch.isin(Y.view(-1),
                                torch.tensor(start_of_think_ids + end_of_think_ids
                                             + start_of_answer_ids + end_of_answer_ids
                                             ).to(args.device))
            # 在 sp_ids 对应的位置增加额外的惩罚
            loss_mask = loss_mask.view(-1)
            loss_mask_sum = loss_mask.sum()
            loss_mask[sp_ids] = 10
            loss_mask = loss_mask.view(Y.size())
            loss = (loss * loss_mask).sum() / loss_mask_sum
            loss += res.aux_loss
            loss = loss / args.accumulation_steps

        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:{:.12f} epoch_Time:{}min:'.format(
                    epoch + 1,
                    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}/reason_{lm_config.dim}{moe_path}.pth'

            if isinstance(model, torch.nn.parallel.DistributedDataParallel):
                state_dict = model.module.state_dict()
            else:
                state_dict = model.state_dict()

            torch.save(state_dict, ckp)
            model.train()


def init_model(lm_config):
    tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
    model = MiniMindLM(lm_config)
    moe_path = '_moe' if lm_config.use_moe else ''
    ckp = f'./out/rlhf_{lm_config.dim}{moe_path}.pth'
    state_dict = torch.load(ckp, map_location=args.device)
    model.load_state_dict(state_dict, strict=False)
    Logger(f'LLM总参数量：{sum(p.numel() for p in model.parameters() if p.requires_grad) / 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="MiniMind Distill Reasoning")
    parser.add_argument("--out_dir", type=str, default="out")
    parser.add_argument("--epochs", type=int, default=1)
    parser.add_argument("--batch_size", type=int, default=8)
    parser.add_argument("--learning_rate", type=float, default=1e-6)
    parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu")
    parser.add_argument("--dtype", type=str, default="bfloat16")
    parser.add_argument("--use_wandb", action="store_true")
    parser.add_argument("--wandb_project", type=str, default="MiniMind-Full-SFT")
    parser.add_argument("--num_workers", type=int, default=1)
    parser.add_argument("--ddp", action="store_true")
    parser.add_argument("--accumulation_steps", type=int, default=1)
    parser.add_argument("--grad_clip", type=float, default=1.0)
    parser.add_argument("--warmup_iters", type=int, default=0)
    parser.add_argument("--log_interval", type=int, default=1)
    parser.add_argument("--save_interval", type=int, default=50)
    parser.add_argument('--local_rank', type=int, default=-1)
    parser.add_argument('--dim', default=512, type=int)
    parser.add_argument('--n_layers', default=8, type=int)
    parser.add_argument('--max_seq_len', default=1024, type=int)
    parser.add_argument('--use_moe', default=False, type=bool)
    parser.add_argument("--data_path", type=str, default="./dataset/r1_mix_1024.jsonl")

    args = parser.parse_args()

    lm_config = LMConfig(dim=args.dim, n_layers=args.n_layers, max_seq_len=args.max_seq_len, use_moe=args.use_moe)
    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 * lm_config.max_seq_len
    torch.manual_seed(1337)
    device_type = "cuda" if "cuda" in args.device else "cpu"

    args.wandb_run_name = f"MiniMind-Distill-Reasoning-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(lm_config)

    train_ds = SFTDataset(args.data_path, tokenizer, max_length=lm_config.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
    )

    scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16']))
    optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)

    if ddp:
        model._ddp_params_and_buffers_to_ignore = {"pos_cis"}
        model = DistributedDataParallel(model, device_ids=[ddp_local_rank])

    iter_per_epoch = len(train_loader)
    for epoch in range(args.epochs):
        train_epoch(epoch, wandb)