-
导入模块：导入了一些必要的模块，包括数学计算、时间处理、文件操作、深度学习框架（如
torch）、以及自定义的LLama Transformer模型相关内容。 -
I/O配置：定义了模型输出路径、评估与日志记录的间隔步数、批次大小、最大序列长度、词汇表大小等。
-
模型配置：设置了模型的隐藏层维度、层数、注意力头的数量、分组数量、Dropout概率等超参数。
-
AdamW优化器配置：配置了优化器的学习率、权重衰减系数、β1和β2参数、梯度累积步数等，使用梯度累积来模拟更大的批次训练。
-
学习率调度：通过线性预热和余弦衰减方式动态调整学习率，保证训练过程中的学习率变化更加平滑。
-
设备与精度设置：判断设备类型（CPU或GPU），并配置混合精度（FP16）训练相关的上下文。
-
数据加载：定义了批次迭代器
iter_batches，每次从训练集和验证集中加载数据，并准备进行模型的前向和反向传播。 -
模型初始化：通过自定义的Transformer类初始化模型，并使用GradScaler处理混合精度训练时的梯度缩放。
-
损失评估函数
estimate_loss：定义了一个评估函数，在训练过程中定期评估训练和验证集上的损失，并根据损失决定是否保存模型检查点。 -
学习率获取函数
get_lr：实现了学习率的预热和衰减策略，根据当前的迭代步数动态调整学习率。 -
训练循环：
- 按照配置的批次大小从数据加载器中获取数据。
- 进行前向传播，计算损失。
- 进行梯度累积和反向传播。
- 使用梯度裁剪和优化器更新模型参数。
- 根据配置的间隔进行日志记录和损失评估，并保存模型检查点。
- 在迭代次数达到预设的最大值时终止训练。
-
日志记录：在训练过程中，程序会定期输出当前的训练步数、损失值、学习率、运行时间以及模型的浮点运算利用率（MFU，模型浮点计算的利用率，表示模型计算效率）。
-
检查点保存：当验证损失下降时，程序会保存模型的检查点，包括模型的状态、优化器的状态和当前的迭代信息，以便在需要时可以恢复训练。
import math
import os
import time
from contextlib import nullcontext
from datetime import datetime
from functools import partialimport torch
from LLama_content.llama_model import Transformer, ModelArgs
from LLama_content.llama_model import Task# -----------------------------------------------------------------------------
# I/O 配置，用于定义输出目录和训练时的日志记录与评估设置
out_dir = "output" # 模型输出保存路径
eval_interval = 2000 # 评估间隔步数
log_interval = 1 # 日志记录间隔步数
eval_iters = 100 # 每次评估时迭代的步数
eval_only = False # 如果为True，脚本在第一次评估后立即退出
always_save_checkpoint = False # 如果为True，在每次评估后总是保存检查点
init_from = "scratch" # 可以选择从头开始训练（'scratch'）或从已有的检查点恢复（'resume'）# 数据配置
batch_size = 128 # 每个微批次的样本数量，如果使用梯度累积，实际批次大小将更大
max_seq_len = 256 # 最大序列长度
vocab_size = 4096 # 自定义词汇表大小# 模型配置
dim = 288 # 模型的隐藏层维度
n_layers = 8 # Transformer的层数
n_heads = 8 # 注意力头的数量
n_group = 4 # 模型分组
multiple_of = 32 # 在某些层的维度必须是该数的倍数
dropout = 0.0 # Dropout概率# AdamW优化器配置
gradient_accumulation_steps = 4 # 梯度累积步数，用于模拟更大的批次
learning_rate = 5e-4 # 最大学习率
max_iters = 100000 # 总的训练迭代次数
weight_decay = 1e-1 # 权重衰减系数
beta1 = 0.9 # AdamW优化器的β1参数
beta2 = 0.95 # AdamW优化器的β2参数
grad_clip = 1.0 # 梯度裁剪阈值，0表示不裁剪# 学习率衰减配置
decay_lr = True # 是否启用学习率衰减
warmup_iters = 1000 # 学习率预热的步数# 系统设置
device = "cuda:0" # 设备选择：'cpu'，'cuda'，'cuda:0'等
dtype = "bfloat16" # 数据类型：'float32'，'bfloat16'，'float16'# -----------------------------------------------------------------------------
# 获取配置参数的键值对，便于后续的日志记录
config_keys = [kfor k, v in globals().items()if not k.startswith("_") and isinstance(v, (int, float, bool, str))
]
config = {k: globals()[k] for k in config_keys} # 保存配置到字典中，便于日志记录
# -----------------------------------------------------------------------------# 固定一些超参数的默认值
lr_decay_iters = max_iters # 学习率衰减步数，设置为等于最大迭代步数
min_lr = 0.0 # 最小学习率，建议为学习率的十分之一
vocab_source = 'custom' # 词汇表来源
master_process = True # 用于区分主进程
seed_offset = 0 # 随机种子偏移量
ddp_world_size = 1 # 分布式数据并行的世界大小
tokens_per_iter = batch_size * max_seq_len # 每次迭代处理的token数# 设置随机种子，确保可重复性
torch.manual_seed(1337 + seed_offset)
torch.backends.cuda.matmul.allow_tf32 = True # 允许在matmul上使用tf32
torch.backends.cudnn.allow_tf32 = True # 允许在cudnn上使用tf32
device_type = "cuda" if "cuda" in device else "cpu" # 用于自动选择设备类型
ptdtype = torch.float16 # 设置训练时使用的数据类型# 混合精度训练相关
ctx = (nullcontext()if device_type == "cpu"else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
)# task-specific setup
iter_batches = partial(Task.iter_batches,batch_size=batch_size,max_seq_len=max_seq_len,vocab_size=vocab_size,vocab_source=vocab_source,device=device,num_workers=0,
)iter_num = 0
best_val_loss = 1e9# model init
model_args = dict(dim=dim,n_layers=n_layers,n_heads=n_heads,n_group=n_group,vocab_size=vocab_size,multiple_of=multiple_of,max_seq_len=max_seq_len,dropout=dropout,
) # 模型参数初始化# ===========================================================
# 模型初始化
gptconf = ModelArgs(**model_args)
model = Transformer(gptconf)model.to(device)# initialize a GradScaler. If enabled=False scaler is a no-op
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == "float16"))# optimizer
optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)# 定义eval流程
@torch.no_grad()
def estimate_loss():out = {}model.eval()for split in ["train", "val"]:batch_iter = iter_batches(split=split)losses = torch.zeros(eval_iters) # keep on CPUfor k in range(eval_iters):X, Y = next(batch_iter)with ctx:logits = model(X, Y)loss = raw_model.last_losslosses[k] = loss.item()out[split] = losses.mean()model.train()return out# 定义学习率
def get_lr(it):# 1) linear warmup for warmup_iters stepsif it < warmup_iters:return learning_rate * it / warmup_iters# 2) if it > lr_decay_iters, return min learning rateif it > lr_decay_iters:return min_lr# 3) in between, use cosine decay down to min learning ratedecay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)assert 0 <= decay_ratio <= 1coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1return min_lr + coeff * (learning_rate - min_lr)# training loop
train_batch_iter = iter_batches(split="train")
X, Y = next(train_batch_iter) # fetch the very first batch
t0 = time.time()
local_iter_num = 0 # number of iterations in the lifetime of this process
raw_model = model # unwrap DDP container if needed
running_mfu = -1.0
while True:# 或许当前step的学习率lr = get_lr(iter_num) if decay_lr else learning_ratefor param_group in optimizer.param_groups:param_group["lr"] = lr# evaluate the loss on train/val sets and write checkpointsif iter_num % eval_interval == 0 and master_process:losses = estimate_loss()print(f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")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,}print(f"saving checkpoint to {out_dir}")torch.save(checkpoint, os.path.join(out_dir, "ckpt.pt"))if iter_num == 0 and eval_only:break# 前向更新过程，使用了梯度累积(检查点)for micro_step in range(gradient_accumulation_steps):with ctx: # 混合精度相关logits = model(X, Y)loss = raw_model.last_lossloss = loss / gradient_accumulation_stepsX, Y = next(train_batch_iter)# 反向传播scaler.scale(loss).backward()# 梯度阶段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 fp16scaler.step(optimizer)scaler.update()# flush the gradients as soon as we can, no need for this memory anymoreoptimizer.zero_grad(set_to_none=True)# timing and loggingt1 = time.time()dt = t1 - t0t0 = t1if iter_num % log_interval == 0 and master_process:# get loss as float, scale up due to the divide above. note: this is a CPU-GPU sync pointlossf = loss.item() * gradient_accumulation_stepsif local_iter_num >= 5: # let the training loop settle a bitmfu = 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 * mfuprint(f"{iter_num} | loss {lossf:.4f} | lr {lr:e} | {dt*1000:.2f}ms | mfu {running_mfu*100:.2f}%" # mfu表示模型浮点运算利用率)iter_num += 1local_iter_num += 1# termination conditionsif iter_num > max_iters:break0 | loss 8.3757 | lr 0.000000e+00 | 36663.83ms | mfu -100.00%
1 | loss 8.3732 | lr 5.000000e-07 | 1476.33ms | mfu -100.00%
2 | loss 8.3785 | lr 1.000000e-06 | 1700.56ms | mfu -100.00%
3 | loss 8.3699 | lr 1.500000e-06 | 1702.70ms | mfu -100.00%
4 | loss 8.3650 | lr 2.000000e-06 | 1697.50ms | mfu -100.00%
5 | loss 8.3470 | lr 2.500000e-06 | 1703.32ms | mfu 1.38%
6 | loss 8.3465 | lr 3.000000e-06 | 1704.92ms | mfu 1.38%
7 | loss 8.3269 | lr 3.500000e-06 | 1701.99ms | mfu 1.38%
8 | loss 8.3189 | lr 4.000000e-06 | 1700.86ms | mfu 1.38%
9 | loss 8.2904 | lr 4.500000e-06 | 1700.20ms | mfu 1.38%
10 | loss 8.2730 | lr 5.000000e-06 | 1702.60ms | mfu 1.38%
11 | loss 8.2581 | lr 5.500000e-06 | 1700.75ms | mfu 1.38%
12 | loss 8.2332 | lr 6.000000e-06 | 1703.53ms | mfu 1.38%
13 | loss 8.2065 | lr 6.500000e-06 | 1718.47ms | mfu 1.38%
14 | loss 8.1800 | lr 7.000000e-06 | 1707.27ms | mfu 1.38%
15 | loss 8.1510 | lr 7.500000e-06 | 1716.64ms | mfu 1.38%
16 | loss 8.1214 | lr 8.000000e-06 | 1721.91ms | mfu 1.38%
17 | loss 8.0870 | lr 8.500000e-06 | 1721.15ms | mfu 1.38%
18 | loss 8.0530 | lr 9.000000e-06 | 1721.06ms | mfu 1.38%
19 | loss 8.0241 | lr 9.500000e-06 | 1733.10ms | mfu 1.37%
20 | loss 7.9939 | lr 1.000000e-05 | 1729.06ms | mfu 1.37%
21 | loss 7.9736 | lr 1.050000e-05 | 1727.64ms | mfu 1.37%
22 | loss 7.9490 | lr 1.100000e-05 | 1734.65ms | mfu 1.37%
23 | loss 7.9211 | lr 1.150000e-05 | 1731.17ms | mfu 1.37%
24 | loss 7.8911 | lr 1.200000e-05 | 1731.48ms | mfu 1.37%
25 | loss 7.8686 | lr 1.250000e-05 | 1736.57ms | mfu 1.37%
26 | loss 7.8458 | lr 1.300000e-05 | 1726.87ms | mfu 1.37%
27 | loss 7.8293 | lr 1.350000e-05 | 1724.82ms | mfu 1.37%
28 | loss 7.8036 | lr 1.400000e-05 | 1720.08ms | mfu 1.37%
29 | loss 7.8019 | lr 1.450000e-05 | 1732.76ms | mfu 1.36%
30 | loss 7.7791 | lr 1.500000e-05 | 1730.16ms | mfu 1.36%
31 | loss 7.7619 | lr 1.550000e-05 | 1724.99ms | mfu 1.36%
32 | loss 7.7674 | lr 1.600000e-05 | 1741.93ms | mfu 1.36%
33 | loss 7.7393 | lr 1.650000e-05 | 1741.69ms | mfu 1.36%
34 | loss 7.7329 | lr 1.700000e-05 | 1739.45ms | mfu 1.36%
35 | loss 7.7254 | lr 1.750000e-05 | 1742.96ms | mfu 1.36%
36 | loss 7.7207 | lr 1.800000e-05 | 1744.78ms | mfu 1.36%
37 | loss 7.7090 | lr 1.850000e-05 | 1745.76ms | mfu 1.36%
38 | loss 7.7068 | lr 1.900000e-05 | 1755.41ms | mfu 1.36%
39 | loss 7.6913 | lr 1.950000e-05 | 1751.07ms | mfu 1.35%
40 | loss 7.6824 | lr 2.000000e-05 | 1760.08ms | mfu 1.35%
41 | loss 7.6860 | lr 2.050000e-05 | 1764.52ms | mfu 1.35%
42 | loss 7.6879 | lr 2.100000e-05 | 1759.85ms | mfu 1.35%
43 | loss 7.6856 | lr 2.150000e-05 | 1773.77ms | mfu 1.35%
44 | loss 7.6762 | lr 2.200000e-05 | 1778.94ms | mfu 1.34%
45 | loss 7.6797 | lr 2.250000e-05 | 1781.31ms | mfu 1.34%
46 | loss 7.6697 | lr 2.300000e-05 | 1772.70ms | mfu 1.34%
47 | loss 7.6638 | lr 2.350000e-05 | 1775.44ms | mfu 1.34%
48 | loss 7.6535 | lr 2.400000e-05 | 1778.56ms | mfu 1.34%
49 | loss 7.6516 | lr 2.450000e-05 | 1780.90ms | mfu 1.34%
50 | loss 7.6531 | lr 2.500000e-05 | 1776.91ms | mfu 1.33%
51 | loss 7.6512 | lr 2.550000e-05 | 1769.89ms | mfu 1.33%
52 | loss 7.6373 | lr 2.600000e-05 | 1775.77ms | mfu 1.33%
53 | loss 7.6428 | lr 2.650000e-05 | 1784.00ms | mfu 1.33%
54 | loss 7.6320 | lr 2.700000e-05 | 1780.11ms | mfu 1.33%
55 | loss 7.6242 | lr 2.750000e-05 | 1784.60ms | mfu 1.33%
56 | loss 7.6152 | lr 2.800000e-05 | 1788.55ms | mfu 1.33%
57 | loss 7.6133 | lr 2.850000e-05 | 1778.49ms | mfu 1.33%
58 | loss 7.6066 | lr 2.900000e-05 | 1779.38ms | mfu 1.33%
59 | loss 7.6035 | lr 2.950000e-05 | 1783.05ms | mfu 1.33%
60 | loss 7.5961 | lr 3.000000e-05 | 1786.11ms | mfu 1.33%
61 | loss 7.5854 | lr 3.050000e-05 | 1779.18ms | mfu 1.32%
62 | loss 7.5616 | lr 3.100000e-05 | 1781.52ms | mfu 1.32%
63 | loss 7.5642 | lr 3.150000e-05 | 1780.07ms | mfu 1.32%
64 | loss 7.5492 | lr 3.200000e-05 | 1783.43ms | mfu 1.32%
65 | loss 7.5362 | lr 3.250000e-05 | 1779.71ms | mfu 1.32%
66 | loss 7.5173 | lr 3.300000e-05 | 1783.60ms | mfu 1.32%
67 | loss 7.4988 | lr 3.350000e-05 | 1786.05ms | mfu 1.32%
68 | loss 7.4706 | lr 3.400000e-05 | 1780.08ms | mfu 1.32%
69 | loss 7.4393 | lr 3.450000e-05 | 1785.92ms | mfu 1.32%
70 | loss 7.4272 | lr 3.500000e-05 | 1781.09ms | mfu 1.32%
71 | loss 7.3885 | lr 3.550000e-05 | 1784.05ms | mfu 1.32%
72 | loss 7.3811 | lr 3.600000e-05 | 1784.42ms | mfu 1.32%
73 | loss 7.3529 | lr 3.650000e-05 | 1789.80ms | mfu 1.32%
74 | loss 7.3472 | lr 3.700000e-05 | 1779.04ms | mfu 1.32%
75 | loss 7.3249 | lr 3.750000e-05 | 1786.06ms | mfu 1.32%
76 | loss 7.3031 | lr 3.800000e-05 | 1779.51ms | mfu 1.32%
77 | loss 7.2921 | lr 3.850000e-05 | 1783.96ms | mfu 1.32%
78 | loss 7.2693 | lr 3.900000e-05 | 1775.51ms | mfu 1.32%
79 | loss 7.2408 | lr 3.950000e-05 | 1779.59ms | mfu 1.32%
80 | loss 7.2256 | lr 4.000000e-05 | 1783.40ms | mfu 1.32%
81 | loss 7.2261 | lr 4.050000e-05 | 1780.04ms | mfu 1.32%
82 | loss 7.1853 | lr 4.100000e-05 | 1781.24ms | mfu 1.32%
83 | loss 7.1707 | lr 4.150000e-05 | 1775.67ms | mfu 1.32%
84 | loss 7.1402 | lr 4.200000e-05 | 1783.91ms | mfu 1.32%
85 | loss 7.1046 | lr 4.250000e-05 | 1780.41ms | mfu 1.32%
86 | loss 7.1195 | lr 4.300000e-05 | 1780.56ms | mfu 1.32%
87 | loss 7.0987 | lr 4.350000e-05 | 1779.48ms | mfu 1.32%
88 | loss 7.0582 | lr 4.400000e-05 | 1780.50ms | mfu 1.32%
89 | loss 7.0510 | lr 4.450000e-05 | 1781.78ms | mfu 1.32%
90 | loss 7.0386 | lr 4.500000e-05 | 1778.67ms | mfu 1.32%
91 | loss 7.0191 | lr 4.550000e-05 | 1783.66ms | mfu 1.32%
92 | loss 7.0163 | lr 4.600000e-05 | 1781.06ms | mfu 1.32%
93 | loss 6.9911 | lr 4.650000e-05 | 1782.13ms | mfu 1.32%
94 | loss 6.9513 | lr 4.700000e-05 | 1782.88ms | mfu 1.32%
95 | loss 6.9741 | lr 4.750000e-05 | 1773.50ms | mfu 1.32%
96 | loss 6.9196 | lr 4.800000e-05 | 1781.74ms | mfu 1.32%
97 | loss 6.9085 | lr 4.850000e-05 | 1787.56ms | mfu 1.32%
98 | loss 6.9231 | lr 4.900000e-05 | 1776.59ms | mfu 1.32%
99 | loss 6.8738 | lr 4.950000e-05 | 1779.54ms | mfu 1.32%
100 | loss 6.8620 | lr 5.000000e-05 | 1782.77ms | mfu 1.32%
