src/mRNA2vec/train_distillation_pt.py

# torchrun --nproc_per_node 8 train_distillation_pt.py --limit=-1 --batch_size=32 --n_layers=4 --use_wandb --ddp --local_rank=0 --epochs=100
import os
from collections import defaultdict

import numpy as np


# Load modules
# env | grep -E "PATH|LIBRARY_PATH|LD_LIBRARY_PATH"
# Your code here
username = os.environ['HOME']
import argparse
import time
import math
import warnings

# import pandas as pd
import numpy as np
# np.float = float
# np.int = int   #module 'numpy' has no attribute 'int'
# np.object = object    #module 'numpy' has no attribute 'object'
# np.bool = bool    #module 'numpy' has no attribute 'bool'

# 获取脚本所在的目录
script_dir = os.path.dirname(os.path.abspath(__file__))

# 切换到脚本所在的目录
os.chdir(script_dir)

import torch
import torch.nn.functional as F
import torch.distributed as dist

from torch import optim, nn
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data import DataLoader, DistributedSampler, random_split
from contextlib import nullcontext
from model.model_ribo import MiniMindLM
from model.LMConfig import LMConfig
from model.tools import EarlyStopping, get_pretraining_args,get_dataset_args
from utils.ernie_rna.dictionary import Dictionary
from utils.ernie_rna.dataset import RNADataset

from src.utils import load_pretrained_ernierna
from src.mRNA2vec.model import mRNA2vec, T5_encoder
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 distillation_loss_fn(student_logits, teacher_logits, temperature=1.0, reduction='batchmean'):
    with torch.no_grad():
        teacher_probs = F.softmax(teacher_logits / temperature, dim=-1).detach()

    student_log_probs = F.log_softmax(student_logits / temperature, dim=-1)

    kl = F.kl_div(
        student_log_probs,
        teacher_probs,
        reduction=reduction
    )
    return (temperature ** 2) * kl


def train_epoch(epoch, wandb, alpha=0.0, temperature=1.0):
    start_time = time.time()
    if teacher_model is not None:
        teacher_model.eval()
        teacher_model.requires_grad_(False)
    model.train()
    for step, (src_data,tgt_data,twod_data,loss_mask) in enumerate(train_loader):
        # print(f'train step:{step}/{len(train_loader)}')
        src_data, tgt_data, twod_data, loss_mask = src_data.to(args.device),tgt_data.to(args.device),twod_data.to(args.device),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

        # 前向传播（学生模型）
        with ctx:
            res = model(src_data,twod_data)
            student_logits = res.logits[...,:-1] # drop mask

        # 教师模型前向传播（只在eval & no_grad）
        if teacher_model is not None:
            with torch.no_grad():
                # teacher_logits = teacher_model(src_data,twod_data,is_twod=args.is_twod).logits
                if "ernierna" in  args.mlm_pretrained_model_path:
                    teacher_logits,attn_map_lst,out_dict = teacher_model(src_data,twod_data,is_twod=args.is_twod)
                    # teacher_logits = out_dict['sentence_logits'] # [11,1205, 25]
                    # teacher_logits = teacher_logits#[:,:,4:8] # [11,1205, 25]
                    vocab_size_student = student_logits.size(-1)  # N 10
                    teacher_logits = teacher_logits[..., :vocab_size_student] # [11*1205, 25]
                elif "mrna2vec" in args.mlm_pretrained_model_path:
                    teacher_logits = teacher_model(src_data)
                    vocab_size_student = student_logits.size(-1)
                    teacher_logits = teacher_logits[..., :vocab_size_student]  # [11*1205, 25]

        # ========== 计算损失 ==========
        # 1) Ground-Truth CE Loss（可选）
        loss_mask_flat = loss_mask.view(-1)
        ce_loss = F.cross_entropy( # 隐含了softmax的计算
            student_logits.view(-1, student_logits.size(-1)),
            tgt_data.view(-1),
            ignore_index=0,
            reduction='none'
        )
        ce_loss = torch.sum(ce_loss * loss_mask_flat) / loss_mask_flat.sum()
        if lm_config_student.use_moe:
            ce_loss += res.aux_loss

        # 2) Distillation Loss（可选）
        if teacher_model is not None:
            # 只在有效token位置做蒸馏
            distill_loss = distillation_loss_fn(
                student_logits.reshape(-1, student_logits.size(-1))[loss_mask_flat == 1],
                teacher_logits.reshape(-1, teacher_logits.size(-1))[loss_mask_flat == 1], # [1394, 9] mask token
                temperature=temperature
            )
        else:
            distill_loss = torch.tensor(0.0, device=args.device)

        # 3) 总损失 = alpha * CE + (1-alpha) * Distill
        loss = alpha * ce_loss + (1 - alpha) * distill_loss

        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 or args.debug) and (not ddp or dist.get_rank() == 0):
            spend_time = time.time() - start_time
            ans = {
                    "tr_loss": loss.item(),
                    "tr_ce_loss": ce_loss.item(),
                    "tr_distill_loss": distill_loss.item() if teacher_model is not None else 0.0,
                    "tr_lr": optimizer.param_groups[-1]['lr'],
                    "tr_step_time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60
                }
            Logger(
                'Epoch:[{}/{}][step:{}/{}] tr_loss:{:.4f} tr_ce_loss:{:.4f} tr_distill_loss:{:.4f} tr_step_time:{} min:'.format(
                    epoch,
                    args.epochs - 1,
                    step,train_loader.__len__(),
                    ans['tr_loss'],
                    ans['tr_ce_loss'],
                    ans['tr_distill_loss'],
                    ans['tr_step_time']
                )
            )
            if (wandb is not None) and (not ddp or dist.get_rank() == 0):
                wandb.log(ans)

        if (step+1) % args.save_interval == 0 and (not ddp or dist.get_rank() == 0):
            early_stopping.save_model(model, ckp.replace('epoch','step'))
            model.train()
        # print('end of a train step')
def val_epoch(epoch, wandb, alpha=0.0, temperature=1.0):
    start_time = time.time()
    if teacher_model is not None:
        teacher_model.eval()
        teacher_model.requires_grad_(False)
    result = defaultdict(list)
    # for step, (X, Y, loss_mask) in enumerate(train_loader):
    for step, (src_data,tgt_data,twod_data,loss_mask) in enumerate(val_loader):
        # print(f'val step:{step}/{len(val_loader)}')
        src_data, tgt_data, twod_data, loss_mask = src_data.to(args.device),tgt_data.to(args.device),twod_data.to(args.device),loss_mask.to(args.device)

        # 前向传播（学生模型）
        with ctx:
            with torch.no_grad():
                res = model(src_data,twod_data)
                student_logits = res.logits[...,:-1] # drop mask

        # 教师模型前向传播（只在eval & no_grad）
        if teacher_model is not None:
            with torch.no_grad():
                # teacher_logits = teacher_model(src_data,twod_data,is_twod=args.is_twod).logits
                teacher_logits,attn_map_lst,out_dict = teacher_model(src_data,twod_data,is_twod=args.is_twod)
                # teacher_logits = out_dict['sentence_logits'] # [11,1205, 25]
                # teacher_logits = teacher_logits#[:,:,4:8] # [11,1205, 25]
                vocab_size_student = student_logits.size(-1)  # N 10
                teacher_logits = teacher_logits[..., :vocab_size_student] # [11*1205, 25]

        # ========== 计算损失 ==========
        # 1) Ground-Truth CE Loss（可选）
        loss_mask_flat = loss_mask.view(-1)
        ce_loss = F.cross_entropy( # 隐含了softmax的计算
            student_logits.view(-1, student_logits.size(-1)),
            tgt_data.view(-1),
            ignore_index=0,
            reduction='none'
        )
        ce_loss = torch.sum(ce_loss * loss_mask_flat) / loss_mask_flat.sum()
        if lm_config_student.use_moe:
            ce_loss += res.aux_loss

        # 2) Distillation Loss（可选）
        if teacher_model is not None:
            # 只在有效token位置做蒸馏
            distill_loss = distillation_loss_fn(
                student_logits.reshape(-1, student_logits.size(-1))[loss_mask_flat == 1],
                teacher_logits.reshape(-1, teacher_logits.size(-1))[loss_mask_flat == 1], # [1394, 9] mask token
                temperature=temperature
            )
        else:
            distill_loss = torch.tensor(0.0, device=args.device)

        # 3) 总损失 = alpha * CE + (1-alpha) * Distill
        loss = alpha * ce_loss + (1 - alpha) * distill_loss
        spend_time = time.time() - start_time
        rs = {
                "val_loss": loss.item(),
                "val_ce_loss": ce_loss.item(),
                "val_distill_loss": distill_loss.item() if teacher_model is not None else 0.0,
                "val_lr": optimizer.param_groups[-1]['lr'],
                "val_step_time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60
            }
        [result[key].append(value) for key,value in rs.items()]
    ans = {key:np.array(result[key]).mean() for key,value in result.items()}
    # print('dist.get_rank()',dist.get_rank())
    if not ddp or dist.get_rank() == 0:
        Logger(
            'Epoch:[{}/{}] val_loss:{:.4f} val_ce_loss:{:.4f} val_distill_loss:{:.4f} val_step_time:{} min:'.format(
                epoch,
                args.epochs - 1,
                ans['val_loss'],
                ans['val_ce_loss'],
                ans['val_distill_loss'],
                ans['val_step_time']
            )
        )

    if (wandb is not None) and (not ddp or dist.get_rank() == 0):
        wandb.log(ans)
    # print('end of val epoch')
    return ans['val_loss'] #val_loss


def init_student_model():
    # tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
    vocab_path = args.arg_overrides['data'] + '/small_dict.txt'
    tokenizer = Dictionary.load(vocab_path)
    tokenizer.mask_index = tokenizer.add_symbol('<mask>')
    # ['<s>', '<pad>', '</s>', '<unk>', 'G', 'A', 'U', 'C', 'N', '<mask>']
    if args.debug:args.n_layers = 1
    lm_config = LMConfig(dim=256, n_layers=args.n_layers, max_seq_len=max_seq_len, vocab_size=len(tokenizer),padding_idx=tokenizer.pad_index) # n_layers 8, <s> <unk><unk><unk> </s>
    model = MiniMindLM(lm_config)
    # moe_path = '_moe' if lm_config.use_moe else ''
    # ckp = f'./out/full_sft_{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'学生模型(LLM)总参数量：{sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f} 百万,vocab_size={len(tokenizer)}')
    model = model.to(args.device)
    print(model)
    return model, tokenizer,lm_config

# def load_pretrained_ernierna(mlm_pretrained_model_path,arg_overrides):
#     rna_models, _, _ = checkpoint_utils.load_model_ensemble_and_task(mlm_pretrained_model_path.split(os.pathsep),arg_overrides=arg_overrides)
#     model_pretrained = rna_models[0]
#     return model_pretrained
def init_teacher_model(args,Logger=None):
    # model = MiniMindLM(lm_config)
    # moe_path = '_moe' if lm_config.use_moe else ''
    # ckp = args.mlm_pretrained_model_path
    vocab_path = args.arg_overrides['data'] + '/dict.txt'
    tokenizer = Dictionary.load(vocab_path)
    tokenizer.add_symbol('<mask>')
    # tokenizer = None
    if 'ernierna' in args.mlm_pretrained_model_path:
        model_pre = load_pretrained_ernierna(args.mlm_pretrained_model_path, args.arg_overrides)
        model = model_pre.encoder

        if args.debug:
            print('debug mode')
            num_layers_to_keep = 1  # 保留12层,todo
            model.sentence_encoder.layers = model.sentence_encoder.layers[
                                                                :num_layers_to_keep]
        # torch.save(model,args.save_dir+'/pretraining0215.pt')
        # print('save small ERNIE-RNA model in',args.save_dir+'/pretraining0215.pt')
        # state_dict = torch.load(ckp, map_location=args.device)
        # model.load_state_dict(state_dict, strict=False)
    elif "mrna2vec" in args.mlm_pretrained_model_path:
        encoder = T5_encoder(
            hidden_size=256,
            num_attention_heads=4,
            num_hidden_layers=4,
        )
        model = mRNA2vec(encoder=encoder)
        ckpt = torch.load(args.mlm_pretrained_model_path, map_location=args.device)
        model.encoder.load_state_dict(ckpt['encoder'], strict=True)

    Logger(f'教师模型(LLM)总参数量：{sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f} 百万, vocab_size={len(tokenizer)}')
    model = model.to(args.device)
    print(model)
    return model,tokenizer


def init_distributed_mode():
    print("init distributed mode,ddp=",ddp)

    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__':
    # 获取 pretraining 和 dataset 的 args
    os.chdir(os.path.dirname(os.path.abspath(__file__)))
    pretraining_parser = get_pretraining_args()
    dataset_parser = get_dataset_args()

    # 合并 args
    parser = argparse.ArgumentParser(parents=[pretraining_parser, dataset_parser], add_help=False,conflict_handler='resolve')
    args = parser.parse_args()

    torch.manual_seed(1337)
    device_type = "cuda" if "cuda" in args.device else "cpu"
    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)
    Logger(f'loading model')

    # 定义学生模型和教师模型
    max_seq_len = args.region*4+5

    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

    # 初始化学生模型和教师模型
    # teacher_model = init_teacher_model(lm_config_teacher)
    teacher_model,tokenizer = init_teacher_model(args,Logger)
    # teacher_model, tokenizer = init_student_model(lm_config_student)
    model, tokenizer,lm_config_student = init_student_model()

    if args.debug:args.limit=320
    # train_ds = SFTDataset(args.data_path, tokenizer, max_length=max_seq_len)
    dataset = RNADataset(args.ffasta, region=args.region,tokenizer=tokenizer,limit=args.limit)

    train_size = int(0.99 * len(dataset))  # 99% 用于训练
    val_size = len(dataset) - train_size  # 1% 用于验证
    train_dataset, val_dataset = random_split(dataset, [train_size, val_size])

    train_sampler = DistributedSampler(train_dataset) if ddp else None
    Logger(f'loading {train_size} training samples from',args.ffasta)
    Logger(f'loading {val_size} validation samples from',args.ffasta)

    train_loader = DataLoader(
        train_dataset,
        batch_size=args.batch_size,
        pin_memory=True,
        drop_last=False,
        shuffle=False,
        num_workers=args.num_workers,
        sampler=train_sampler
    )

    val_loader = DataLoader(
        val_dataset,
        batch_size=args.batch_size,
        pin_memory=True,
        drop_last=False,
        shuffle=False,
        num_workers=args.num_workers
    )

    args.wandb_run_name = f"{args.wandb_project}-EP-{args.epochs}-BS-{args.batch_size}-LR-{args.learning_rate}-{len(train_loader.dataset)/ 1e6:.0f}"
    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,mode="offline")
    else:
        wandb = None
    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])
    Logger('local_rank',args.local_rank,'ddp',ddp,'ddp_local_rank',ddp_local_rank)

    iter_per_epoch = len(train_loader)
    epoch = 0

    # 定义 EarlyStopping 实例
    moe_path = '_moe' if lm_config_student.use_moe else ''
    ckp = f'{args.save_dir}/full_dist_{lm_config_student.dim}{moe_path}_epoch.pth'
    Logger(f'save model to', os.path.abspath(ckp))
    early_stopping = EarlyStopping(patience=5, verbose=True, path=ckp)
    for epoch in range(args.epochs):
        if ddp: train_loader.sampler.set_epoch(epoch)
        # print(f'start training epoch: {epoch}/{args.epochs}')
        train_epoch(epoch, wandb,alpha=args.celoss_alpha)
        current_loss = val_epoch(epoch, wandb, alpha=args.celoss_alpha)
        if ddp:
            # 分布式训练逻辑
            if ddp_local_rank == 0:
                early_stopping(current_loss, model)  # 如果监控的是SPR，直接传入-SPR即可
                # 广播 should_stop 的值到其他进程
                to_broadcast = torch.tensor([early_stopping.early_stop], dtype=torch.float32, device=args.device)
                dist.broadcast(to_broadcast, 0)
            else:
                # 非主进程等待主进程广播
                to_broadcast = torch.tensor([False], dtype=torch.float32, device=args.device)
                dist.broadcast(to_broadcast, 0)
                early_stopping.early_stop = bool(to_broadcast.item())
        else:
            # 单机单卡训练逻辑
            early_stopping(current_loss, model)  # 如果监控的是SPR，直接传入-SPR即可
        if early_stopping.early_stop:break

    print('the end')