iomgaa/Minimind
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

添加注释，方便学习者快速理解

Browse Source
This commit is contained in:
Ben 2024-09-18 21:53:39 +08:00
parent 61cb61a46a
commit 2dceaf4a92
4 changed files with 337 additions and 319 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

151
1-pretrain.py
View File

@ -14,56 +14,62 @@ from model.model import Transformer
from model.LMConfig import LMConfig from model.LMConfig import LMConfig
from model.dataset import PretrainDataset from model.dataset import PretrainDataset
# 忽略警告信息
warnings.filterwarnings('ignore') warnings.filterwarnings('ignore')
# 定义日志打印函数，仅在主进程（rank 0）打印日志信息
def Logger(content): def Logger(content):
if not ddp or dist.get_rank() == 0: if not ddp or dist.get_rank() == 0:
print(content) print(content)
# 定义学习率调度函数，根据当前迭代次数计算学习率，采用余弦退火策略
def get_lr(it, all): def get_lr(it, all):
warmup_iters = 0 warmup_iters = 0 # 预热迭代次数
lr_decay_iters = all lr_decay_iters = all # 学习率衰减的总迭代次数
min_lr = learning_rate / 10 min_lr = learning_rate / 10 # 最小学习率
# 如果当前迭代次数小于预热迭代次数，使用线性预热策略
if it < warmup_iters: if it < warmup_iters:
return learning_rate * it / warmup_iters return learning_rate * it / warmup_iters
# 如果当前迭代次数大于衰减迭代次数，返回最小学习率
if it > lr_decay_iters: if it > lr_decay_iters:
return min_lr return min_lr
# 计算衰减系数，使用余弦退火策略
decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters) decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
assert 0 <= decay_ratio <= 1 assert 0 <= decay_ratio <= 1
coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
return min_lr + coeff * (learning_rate - min_lr) return min_lr + coeff * (learning_rate - min_lr)
# 定义训练 epoch 的函数
def train_epoch(epoch, accumulation_steps=8): def train_epoch(epoch, accumulation_steps=8):
start_time = time.time() start_time = time.time() # 记录开始时间
for step, (X, Y) in enumerate(train_loader): for step, (X, Y) in enumerate(train_loader): # 遍历数据加载器
X = X.to(device) X = X.to(device) # 将输入数据移动到设备上
Y = Y.to(device) Y = Y.to(device) # 将目标数据移动到设备上
lr = get_lr(epoch * iter_per_epoch + step, epochs * iter_per_epoch) lr = get_lr(epoch * iter_per_epoch + step, epochs * iter_per_epoch) # 计算当前学习率
for param_group in optimizer.param_groups: for param_group in optimizer.param_groups:
param_group['lr'] = lr param_group['lr'] = lr # 设置优化器的学习率
with ctx: with ctx: # 使用混合精度训练（如果设备是 GPU）
out = model(X, Y) out = model(X, Y) # 前向传播，计算输出
loss = out.last_loss / accumulation_steps loss = out.last_loss / accumulation_steps # 计算损失，并进行梯度累积
scaler.scale(loss).backward() scaler.scale(loss).backward() # 反向传播，计算梯度
# 每 accumulation_steps 步进行一次梯度更新
if (step + 1) % accumulation_steps == 0: if (step + 1) % accumulation_steps == 0:
scaler.unscale_(optimizer) scaler.unscale_(optimizer) # 反缩放梯度
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) # 梯度裁剪
scaler.step(optimizer) scaler.step(optimizer) # 更新模型参数
scaler.update() scaler.update() # 更新缩放器
optimizer.zero_grad(set_to_none=True) optimizer.zero_grad(set_to_none=True) # 清空梯度
# 每 100 步打印一次训练信息
if step % 100 == 0: if step % 100 == 0:
spend_time = time.time() - start_time spend_time = time.time() - start_time # 计算已用时间
Logger( Logger(
'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.7f} epoch_Time:{}min:'.format( 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.7f} epoch_Time:{}min:'.format(
epoch, epoch,
@ -74,26 +80,27 @@ def train_epoch(epoch, accumulation_steps=8):
optimizer.param_groups[-1]['lr'], optimizer.param_groups[-1]['lr'],
spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60)) spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))
# 每 1000 步保存一次模型
if (step + 1) % 1000 == 0 and (not ddp or dist.get_rank() == 0): if (step + 1) % 1000 == 0 and (not ddp or dist.get_rank() == 0):
model.eval() model.eval() # 切换到评估模式
# torch.save(model.state_dict(), '{}/iter_{}.pth'.format(save_dir, int(step + epoch * iter_per_epoch))) # torch.save(model.state_dict(), '{}/iter_{}.pth'.format(save_dir, int(step + epoch * iter_per_epoch)))
moe_path = '_moe' if lm_config.use_moe else '' moe_path = '_moe' if lm_config.use_moe else '' # 根据是否使用 MoE 设置保存路径
ckp = f'{save_dir}/pretrain_{lm_config.dim}{moe_path}.pth' ckp = f'{save_dir}/pretrain_{lm_config.dim}{moe_path}.pth'
if isinstance(model, torch.nn.parallel.DistributedDataParallel): if isinstance(model, torch.nn.parallel.DistributedDataParallel):
state_dict = model.module.state_dict() state_dict = model.module.state_dict() # 获取模型状态字典
else: else:
state_dict = model.state_dict() state_dict = model.state_dict()
torch.save(state_dict, ckp) torch.save(state_dict, ckp) # 保存模型
model.train() model.train() # 切换回训练模式
# 定义初始化模型的函数
def init_model(): def init_model():
def count_parameters(model): def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad) return sum(p.numel() for p in model.parameters() if p.requires_grad) # 计算模型可训练参数的数量
# model init # 初始化模型
model = Transformer(lm_config).to(device) model = Transformer(lm_config).to(device)
moe_path = '_moe' if lm_config.use_moe else '' moe_path = '_moe' if lm_config.use_moe else ''
# ckp = f'{save_dir}/pretrain_{lm_config.dim}{moe_path}.pth' # ckp = f'{save_dir}/pretrain_{lm_config.dim}{moe_path}.pth'
@ -105,57 +112,57 @@ def init_model():
# state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k) # state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
# model.load_state_dict(state_dict, strict=False) # model.load_state_dict(state_dict, strict=False)
Logger(f'LLM总参数量：{count_parameters(model) / 1e6:.3f} 百万') Logger(f'LLM总参数量：{count_parameters(model) / 1e6:.3f} 百万') # 打印模型总参数量
return model return model
# 定义初始化分布式训练环境的函数
def init_distributed_mode(): def init_distributed_mode():
if not ddp: return if not ddp: return
global ddp_local_rank, DEVICE global ddp_local_rank, DEVICE
dist.init_process_group(backend="nccl") dist.init_process_group(backend="nccl") # 初始化分布式进程组，使用 NCCL 后端
ddp_rank = int(os.environ["RANK"]) ddp_rank = int(os.environ["RANK"]) # 获取当前进程的 rank
ddp_local_rank = int(os.environ["LOCAL_RANK"]) ddp_local_rank = int(os.environ["LOCAL_RANK"]) # 获取当前进程的本地 rank
ddp_world_size = int(os.environ["WORLD_SIZE"]) ddp_world_size = int(os.environ["WORLD_SIZE"]) # 获取分布式训练的总进程数
DEVICE = f"cuda:{ddp_local_rank}" DEVICE = f"cuda:{ddp_local_rank}" # 设置当前设备的 CUDA 设备
torch.cuda.set_device(DEVICE) torch.cuda.set_device(DEVICE) # 设置当前设备的 CUDA 设备
# torchrun --nproc_per_node 2 1-pretrain.py # torchrun --nproc_per_node 2 1-pretrain.py
# I/O # I/O
if __name__ == "__main__": if __name__ == "__main__":
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
lm_config = LMConfig() lm_config = LMConfig() # 加载配置文件
max_seq_len = lm_config.max_seq_len max_seq_len = lm_config.max_seq_len # 获取最大序列长度
out_dir = 'out' out_dir = 'out' # 设置输出目录
epochs = 20 epochs = 20 # 设置训练 epoch 数
batch_size = 64 batch_size = 64 # 设置批量大小
learning_rate = 2e-4 learning_rate = 2e-4 # 设置初始学习率
device = 'cuda:0' device = 'cuda:0' # 设置设备为 CUDA:0
dtype = 'bfloat16' dtype = 'bfloat16' # 设置数据类型为 bfloat16
save_dir = os.path.join(out_dir) save_dir = os.path.join(out_dir) # 设置模型保存目录
os.makedirs(save_dir, exist_ok=True) os.makedirs(save_dir, exist_ok=True) # 创建模型保存目录
os.makedirs(out_dir, exist_ok=True) os.makedirs(out_dir, exist_ok=True) # 创建输出目录
tokens_per_iter = batch_size * max_seq_len tokens_per_iter = batch_size * max_seq_len # 计算每个迭代处理的 token 数量
torch.manual_seed(1337) torch.manual_seed(1337) # 设置随机种子
device_type = device if "cuda" in device else "cpu" device_type = device if "cuda" in device else "cpu" # 设置设备类型
ctx = ( ctx = (
nullcontext() nullcontext() # 如果设备是 CPU，使用 nullcontext
if device_type == "cpu" if device_type == "cpu"
else torch.cuda.amp.autocast() else torch.cuda.amp.autocast() # 如果设备是 GPU，使用混合精度训练
) )
ddp = int(os.environ.get("RANK", -1)) != -1 # is this a ddp run? ddp = int(os.environ.get("RANK", -1)) != -1 # 判断是否为分布式训练
ddp_local_rank, DEVICE = 0, "cuda:0" ddp_local_rank, DEVICE = 0, "cuda:0" # 初始化分布式训练的本地 rank 和设备
if ddp: if ddp:
init_distributed_mode() init_distributed_mode() # 初始化分布式训练环境
device = torch.device(DEVICE) device = torch.device(DEVICE) # 设置设备
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# -----init dataloader------ # -----init dataloader------
data_path_list = ['./dataset/pretrain_data.bin'] data_path_list = ['./dataset/pretrain_data.bin'] # 设置数据路径
train_ds = PretrainDataset(data_path_list, max_length=max_seq_len, memmap=True) train_ds = PretrainDataset(data_path_list, max_length=max_seq_len, memmap=True) # 初始化数据集
train_sampler = DistributedSampler(train_ds) if ddp else None train_sampler = DistributedSampler(train_ds) if ddp else None # 如果是分布式训练，使用分布式采样器
num_workers = 8 # 可以根据系统的 CPU 核心数来调整 num_workers = 8 # 设置数据加载器的 num_workers
train_loader = DataLoader( train_loader = DataLoader(
train_ds, train_ds,
batch_size=batch_size, batch_size=batch_size,
@ -164,27 +171,27 @@ if __name__ == "__main__":
shuffle=False, shuffle=False,
num_workers=num_workers, num_workers=num_workers,
sampler=train_sampler sampler=train_sampler
) ) # 初始化数据加载器
# init model # init model
model = init_model() model = init_model() # 初始化模型
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == dtype)) scaler = torch.cuda.amp.GradScaler(enabled=(dtype == dtype)) # 初始化梯度缩放器
# optimizer # optimizer
optimizer = optim.Adam(model.parameters(), lr=learning_rate) optimizer = optim.Adam(model.parameters(), lr=learning_rate) # 初始化优化器
# compile the model # compile the model
if False and platform.system() != 'Windows' and float(torch.__version__.split('.')[0]) >= 2: if False and platform.system() != 'Windows' and float(torch.__version__.split('.')[0]) >= 2:
Logger("compiling the model... (takes a ~minute)") Logger("compiling the model... (takes a ~minute)")
unoptimized_model = model unoptimized_model = model
model = torch.compile(model) model = torch.compile(model) # 编译模型（如果条件满足）
if ddp: if ddp:
# Ignore the freqs_cis buffer so that DDP does not broadcast it at # Ignore the freqs_cis buffer so that DDP does not broadcast it at
# construction time since NCCL does not support ComplexFloat # construction time since NCCL does not support ComplexFloat
model._ddp_params_and_buffers_to_ignore = {"pos_cis"} model._ddp_params_and_buffers_to_ignore = {"pos_cis"} # 设置 DDP 忽略的参数和缓冲区
model = DistributedDataParallel(model, device_ids=[ddp_local_rank]) model = DistributedDataParallel(model, device_ids=[ddp_local_rank]) # 使用 DDP 包装模型
# training loop # training loop
iter_per_epoch = len(train_loader) iter_per_epoch = len(train_loader) # 计算每个 epoch 的迭代次数
for epoch in range(epochs): for epoch in range(epochs): # 遍历每个 epoch
train_epoch(epoch) train_epoch(epoch) # 训练一个 epoch

90
model/LMConfig.py
View File

@ -1,58 +1,58 @@
from transformers import PretrainedConfig from transformers import PretrainedConfig
from typing import List from typing import List
# 定义 LMConfig 类，继承自 PretrainedConfig
class LMConfig(PretrainedConfig): class LMConfig(PretrainedConfig):
model_type = "minimind" model_type = "minimind" # 设置模型类型为 "minimind"
def __init__( def __init__(
self, self,
dim: int = 512, dim: int = 512, # 模型维度，默认为 512
n_layers: int = 8, n_layers: int = 8, # Transformer 层数，默认为 8
n_heads: int = 16, n_heads: int = 16, # 注意力头数，默认为 16
n_kv_heads: int = 8, n_kv_heads: int = 8, # KV 头数，默认为 8
vocab_size: int = 6400, vocab_size: int = 6400, # 词汇表大小，默认为 6400
hidden_dim: int = None, hidden_dim: int = None, # 隐藏层维度，默认为 None
multiple_of: int = 64, multiple_of: int = 64, # 隐藏层维度的倍数，默认为 64
norm_eps: float = 1e-5, norm_eps: float = 1e-5, # 归一化层的 epsilon 值，默认为 1e-5
max_seq_len: int = 512, max_seq_len: int = 512, # 最大序列长度，默认为 512
dropout: float = 0.0, dropout: float = 0.0, # Dropout 概率，默认为 0.0
flash_attn: bool = True, flash_attn: bool = True, # 是否使用 Flash Attention，默认为 True
#################################################### ####################################################
# Here are the specific configurations of MOE # 以下是 MOE（Mixture of Experts）的特定配置
# When use_moe is false, the following is invalid # 当 use_moe 为 False 时，以下配置无效
#################################################### ####################################################
use_moe: bool = False, use_moe: bool = False, # 是否使用 MOE，默认为 False
num_experts_per_tok=2, num_experts_per_tok=2, # 每个 token 选择的专家数量，默认为 2
n_routed_experts=4, n_routed_experts=4, # 总的专家数量，默认为 4
n_shared_experts: bool = True, n_shared_experts: bool = True, # 是否使用共享专家，默认为 True
scoring_func='softmax', scoring_func='softmax', # 评分函数，默认为 'softmax'
aux_loss_alpha=0.01, aux_loss_alpha=0.01, # 辅助损失的 alpha 参数，默认为 0.01
seq_aux=True, seq_aux=True, # 是否在序列级别上计算辅助损失，默认为 True
norm_topk_prob=True, norm_topk_prob=True, # 是否标准化 top-k 概率，默认为 True
**kwargs, **kwargs,
): ):
self.dim = dim self.dim = dim # 设置模型维度
self.n_layers = n_layers self.n_layers = n_layers # 设置 Transformer 层数
self.n_heads = n_heads self.n_heads = n_heads # 设置注意力头数
self.n_kv_heads = n_kv_heads self.n_kv_heads = n_kv_heads # 设置 KV 头数
self.vocab_size = vocab_size self.vocab_size = vocab_size # 设置词汇表大小
self.hidden_dim = hidden_dim self.hidden_dim = hidden_dim # 设置隐藏层维度
self.multiple_of = multiple_of self.multiple_of = multiple_of # 设置隐藏层维度的倍数
self.norm_eps = norm_eps self.norm_eps = norm_eps # 设置归一化层的 epsilon 值
self.max_seq_len = max_seq_len self.max_seq_len = max_seq_len # 设置最大序列长度
self.dropout = dropout self.dropout = dropout # 设置 Dropout 概率
self.flash_attn = flash_attn self.flash_attn = flash_attn # 设置是否使用 Flash Attention
#################################################### ####################################################
# Here are the specific configurations of MOE # 以下是 MOE（Mixture of Experts）的特定配置
# When use_moe is false, the following is invalid # 当 use_moe 为 False 时，以下配置无效
#################################################### ####################################################
self.use_moe = use_moe self.use_moe = use_moe # 设置是否使用 MOE
self.num_experts_per_tok = num_experts_per_tok # 每个token选择的专家数量 self.num_experts_per_tok = num_experts_per_tok # 设置每个 token 选择的专家数量
self.n_routed_experts = n_routed_experts # 总的专家数量 self.n_routed_experts = n_routed_experts # 设置总的专家数量
self.n_shared_experts = n_shared_experts # 共享专家 self.n_shared_experts = n_shared_experts # 设置是否使用共享专家
self.scoring_func = scoring_func # 评分函数，默认为'softmax' self.scoring_func = scoring_func # 设置评分函数
self.aux_loss_alpha = aux_loss_alpha # 辅助损失的alpha参数 self.aux_loss_alpha = aux_loss_alpha # 设置辅助损失的 alpha 参数
self.seq_aux = seq_aux # 是否在序列级别上计算辅助损失 self.seq_aux = seq_aux # 设置是否在序列级别上计算辅助损失
self.norm_topk_prob = norm_topk_prob # 是否标准化top-k概率 self.norm_topk_prob = norm_topk_prob # 设置是否标准化 top-k 概率
super().__init__(**kwargs) super().__init__(**kwargs) # 调用父类 PretrainedConfig 的初始化方法

80
model/dataset.py
View File

@ -9,79 +9,79 @@ import torch
from sklearn.model_selection import train_test_split from sklearn.model_selection import train_test_split
import os import os
os.environ["TOKENIZERS_PARALLELISM"] = "false" os.environ["TOKENIZERS_PARALLELISM"] = "false" # 禁用 tokenizer 的并行处理
# 定义 PretrainDataset 类，继承自 Dataset
class PretrainDataset(Dataset): class PretrainDataset(Dataset):
def __init__(self, data_path_lst, max_length=512, memmap=False): def __init__(self, data_path_lst, max_length=512, memmap=False):
super().__init__() super().__init__()
# # 如果使用内存映射（memmap）
if memmap: if memmap:
with open(data_path_lst[0], 'r') as f: with open(data_path_lst[0], 'r') as f:
nbytes = f.seek(0, 2) nbytes = f.seek(0, 2) # 获取文件总字节数
flen = f.tell() // np.dtype('uint16').itemsize flen = f.tell() // np.dtype('uint16').itemsize # 计算文件长度
self.data = np.memmap(data_path_lst[0], dtype=np.dtype('uint16'), shape=(flen // max_length, max_length)) self.data = np.memmap(data_path_lst[0], dtype=np.dtype('uint16'), shape=(flen // max_length, max_length)) # 使用内存映射加载数据
else: else:
data_lst = [] data_lst = []
for data_path in data_path_lst: for data_path in data_path_lst:
with open(data_path, 'rb') as f: with open(data_path, 'rb') as f:
data = np.fromfile(f, dtype=np.uint16) data = np.fromfile(f, dtype=np.uint16) # 从文件中读取数据
data_lst.append(data) data_lst.append(data)
data = np.concatenate(data_lst) data = np.concatenate(data_lst) # 合并所有数据
data = data[:max_length * int(len(data) / max_length)] data = data[:max_length * int(len(data) / max_length)] # 截取数据
# np.random.shuffle(data) # np.random.shuffle(data) # 打乱数据（注释掉了）
self.data = data.reshape(-1, max_length) self.data = data.reshape(-1, max_length) # 将数据重塑为 (样本数, 最大长度) 的形状
# # 打印数据形状
print("memmap:{} train data.shape:{}".format(memmap, self.data.shape)) print("memmap:{} train data.shape:{}".format(memmap, self.data.shape))
print("downloading finished.....") print("downloading finished.....")
def __len__(self): def __len__(self):
return self.data.shape[0] return self.data.shape[0] # 返回数据集的长度
def __getitem__(self, index: int): def __getitem__(self, index: int):
# # 获取指定索引的样本
sample = self.data[index] sample = self.data[index]
X = np.array(sample[:-1]).astype(np.int64) X = np.array(sample[:-1]).astype(np.int64) # 输入数据（去掉最后一个 token）
Y = np.array(sample[1:]).astype(np.int64) Y = np.array(sample[1:]).astype(np.int64) # 目标数据（去掉第一个 token）
return torch.from_numpy(X), torch.from_numpy(Y)
return torch.from_numpy(X), torch.from_numpy(Y) # 返回 PyTorch 张量
# 定义 SFTDataset 类，继承自 Dataset
class SFTDataset(Dataset): class SFTDataset(Dataset):
def __init__(self, df, tokenizer, max_length=1024, prompt_max_len=512, answer_max_len=256): def __init__(self, df, tokenizer, max_length=1024, prompt_max_len=512, answer_max_len=256):
super().__init__() super().__init__()
self.df = df self.df = df # 数据框
self.max_length = max_length self.max_length = max_length # 最大序列长度
self.prompt_max_len = prompt_max_len self.prompt_max_len = prompt_max_len # 提示的最大长度
self.answer_max_len = answer_max_len self.answer_max_len = answer_max_len # 回答的最大长度
# #
self.tokenizer = tokenizer self.tokenizer = tokenizer # 分词器
self.padding = 0 # self.tokenizer.special_tokens['<pad>'] self.padding = 0 # 填充 token ID
self.bos_id = self.tokenizer('<s>assistant').data['input_ids'] self.bos_id = self.tokenizer('<s>assistant').data['input_ids'] # 开始 token ID
def __len__(self): def __len__(self):
return self.df.shape[0] return self.df.shape[0] # 返回数据集的长度
def find_sublist_index(self, main_list, sub_list) -> int: def find_sublist_index(self, main_list, sub_list) -> int:
last_index = -1 last_index = -1
for i in range(len(main_list) - len(sub_list) + 1): for i in range(len(main_list) - len(sub_list) + 1):
if main_list[i:i + len(sub_list)] == sub_list: if main_list[i:i + len(sub_list)] == sub_list:
last_index = i last_index = i
return last_index return last_index # 查找子列表在主列表中的最后一个索引
def safe_eval(self, s): def safe_eval(self, s):
try: try:
res = eval(s) res = eval(s)
except Exception as e: except Exception as e:
return [] return []
return res return res # 安全地执行 eval 函数
def __getitem__(self, index: int): def __getitem__(self, index: int):
# # 获取指定索引的样本
sample = self.df.iloc[index] sample = self.df.iloc[index]
history = self.safe_eval(sample['history']) history = self.safe_eval(sample['history']) # 获取历史对话
q = str(sample['q']) q = str(sample['q']) # 获取问题
a = str(sample['a']) a = str(sample['a']) # 获取回答
messages = [] messages = []
for history_message in history: for history_message in history:
@ -102,29 +102,29 @@ class SFTDataset(Dataset):
messages, messages,
tokenize=False, tokenize=False,
add_generation_prompt=True add_generation_prompt=True
) ) # 生成新的提示
input_id = self.tokenizer(new_prompt).data['input_ids'][:self.max_length] input_id = self.tokenizer(new_prompt).data['input_ids'][:self.max_length] # 分词并截取
# 实际长度 # 实际长度
question_length = self.find_sublist_index(input_id, self.bos_id) + len(self.bos_id) question_length = self.find_sublist_index(input_id, self.bos_id) + len(self.bos_id)
# 没满最大长度的剩余部分 # 没满最大长度的剩余部分
padding_len = self.max_length - len(input_id) padding_len = self.max_length - len(input_id)
input_id = input_id + [self.padding] * padding_len input_id = input_id + [self.padding] * padding_len # 填充到最大长度
mask_len = len(input_id) - question_length - padding_len mask_len = len(input_id) - question_length - padding_len
# 0表示不计算损失 # 0表示不计算损失
loss_mask = [0] * question_length + [1] * (mask_len) + [0] * padding_len loss_mask = [0] * question_length + [1] * (mask_len) + [0] * padding_len
input_id = np.array(input_id) input_id = np.array(input_id)
X = np.array(input_id[:-1]).astype(np.int64) X = np.array(input_id[:-1]).astype(np.int64) # 输入数据（去掉最后一个 token）
Y = np.array(input_id[1:]).astype(np.int64) Y = np.array(input_id[1:]).astype(np.int64) # 目标数据（去掉第一个 token）
loss_mask = np.array(loss_mask[1:]).astype(np.int64) loss_mask = np.array(loss_mask[1:]).astype(np.int64) # 损失掩码
X_tensor = torch.from_numpy(X) X_tensor = torch.from_numpy(X)
Y_tensor = torch.from_numpy(Y) Y_tensor = torch.from_numpy(Y)
loss_mask_tensor = torch.from_numpy(loss_mask) loss_mask_tensor = torch.from_numpy(loss_mask)
return X_tensor, Y_tensor, loss_mask_tensor return X_tensor, Y_tensor, loss_mask_tensor # 返回 PyTorch 张量
# 主函数
if __name__ == "__main__": if __name__ == "__main__":
pass pass

333
model/model.py
View File

@ -10,29 +10,29 @@ from torch import nn
from transformers import PreTrainedModel from transformers import PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast from transformers.modeling_outputs import CausalLMOutputWithPast
# 定义 RMSNorm 类，实现一种归一化方法，类似于 LayerNorm，但计算方式不同
class RMSNorm(torch.nn.Module): class RMSNorm(torch.nn.Module):
def __init__(self, dim: int, eps: float): def __init__(self, dim: int, eps: float):
super().__init__() super().__init__()
self.eps = eps self.eps = eps # 设置 epsilon，防止除零错误
self.weight = nn.Parameter(torch.ones(dim)) self.weight = nn.Parameter(torch.ones(dim)) # 初始化权重参数
def _norm(self, x): def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) # 计算 RMSNorm
def forward(self, x): def forward(self, x):
output = self._norm(x.float()).type_as(x) output = self._norm(x.float()).type_as(x) # 应用 RMSNorm
return output * self.weight return output * self.weight # 乘以权重参数
# 定义 precompute_pos_cis 函数，用于预计算位置编码的复数形式
def precompute_pos_cis(dim: int, end: int, theta: float = 10000.0): def precompute_pos_cis(dim: int, end: int, theta: float = 10000.0):
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) # 计算频率
t = torch.arange(end, device=freqs.device) # type: ignore t = torch.arange(end, device=freqs.device) # 生成时间序列
freqs = torch.outer(t, freqs).float() # type: ignore freqs = torch.outer(t, freqs).float() # 计算外积
pos_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64 pos_cis = torch.polar(torch.ones_like(freqs), freqs) # 计算复数形式的位置编码
return pos_cis return pos_cis
# 定义 apply_rotary_emb 函数，用于应用旋转位置编码
def apply_rotary_emb(xq, xk, pos_cis): def apply_rotary_emb(xq, xk, pos_cis):
def unite_shape(pos_cis, x): def unite_shape(pos_cis, x):
ndim = x.ndim ndim = x.ndim
@ -41,14 +41,14 @@ def apply_rotary_emb(xq, xk, pos_cis):
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
return pos_cis.view(*shape) return pos_cis.view(*shape)
xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2)) xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2)) # 将 xq 转换为复数形式
xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) # 将 xk 转换为复数形式
pos_cis = unite_shape(pos_cis, xq_) pos_cis = unite_shape(pos_cis, xq_) # 调整 pos_cis 的形状
xq_out = torch.view_as_real(xq_ * pos_cis).flatten(3) xq_out = torch.view_as_real(xq_ * pos_cis).flatten(3) # 应用旋转位置编码
xk_out = torch.view_as_real(xk_ * pos_cis).flatten(3) xk_out = torch.view_as_real(xk_ * pos_cis).flatten(3) # 应用旋转位置编码
return xq_out.type_as(xq), xk_out.type_as(xk) return xq_out.type_as(xq), xk_out.type_as(xk) # 返回结果
# 定义 repeat_kv 函数，用于重复 KV 头的值
def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor: def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
"""torch.repeat_interleave(x, dim=2, repeats=n_rep)""" """torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
bs, slen, n_kv_heads, head_dim = x.shape bs, slen, n_kv_heads, head_dim = x.shape
@ -60,130 +60,130 @@ def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
.reshape(bs, slen, n_kv_heads * n_rep, head_dim) .reshape(bs, slen, n_kv_heads * n_rep, head_dim)
) )
# 定义 Attention 类，实现自注意力机制
class Attention(nn.Module): class Attention(nn.Module):
def __init__(self, args: LMConfig): def __init__(self, args: LMConfig):
super().__init__() super().__init__()
self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads # 设置 KV 头的数量
assert args.n_heads % self.n_kv_heads == 0 assert args.n_heads % self.n_kv_heads == 0 # 确保 KV 头的数量是总头数的因数
self.n_local_heads = args.n_heads self.n_local_heads = args.n_heads # 设置本地头的数量
self.n_local_kv_heads = self.n_kv_heads self.n_local_kv_heads = self.n_kv_heads # 设置本地 KV 头的数量
self.n_rep = self.n_local_heads // self.n_local_kv_heads self.n_rep = self.n_local_heads // self.n_local_kv_heads # 计算重复次数
self.head_dim = args.dim // args.n_heads self.head_dim = args.dim // args.n_heads # 计算每个头的维度
self.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False) self.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False) # 初始化 Q 矩阵
self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) # 初始化 K 矩阵
self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) # 初始化 V 矩阵
self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False) self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False) # 初始化输出矩阵
self.k_cache, self.v_cache = None, None self.k_cache, self.v_cache = None, None # 初始化 KV 缓存
self.attn_dropout = nn.Dropout(args.dropout) self.attn_dropout = nn.Dropout(args.dropout) # 初始化注意力 dropout
self.resid_dropout = nn.Dropout(args.dropout) self.resid_dropout = nn.Dropout(args.dropout) # 初始化残差 dropout
self.dropout = args.dropout self.dropout = args.dropout # 设置 dropout 概率
self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn # 判断是否使用 Flash Attention
if not self.flash: if not self.flash:
# print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0") # print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")
mask = torch.full((1, 1, args.max_seq_len, args.max_seq_len), float("-inf")) mask = torch.full((1, 1, args.max_seq_len, args.max_seq_len), float("-inf")) # 初始化掩码
mask = torch.triu(mask, diagonal=1) mask = torch.triu(mask, diagonal=1) # 生成上三角掩码
self.register_buffer("mask", mask) self.register_buffer("mask", mask) # 注册掩码
def forward(self, x: torch.Tensor, pos_cis: torch.Tensor, use_kv_cache=False): def forward(self, x: torch.Tensor, pos_cis: torch.Tensor, use_kv_cache=False):
bsz, seqlen, _ = x.shape bsz, seqlen, _ = x.shape
if use_kv_cache and self.eval(): if use_kv_cache and self.eval(): # 如果使用 KV 缓存且在评估模式下
if self.k_cache is None or self.k_cache.shape[1] != x.shape[1] - 1: if self.k_cache is None or self.k_cache.shape[1] != x.shape[1] - 1:
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) # 计算 Q, K, V
else: else:
token = x[:, -1:, :] token = x[:, -1:, :] # 获取最后一个 token
xq = torch.cat((torch.zeros_like(x[:, :-1, :]), self.wq(token)), dim=1) xq = torch.cat((torch.zeros_like(x[:, :-1, :]), self.wq(token)), dim=1) # 更新 Q
xk = torch.cat((self.k_cache, self.wk(token)), dim=1) xk = torch.cat((self.k_cache, self.wk(token)), dim=1) # 更新 K
xv = torch.cat((self.v_cache, self.wv(token)), dim=1) xv = torch.cat((self.v_cache, self.wv(token)), dim=1) # 更新 V
self.k_cache, self.v_cache = xk, xv self.k_cache, self.v_cache = xk, xv # 更新 KV 缓存
else: else:
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) # 计算 Q, K, V
xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim) xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim) # 调整 Q 的形状
xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) # 调整 K 的形状
xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) # 调整 V 的形状
xq, xk = apply_rotary_emb(xq, xk, pos_cis) xq, xk = apply_rotary_emb(xq, xk, pos_cis) # 应用旋转位置编码
xk = repeat_kv(xk, self.n_rep) # (bs, seqlen, n_local_heads, head_dim) xk = repeat_kv(xk, self.n_rep) # 重复 K 的值
xv = repeat_kv(xv, self.n_rep) # (bs, seqlen, n_local_heads, head_dim) xv = repeat_kv(xv, self.n_rep) # 重复 V 的值
xq = xq.transpose(1, 2) xq = xq.transpose(1, 2) # 调整 Q 的形状
xk = xk.transpose(1, 2) xk = xk.transpose(1, 2) # 调整 K 的形状
xv = xv.transpose(1, 2) xv = xv.transpose(1, 2) # 调整 V 的形状
if self.flash: if self.flash:
output = torch.nn.functional.scaled_dot_product_attention(xq, xk, xv, attn_mask=None, output = torch.nn.functional.scaled_dot_product_attention(xq, xk, xv, attn_mask=None,
dropout_p=self.dropout if self.training else 0.0, dropout_p=self.dropout if self.training else 0.0,
is_causal=True) is_causal=True) # 使用 Flash Attention
else: else:
scores = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_dim) scores = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_dim) # 计算注意力分数
assert hasattr(self, 'mask') assert hasattr(self, 'mask')
scores = scores + self.mask[:, :, :seqlen, :seqlen] # (bs, n_local_heads, seqlen, cache_len + seqlen) scores = scores + self.mask[:, :, :seqlen, :seqlen] # 应用掩码
scores = F.softmax(scores.float(), dim=-1).type_as(xq) scores = F.softmax(scores.float(), dim=-1).type_as(xq) # 计算 softmax
scores = self.attn_dropout(scores) scores = self.attn_dropout(scores) # 应用注意力 dropout
output = torch.matmul(scores, xv) # (bs, n_local_heads, seqlen, head_dim) output = torch.matmul(scores, xv) # 计算输出
output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1) output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1) # 调整输出的形状
output = self.wo(output)
output = self.resid_dropout(output)
return output
output = self.wo(output) # 应用输出矩阵
output = self.resid_dropout(output) # 应用残差 dropout
return output # 返回输出
# 定义 FeedForward 类，实现前馈神经网络
class FeedForward(nn.Module): class FeedForward(nn.Module):
def __init__(self, dim: int, hidden_dim: int, multiple_of: int, dropout: float): def __init__(self, dim: int, hidden_dim: int, multiple_of: int, dropout: float):
super().__init__() super().__init__()
if hidden_dim is None: if hidden_dim is None:
hidden_dim = 4 * dim hidden_dim = 4 * dim # 设置隐藏层维度
hidden_dim = int(2 * hidden_dim / 3) hidden_dim = int(2 * hidden_dim / 3) # 调整隐藏层维度
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) # 调整隐藏层维度
self.w1 = nn.Linear(dim, hidden_dim, bias=False) self.w1 = nn.Linear(dim, hidden_dim, bias=False) # 初始化第一层线性变换
self.w2 = nn.Linear(hidden_dim, dim, bias=False) self.w2 = nn.Linear(hidden_dim, dim, bias=False) # 初始化第二层线性变换
self.w3 = nn.Linear(dim, hidden_dim, bias=False) self.w3 = nn.Linear(dim, hidden_dim, bias=False) # 初始化第三层线性变换
self.dropout = nn.Dropout(dropout) self.dropout = nn.Dropout(dropout) # 初始化 dropout
def forward(self, x): def forward(self, x):
return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x))) return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x))) # 前向传播
# 定义 MoEGate 类，实现专家混合（MoE）的门控机制
class MoEGate(nn.Module): class MoEGate(nn.Module):
def __init__(self, config: LMConfig): def __init__(self, config: LMConfig):
super().__init__() super().__init__()
self.config = config self.config = config
self.top_k = config.num_experts_per_tok self.top_k = config.num_experts_per_tok # 设置每个 token 选择的专家数量
self.n_routed_experts = config.n_routed_experts self.n_routed_experts = config.n_routed_experts # 设置路由专家的数量
self.scoring_func = config.scoring_func self.scoring_func = config.scoring_func # 设置评分函数
self.alpha = config.aux_loss_alpha self.alpha = config.aux_loss_alpha # 设置辅助损失的权重
self.seq_aux = config.seq_aux self.seq_aux = config.seq_aux # 设置序列辅助损失
self.norm_topk_prob = config.norm_topk_prob self.norm_topk_prob = config.norm_topk_prob # 设置是否归一化 top-k 概率
self.gating_dim = config.dim self.gating_dim = config.dim # 设置门控维度
self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim))) self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim))) # 初始化权重参数
self.reset_parameters() self.reset_parameters() # 重置参数
def reset_parameters(self) -> None: def reset_parameters(self) -> None:
import torch.nn.init as init import torch.nn.init as init
init.kaiming_uniform_(self.weight, a=math.sqrt(5)) init.kaiming_uniform_(self.weight, a=math.sqrt(5)) # 使用 Kaiming 初始化权重
def forward(self, hidden_states): def forward(self, hidden_states):
bsz, seq_len, h = hidden_states.shape bsz, seq_len, h = hidden_states.shape
hidden_states = hidden_states.view(-1, h) hidden_states = hidden_states.view(-1, h) # 调整隐藏状态的形状
logits = F.linear(hidden_states, self.weight, None) logits = F.linear(hidden_states, self.weight, None) # 计算 logits
if self.scoring_func == 'softmax': if self.scoring_func == 'softmax':
scores = logits.softmax(dim=-1) scores = logits.softmax(dim=-1) # 计算 softmax 评分
else: else:
raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}') raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False) topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False) # 选择 top-k 专家
if self.top_k > 1 and self.norm_topk_prob: if self.top_k > 1 and self.norm_topk_prob:
denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20 denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20 # 计算归一化分母
topk_weight = topk_weight / denominator topk_weight = topk_weight / denominator # 归一化 top-k 概率
if self.training and self.alpha > 0.0: if self.training and self.alpha > 0.0:
scores_for_aux = scores scores_for_aux = scores
@ -204,9 +204,9 @@ class MoEGate(nn.Module):
aux_loss = (Pi * fi).sum() * self.alpha aux_loss = (Pi * fi).sum() * self.alpha
else: else:
aux_loss = None aux_loss = None
return topk_idx, topk_weight, aux_loss return topk_idx, topk_weight, aux_loss # 返回 top-k 专家索引、权重和辅助损失
# 定义 MOEFeedForward 类，实现专家混合（MoE）的前馈神经网络
class MOEFeedForward(nn.Module): class MOEFeedForward(nn.Module):
def __init__(self, config: LMConfig): def __init__(self, config: LMConfig):
super().__init__() super().__init__()
@ -219,16 +219,16 @@ class MOEFeedForward(nn.Module):
dropout=config.dropout, dropout=config.dropout,
) )
for _ in range(config.n_routed_experts) for _ in range(config.n_routed_experts)
]) ]) # 初始化专家列表
self.gate = MoEGate(config) self.gate = MoEGate(config) # 初始化门控机制
if config.n_shared_experts is not None: if config.n_shared_experts is not None:
self.shared_experts = FeedForward( self.shared_experts = FeedForward(
dim=config.dim, dim=config.dim,
hidden_dim=config.hidden_dim, hidden_dim=config.hidden_dim,
multiple_of=config.multiple_of, multiple_of=config.multiple_of,
dropout=config.dropout, dropout=config.dropout,
) ) # 初始化共享专家
def forward(self, x): def forward(self, x):
identity = x identity = x
@ -281,35 +281,46 @@ class MOEFeedForward(nn.Module):
return expert_cache return expert_cache
# 定义 TransformerBlock 类，实现 Transformer 的一个块，包括自注意力和前馈神经网络
class TransformerBlock(nn.Module): class TransformerBlock(nn.Module):
def __init__(self, layer_id: int, args: LMConfig): def __init__(self, layer_id: int, args: LMConfig):
super().__init__() super().__init__()
self.n_heads = args.n_heads self.n_heads = args.n_heads
self.dim = args.dim self.dim = args.dim
self.head_dim = args.dim // args.n_heads self.head_dim = args.dim // args.n_heads
self.attention = Attention(args) self.attention = Attention(args) # 初始化自注意力机制
self.layer_id = layer_id self.layer_id = layer_id
self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps) self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps) # 初始化注意力归一化
self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps) self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps) # 初始化前馈神经网络归一化
if args.use_moe: if args.use_moe:
self.feed_forward = MOEFeedForward(args) self.feed_forward = MOEFeedForward(args) # 初始化专家混合前馈神经网络
else: else:
self.feed_forward = FeedForward( self.feed_forward = FeedForward(
dim=args.dim, dim=args.dim,
hidden_dim=args.hidden_dim, hidden_dim=args.hidden_dim,
multiple_of=args.multiple_of, multiple_of=args.multiple_of,
dropout=args.dropout, dropout=args.dropout,
) ) # 初始化前馈神经网络
def forward(self, x, pos_cis, use_kv_cache=False): def forward(self, x, pos_cis, use_kv_cache=False):
h = x + self.attention(self.attention_norm(x), pos_cis, use_kv_cache) h = x + self.attention(self.attention_norm(x), pos_cis, use_kv_cache) # 计算自注意力
out = h + self.feed_forward(self.ffn_norm(h)) out = h + self.feed_forward(self.ffn_norm(h)) # 计算前馈神经网络
return out return out # 返回输出
# 定义 Transformer 类，实现整个 Transformer 模型
class Transformer(PreTrainedModel):
config_class = LMConfig
last_loss: Optional[torch.Tensor]
def __init__(self, params: LMConfig = None):
super().__init__(params)
if not params:
params = LMConfig()
self.params = params
self.vocab_size = params.vocab_size
self.n_layers = params.n_layers
class Transformer(PreTrainedModel): class Transformer(PreTrainedModel):
config_class = LMConfig config_class = LMConfig
last_loss: Optional[torch.Tensor] last_loss: Optional[torch.Tensor]
@ -322,99 +333,99 @@ class Transformer(PreTrainedModel):
self.vocab_size = params.vocab_size self.vocab_size = params.vocab_size
self.n_layers = params.n_layers self.n_layers = params.n_layers
self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim) self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim) # 初始化词嵌入层
self.dropout = nn.Dropout(params.dropout) self.dropout = nn.Dropout(params.dropout) # 初始化 dropout 层
self.layers = torch.nn.ModuleList() self.layers = torch.nn.ModuleList() # 初始化 Transformer 块列表
for layer_id in range(self.n_layers): for layer_id in range(self.n_layers):
self.layers.append(TransformerBlock(layer_id, params)) self.layers.append(TransformerBlock(layer_id, params)) # 添加 Transformer 块
self.norm = RMSNorm(params.dim, eps=params.norm_eps) self.norm = RMSNorm(params.dim, eps=params.norm_eps) # 初始化归一化层
self.output = nn.Linear(params.dim, params.vocab_size, bias=False) self.output = nn.Linear(params.dim, params.vocab_size, bias=False) # 初始化输出层
self.tok_embeddings.weight = self.output.weight self.tok_embeddings.weight = self.output.weight # 共享词嵌入和输出层的权重
pos_cis = precompute_pos_cis(self.params.dim // self.params.n_heads, self.params.max_seq_len) pos_cis = precompute_pos_cis(self.params.dim // self.params.n_heads, self.params.max_seq_len) # 预计算位置编码
self.register_buffer("pos_cis", pos_cis, persistent=False) self.register_buffer("pos_cis", pos_cis, persistent=False) # 注册位置编码缓冲区
self.apply(self._init_weights) self.apply(self._init_weights) # 初始化模型权重
for pn, p in self.named_parameters(): for pn, p in self.named_parameters():
if pn.endswith('w3.weight') or pn.endswith('wo.weight'): if pn.endswith('w3.weight') or pn.endswith('wo.weight'):
torch.nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * params.n_layers)) torch.nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * params.n_layers)) # 对特定权重进行初始化
self.last_loss = None self.last_loss = None # 初始化最后一个损失
self.OUT = CausalLMOutputWithPast() self.OUT = CausalLMOutputWithPast() # 初始化输出对象
def _init_weights(self, module): def _init_weights(self, module):
if isinstance(module, nn.Linear): if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) # 初始化线性层的权重
if module.bias is not None: if module.bias is not None:
torch.nn.init.zeros_(module.bias) torch.nn.init.zeros_(module.bias) # 初始化线性层的偏置
elif isinstance(module, nn.Embedding): elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) # 初始化嵌入层的权重
def forward(self, tokens: Optional[torch.Tensor] = None, targets: Optional[torch.Tensor] = None, def forward(self, tokens: Optional[torch.Tensor] = None, targets: Optional[torch.Tensor] = None,
use_kv_cache=False, **keyargs): use_kv_cache=False, **keyargs):
if 'input_ids' in keyargs: if 'input_ids' in keyargs:
tokens = keyargs['input_ids'] tokens = keyargs['input_ids'] # 如果传入了 input_ids，则使用 input_ids
if 'attention_mask' in keyargs: if 'attention_mask' in keyargs:
targets = keyargs['attention_mask'] targets = keyargs['attention_mask'] # 如果传入了 attention_mask，则使用 attention_mask
_bsz, seqlen = tokens.shape _bsz, seqlen = tokens.shape # 获取批量大小和序列长度
h = self.tok_embeddings(tokens) h = self.tok_embeddings(tokens) # 获取词嵌入
h = self.dropout(h) h = self.dropout(h) # 应用 dropout
pos_cis = self.pos_cis[:seqlen] pos_cis = self.pos_cis[:seqlen] # 获取对应序列长度的位置编码
for idx, layer in enumerate(self.layers): for idx, layer in enumerate(self.layers):
h = layer(h, pos_cis, use_kv_cache) h = layer(h, pos_cis, use_kv_cache) # 逐层应用 Transformer 块
h = self.norm(h) h = self.norm(h) # 应用归一化
if targets is not None: if targets is not None:
logits = self.output(h) logits = self.output(h) # 计算 logits
self.last_loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1) self.last_loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1) # 计算交叉熵损失
else: else:
logits = self.output(h[:, [-1], :]) logits = self.output(h[:, [-1], :]) # 计算最后一个 token 的 logits
self.last_loss = None self.last_loss = None # 没有目标时，损失为 None
self.OUT.__setitem__('logits', logits) self.OUT.__setitem__('logits', logits) # 设置输出对象的 logits
self.OUT.__setitem__('last_loss', self.last_loss) self.OUT.__setitem__('last_loss', self.last_loss) # 设置输出对象的 last_loss
return self.OUT return self.OUT # 返回输出对象
@torch.inference_mode() @torch.inference_mode() # 推理模式
def generate(self, idx, eos, max_new_tokens, temperature=0.7, top_k=None, stream=True, repetition_penalty=1., def generate(self, idx, eos, max_new_tokens, temperature=0.7, top_k=None, stream=True, repetition_penalty=1.,
use_kv_cache=True): use_kv_cache=True):
index = idx.shape[1] index = idx.shape[1] # 获取当前序列长度
while idx.shape[1] < max_new_tokens - 1: while idx.shape[1] < max_new_tokens - 1: # 当生成的 token 数量小于最大数量时
inference_res = self(idx, use_kv_cache=use_kv_cache) inference_res = self(idx, use_kv_cache=use_kv_cache) # 进行前向传播
logits = inference_res.logits logits = inference_res.logits # 获取 logits
logits = logits[:, -1, :] logits = logits[:, -1, :] # 获取最后一个 token 的 logits
for token in set(idx.tolist()[0]): for token in set(idx.tolist()[0]): # 对重复 token 进行惩罚
logits[:, token] /= repetition_penalty logits[:, token] /= repetition_penalty
if temperature == 0.0: if temperature == 0.0: # 如果温度为 0，直接选择概率最高的 token
_, idx_next = torch.topk(logits, k=1, dim=-1) _, idx_next = torch.topk(logits, k=1, dim=-1)
else: else:
logits = logits / temperature logits = logits / temperature # 调整 logits
if top_k is not None: if top_k is not None: # 如果设置了 top-k 采样
v, _ = torch.topk(logits, min(top_k, logits.size(-1))) v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < v[:, [-1]]] = -float('Inf') logits[logits < v[:, [-1]]] = -float('Inf') # 将小于 top-k 的 logits 设为负无穷
probs = F.softmax(logits, dim=-1) probs = F.softmax(logits, dim=-1) # 计算概率
idx_next = torch.multinomial(probs, num_samples=1, generator=None) idx_next = torch.multinomial(probs, num_samples=1, generator=None) # 采样下一个 token
if idx_next == eos: if idx_next == eos: # 如果生成的 token 是结束符，停止生成
break break
idx = torch.cat((idx, idx_next), dim=1) idx = torch.cat((idx, idx_next), dim=1) # 将生成的 token 添加到序列中
if stream: if stream: # 如果需要流式输出
yield idx[:, index:] yield idx[:, index:] # 返回生成的 token
if not stream: if not stream: # 如果不需要流式输出
yield idx[:, index:] yield idx[:, index:] # 返回生成的 token
@torch.inference_mode() @torch.inference_mode() # 推理模式
def eval_answer(self, idx): def eval_answer(self, idx):
idx_cond = idx if idx.size(1) <= self.params.max_seq_len else idx[:, -self.params.max_seq_len:] idx_cond = idx if idx.size(1) <= self.params.max_seq_len else idx[:, -self.params.max_seq_len:] # 截取序列
inference_res = self(idx_cond) inference_res = self(idx_cond) # 进行前向传播
logits = inference_res.logits logits = inference_res.logits # 获取 logits
logits = logits[:, -1, :] logits = logits[:, -1, :] # 获取最后一个 token 的 logits
return logits return logits # 返回 logits