train_gpt2.py

import os
import numpy as np
from dataclasses import dataclass
import inspect
import tiktoken
import torch
import torch.nn as nn
from torch.nn import functional as F
from utils import Logger, VariableTracker
from hellaswag import HellaSwagEvaluator
from lambada_openai import LambadaEvaluator
from storycloze import StoryClozeEvaluator

class CausalSelfAttention(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.n_embd % config.n_head == 0
        # key, query, value projections for all heads, but in a batch
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
        # output projection
        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
        self.c_proj.NANOGPT_SCALE_INIT = 1 # 用于除以一个缩放因子
        # regularization
        self.n_head = config.n_head
        self.n_embd = config.n_embd

    def forward(self, x):
        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
        # nh is "number of heads", hs is "head size", and C (number of channels) = nh * hs
        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer
        qkv = self.c_attn(x)
        q, k, v = qkv.split(self.n_embd, dim=2)
        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        y = F.scaled_dot_product_attention(q, k, v, is_causal=True) # flash attention
        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
        # output projection
        y = self.c_proj(y)
        return y


class MLP(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd)
        self.gelu = nn.GELU(approximate='tanh')
        self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd)
        self.c_proj.NANOGPT_SACLE_INIT = 1 # 用于除以一个缩放因子

    def forward(self, x):
        x = self.c_fc(x)
        x = self.gelu(x)
        x = self.c_proj(x)
        return x

class Block(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.ln_1 = nn.LayerNorm(config.n_embd)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = nn.LayerNorm(config.n_embd)
        self.mlp = MLP(config)
    
    def forward(self, x):
        x = x + self.attn(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x

@dataclass
class GPTConfig:
    block_size: int = 2048 # max sequence length
    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token
    n_layer: int = 12 # number of layers
    n_head: int = 12 # number of heads
    n_embd: int = 768 # embedding dimension

class GPT(nn.Module):
    
    def __init__(self, config):
        super().__init__()
        self.config = config

        self.transformer = nn.ModuleDict(dict(
            wte = nn.Embedding(config.vocab_size, config.n_embd),# token embedding
            wpe = nn.Embedding(config.block_size, config.n_embd),# position embedding
            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),# 将Block模块重复n_layer次
            ln_f = nn.LayerNorm(config.n_embd),
        ))
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)# 最后的线性输出层
        
        # 权重共享(transformer原文： In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [ 24 ])
        self.transformer.wte.weight = self.lm_head.weight

        # 权重初始化
        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            std = 0.02
            if hasattr(module, 'NANOGPT_SCALE_INIT'):
                std *= (2 * self.config.n_layer)**-0.5
            torch.nn.init.normal_(module.weight, mean=0.0, std=std)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, idx, targets=None):
        # idx is of shape (B, T)
        B, T = idx.size()
        assert T <= self.config.block_size, f"Cannot forward sequence of length {T}, block size is only {self.config.block_size}"
        # forward the token and posisition embeddings
        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)
        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)
        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
        x = tok_emb + pos_emb
        # forward the blocks of the transformer
        for block in self.transformer.h:
            x = block(x)
        # forward the final layernorm and the classifier
        x = self.transformer.ln_f(x)
        logits = self.lm_head(x) # (B, T, vocab_size)
        loss = None
        if targets is not None:
            # logits.view(-1, logits.size(-1))：将所有批次的logits展平，然后和targets展平
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
        return logits, loss

    @classmethod
    def from_pretrained(cls, model_type, model_path):
        """Loads pretrained GPT-2 model weights from huggingface"""
        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
        from transformers import GPT2LMHeadModel
        logger.mes("loading weights from pretrained gpt: %s" % model_type)

        # n_layer, n_head and n_embd are determined from model_type
        config_args = {
            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
        }[model_type]
        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
        # create a from-scratch initialized minGPT model
        config = GPTConfig(**config_args)
        model = GPT(config)
        sd = model.state_dict()
        sd_keys = sd.keys()
        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param

        # init a huggingface/transformers model
        model_hf = GPT2LMHeadModel.from_pretrained(model_path)
        sd_hf = model_hf.state_dict()

        # copy while ensuring all of the parameters are aligned and match in names and shapes
        sd_keys_hf = sd_hf.keys()
        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
        # this means that we have to transpose these weights when we import them
        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
        for k in sd_keys_hf:
            if any(k.endswith(w) for w in transposed):
                # special treatment for the Conv1D weights we need to transpose
                assert sd_hf[k].shape[::-1] == sd[k].shape
                with torch.no_grad():
                    sd[k].copy_(sd_hf[k].t())
            else:
                # vanilla copy over the other parameters
                assert sd_hf[k].shape == sd[k].shape
                with torch.no_grad():
                    sd[k].copy_(sd_hf[k])

        return model

    def configure_optimizers(self, weight_decay, learning_rate, device_type):
        # start with all of the candidate parameters (that require grad)
        param_dict = {pn: p for pn, p in self.named_parameters()}
        param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}
        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
        optim_groups = [
            {'params': decay_params, 'weight_decay': weight_decay},
            {'params': nodecay_params, 'weight_decay': 0.0}
        ]
        num_decay_params = sum(p.numel() for p in decay_params)
        num_nodecay_params = sum(p.numel() for p in nodecay_params)
        logger.mes(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
        logger.mes(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
        # Create AdamW optimizer and use the fused version if it is available
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device_type == "cuda"
        logger.mes(f"using fused AdamW: {use_fused}")
        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=(0.9, 0.95), eps=1e-8, fused=use_fused)
        return optimizer

class DataLoaderLite:
    def __init__(self, B, T, process_rank, num_processes, split):
        self.B = B
        self.T = T
        self.process_rank = process_rank
        self.num_processes = num_processes
        assert split in {'train', 'val'}

        # get the shard filenames
        data_root = "edu_fineweb10B"
        shards = os.listdir(data_root)
        shards = [s for s in shards if split in s] # 只读取train或val
        shards = sorted(shards)
        shards = [os.path.join(data_root, s) for s in shards] # 得到数据读取的地址
        self.shards = shards
        assert len(shards) > 0, f"no shards found for split {split}"
        logger.mes(f"found {len(shards)} shards for split {split}")
        self.reset()

    def _load_tokens(self, filename):
        npt = np.load(filename)
        npt = npt.astype(np.int32) # added after video
        ptt = torch.tensor(npt, dtype=torch.long)
        return ptt

    def reset(self):
        # state, init at shard zero
        self.current_shard = 0
        self.tokens = self._load_tokens(self.shards[self.current_shard])
        self.current_position = self.B * self.T * self.process_rank # self.process_rank取值0-7

    def next_batch(self):
        B, T = self.B, self.T
        buf = self.tokens[self.current_position : self.current_position+B*T+1]
        x = (buf[:-1]).view(B, T) # inputs
        y = (buf[1:]).view(B, T) # targets
        # advance the position in the tensor
        self.current_position += B * T * self.num_processes
        # if loading the next batch would be out of bounds, advance to next shard
        if self.current_position + (B * T * self.num_processes + 1) > len(self.tokens):
            self.current_shard = (self.current_shard + 1) % len(self.shards)
            self.tokens = self._load_tokens(self.shards[self.current_shard])
            self.current_position = B * T * self.process_rank
        return x, y

#--------------------------------------------------------------------------------
# run the training loop
# simple launch:
# python train_gpt2.py
# DDP launch for e.g. 8 GPUs:
# torchrun --standalone --nproc_per_node=8 train_gpt2.py

import time
import math
from datetime import datetime
from torch.distributed import init_process_group, destroy_process_group
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist

# set up DDP (distributed data parallel).
# torchrun command sets the env variables RANK, LOCAL_RANK, and WORLD_SIZE
ddp = int(os.environ.get('RANK', -1)) != -1 # is this a ddp run?
if ddp:
    # os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"  #（保证程序cuda序号与实际cuda序号对应）
    # os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3" # 使用第几个cuda
    # use of DDP atm demands CUDA, we set the device appropriately according to rank
    assert torch.cuda.is_available(), "for now i think we need CUDA for DDP"
    init_process_group(backend='nccl') # 初始化进程组，使用nccl作为后端。
    ddp_rank = int(os.environ['RANK']) # 进程号
    ddp_local_rank = int(os.environ['LOCAL_RANK']) # 本地进程号
    ddp_world_size = int(os.environ['WORLD_SIZE']) # 总进程数
    device = f'cuda:{ddp_local_rank}' # 使进程编号与cuda设备编号对应
    torch.cuda.set_device(device)
    master_process = ddp_rank == 0 # this process will do logging, checkpointing etc.
else:
    # vanilla, non-DDP run
    ddp_rank = 0
    ddp_local_rank = 0
    ddp_world_size = 1
    master_process = True

    # 注意：这种方式会将cuda:x映射到cuda:0，从而导致查询到的cuda还是cuda:0，而实际使用的是cuda:x
    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"  #（保证程序cuda序号与实际cuda序号对应）
    os.environ["CUDA_VISIBLE_DEVICES"] = "4" # 使用第几个cuda
    device = "cuda"

# RECORD_CACHE_DIR 是记录训练过程的文件夹，其中包含每次训练的日志 权重文件 变量
# RECORD_DIR 是当前训练循环的记录文件夹
RECORD_CACHE_DIR = os.path.join(os.path.dirname(__file__), "gpt2_training_record")
RECORD_DIR = os.path.join(RECORD_CACHE_DIR, datetime.now().strftime("%Y-%m-%d_%H-%M-%S"))
if master_process:
    os.makedirs(RECORD_CACHE_DIR, exist_ok=True)
    os.makedirs(os.path.join(RECORD_DIR, "checkpoints"), exist_ok=True)
    os.makedirs(os.path.join(RECORD_DIR, "variables"), exist_ok=True)

HELLASWAG_CACHE_DIR = os.path.join(os.path.dirname(__file__), "hellaswag")
HELLASWAG_PATH = os.path.join(HELLASWAG_CACHE_DIR, f"hellaswag_val.jsonl")

STORYCLOZE_CACHE_DIR = os.path.join(os.path.dirname(__file__), "StoryCloze")
STORYCLOZE_PATH = os.path.join(STORYCLOZE_CACHE_DIR, "cloze_test_val__winter2018-cloze_test_ALL_val - 1 - 1.csv")

LAMBADA_CACHE_DIR = os.path.join(os.path.dirname(__file__), "lambada_openai")
LAMBADA_PATH = os.path.join(LAMBADA_CACHE_DIR, "lambada_test_en.jsonl")

logger = Logger(log_directory=RECORD_DIR, use_logging=master_process) # 只有主进程进行日志记录
vt = VariableTracker(keys=['step', 'train_loss','val_loss','hellaswg_acc', 'storycloze_acc', 'lambada_openai','lr'], master_process=master_process) # 只有主进程进行数据添加

# added after video, pytorch can be serious about it's device vs. device_type distinction
# device是设备，device_type是设备类型，device_type被用于autocast中
device_type = "cuda" if device.startswith("cuda") else "cpu"

torch.manual_seed(1337)
torch.cuda.manual_seed(1337)

enc = tiktoken.get_encoding("gpt2")

total_batch_size = 524288 # 2**19, ~0.5M, in number of tokens
B = 2 # micro batch size
T = 2048 # sequence length
assert total_batch_size % (B * T * ddp_world_size) == 0, "make sure total_batch_size is divisible by B * T * ddp_world_size"
grad_accum_steps = total_batch_size // (B * T * ddp_world_size)
logger.mes(f"total batch size: {total_batch_size}")
logger.mes(f"gradient accumulation steps: {grad_accum_steps}")

train_loader = DataLoaderLite(B=B, T=T, process_rank=ddp_rank, num_processes=ddp_world_size, split='train')
val_loader = DataLoaderLite(B=B, T=T, process_rank=ddp_rank, num_processes=ddp_world_size, split="val")

torch.set_float32_matmul_precision('high')

# 获得logits
model = GPT(GPTConfig(vocab_size=50304))
model.to(device)
use_compile = False # torch.compile interferes with HellaSwag eval and Generation. TODO fix
if use_compile:
    model = torch.compile(model)
if ddp:
    model = DDP(model, device_ids=[ddp_local_rank])
raw_model = model.module if ddp else model # model经过DDP封装后，取出module

max_lr = 6e-4
min_lr = max_lr * 0.1
warmup_steps = 715
# 10B tokens，每一个 step 是 0.5M tokens，19073个 step 刚好可以遍历完 10B tokens
max_steps = 19073 * 4 # 19,073 steps is ~1 epoch, if data is 10B tokens and batch size 0.5M tokens
def get_lr(it):
    # 1) 预热期，当iter数小于warmup_steps时，lr线性增加
    if it < warmup_steps:
        return max_lr * (it+1) / warmup_steps 
    # 2) 稳定期，当iter数大于max_steps时，返回最小lr
    if it > max_steps:
        return min_lr
    # 3) 中间态，使用余弦退火算法
    decay_ratio = (it - warmup_steps) / (max_steps - warmup_steps)
    assert 0 <= decay_ratio <= 1
    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff starts at 1 and goes to 0
    return min_lr + coeff * (max_lr - min_lr)

# optimize
# optimizer = torch.optim.AdamW(model.parameters(), lr=3e-4, betas=(0.9, 0.95),eps=1e-8)
optimizer = raw_model.configure_optimizers(weight_decay=0.1, learning_rate=max_lr, device_type=device)

for step in range(max_steps):
    vt.append("step", step)
    t0 = time.time()
    last_step = (step == max_steps - 1)

    # once in a while evaluate our validation loss
    if step % 250 == 0 or last_step:
        model.eval()
        val_loader.reset()
        with torch.no_grad():
            val_loss_accum = 0.0
            val_loss_steps = 20
            for _ in range(val_loss_steps):
                x, y = val_loader.next_batch()
                x, y = x.to(device), y.to(device)
                with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
                    logits, loss = model(x, y)
                loss = loss / val_loss_steps
                val_loss_accum += loss.detach()
        if ddp:
            dist.all_reduce(val_loss_accum, op=dist.ReduceOp.AVG)
        if master_process:
            logger.mes(f"validation loss: {val_loss_accum.item():.4f}")
            vt.append("val_loss", val_loss_accum.item())
            if step > 0 and (step % 5000 == 0 or last_step):
                # optionally write model checkpoints
                checkpoint_path = os.path.join(RECORD_DIR, "checkpoints", f"model_{step:05d}.pt")
                checkpoint = {
                    'model': raw_model.state_dict(),
                    'config': raw_model.config,
                    'step': step,
                    'val_loss': val_loss_accum.item()
                }
                # you might also want to add optimizer.state_dict() and
                # rng seeds etc., if you wanted to more exactly resume training
                torch.save(checkpoint, checkpoint_path)
                vt.save(os.path.join(RECORD_DIR, "variables", f"variables_{step:05d}.npy"))
    else:
        vt.append("val_loss", None)

    # evaluate hellaswag-------------------------------------
    if (step % 250 == 0 or last_step) and (not use_compile):
        hellaswag_evaluator = HellaSwagEvaluator()
        num_correct_norm = 0
        num_total = 0
        for i, example in enumerate(hellaswag_evaluator.iterate_examples(HELLASWAG_PATH)):
            # only process examples where i % ddp_world_size == ddp_rank
            if i % ddp_world_size != ddp_rank:
                continue
            # render the example into tokens and labels
            _, tokens, mask, label = hellaswag_evaluator.render_example(example, enc)
            tokens = tokens.to(device)
            mask = mask.to(device)
            # get the logits
            with torch.no_grad():
                with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
                    logits, _ = model(tokens)
                _, pred_norm = hellaswag_evaluator.evaluate_in_training(tokens, mask, logits)
            num_total += 1
            num_correct_norm += int(pred_norm == label)
        # reduce the stats across all processes
        if ddp:
            num_total = torch.tensor(num_total, dtype=torch.long, device=device)
            num_correct_norm = torch.tensor(num_correct_norm, dtype=torch.long, device=device)
            dist.all_reduce(num_total, op=dist.ReduceOp.SUM)
            dist.all_reduce(num_correct_norm, op=dist.ReduceOp.SUM)
            num_total = num_total.item()
            num_correct_norm = num_correct_norm.item()
        acc_norm = num_correct_norm / num_total
        logger.mes(f"HellaSwag accuracy: {num_correct_norm}/{num_total}={acc_norm:.4f}")
        vt.append("hellaswg_acc", acc_norm)
    else:
        vt.append("hellaswg_acc", None)
    
    # evaluate storycloze-------------------------------------
    if (step % 250 == 0 or last_step) and (not use_compile):
        storycloze_evaluator = StoryClozeEvaluator()
        num_correct_norm = 0
        num_total = 0
        for i, example in enumerate(storycloze_evaluator.iterate_examples(STORYCLOZE_PATH)):
            if i % ddp_world_size != ddp_rank:
                continue
            tokens, mask, label = storycloze_evaluator.render_example(example, enc)
            tokens, mask = tokens.to(device), mask.to(device)
            label = label - 1 # label只有1和2，使其从0开始
            with torch.no_grad():
                with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
                    logits, _ = model(tokens)
                _, pred_norm = storycloze_evaluator.evaluate_in_training(tokens, mask, logits)
            num_total += 1
            num_correct_norm += int(pred_norm == label)
        if ddp:
            num_total = torch.tensor(num_total, dtype=torch.long, device=device)
            num_correct_norm = torch.tensor(num_correct_norm, dtype=torch.long, device=device)
            dist.all_reduce(num_total, op=dist.ReduceOp.SUM)
            dist.all_reduce(num_correct_norm, op=dist.ReduceOp.SUM)
            num_total = num_total.item()
            num_correct_norm = num_correct_norm.item()
        acc_norm = num_correct_norm / num_total
        logger.mes(f"StoryCloze accuracy: {num_correct_norm}/{num_total}={acc_norm:.4f}")
        vt.append("storycloze_acc", acc_norm)
    else:
        vt.append("storycloze_acc", None)

    # evaluate lambada_openai-----------------------------------
    if (step % 250 == 0 or last_step) and (not use_compile):
        lambada_evaluator = LambadaEvaluator()
        num_correct = 0
        num_total = 0
        for i, example in enumerate(lambada_evaluator.iterate_examples(LAMBADA_PATH)):
            if i % ddp_world_size != ddp_rank:
                continue
            tokens = lambada_evaluator.render_example(example, enc)
            tokens = tokens.to(device)
            with torch.no_grad():
                with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
                    logits, _ = model(tokens)
                predicted_tokens = lambada_evaluator.evaluate_in_training(logits)
            num_total += 1
            num_correct += int(predicted_tokens == tokens.view(-1)[-1])
        if ddp:
            num_total = torch.tensor(num_total, dtype=torch.long, device=device)
            num_correct = torch.tensor(num_correct, dtype=torch.long, device=device)
            dist.all_reduce(num_total, op=dist.ReduceOp.SUM)
            dist.all_reduce(num_correct, op=dist.ReduceOp.SUM)
            num_total = num_total.item()
            num_correct = num_correct.item()
        acc = num_correct / num_total
        logger.mes(f"Lambada_openai accuracy: {num_correct}/{num_total}={acc:.4f}")
        vt.append("lambada_openai", acc)
    else:
        vt.append("lambada_openai", None)

    # do one step of the optimization
    model.train()
    optimizer.zero_grad() # 梯度清零
    loss_accum = 0.0
    for micro_step in range(grad_accum_steps):
        x, y = train_loader.next_batch()
        x, y = x.to(device), y.to(device)
        if ddp:
            model.require_backward_grad_sync = (micro_step == grad_accum_steps-1) # 
        # 使用混合精度进行训练
        with torch.autocast(device_type=device_type,dtype=torch.bfloat16):
            logits, loss = model(x, y)
        loss = loss/grad_accum_steps
        loss_accum += loss.detach()
        loss.backward() # 计算梯度
    if ddp:
        dist.all_reduce(loss_accum, op=dist.ReduceOp.AVG) # 在分布式计算环境中将所有进程的loss_accum求平均值，并将结果放回每个进程
    norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
    lr = get_lr(step) # 动态修改学习率
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    optimizer.step()
    torch.cuda.synchronize() # wait for the GPU to finish work
    t1 = time.time()
    dt = (t1-t0) # 计算单个epoch耗时（单位秒）
    tokens_processed = train_loader.B * train_loader.T * grad_accum_steps * ddp_world_size
    tokens_per_sec = tokens_processed / dt
    logger.mes(f"step{step:5d} | loss: {loss_accum.item():.8f} | lr: {lr:.4e} | norm: {norm:.4f} | tok/sec: {tokens_per_sec:.2f}")
    vt.append("train_loss", loss_accum.item())
    vt.append("lr", lr)

if ddp:
    destroy_process_group()
# 退出程序
import sys; sys.exit(0)