547 lines
25 KiB
Python
547 lines
25 KiB
Python
# 📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘
|
||
# MiniMind Config
|
||
# 📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘
|
||
|
||
# MiniMind模型配置类，继承自transformers的PretrainedConfig
|
||
# 包含了模型的所有超参数设置，包括基础参数和MoE（混合专家）相关参数
|
||
|
||
from transformers import PretrainedConfig
|
||
|
||
|
||
class MiniMindConfig(PretrainedConfig):
|
||
# 模型类型标识符
|
||
model_type = "minimind"
|
||
|
||
def __init__(
|
||
self,
|
||
# 基础模型参数
|
||
dropout: float = 0.0, # dropout比率，用于防止过拟合
|
||
bos_token_id: int = 1, # 序列开始标记的ID
|
||
eos_token_id: int = 2, # 序列结束标记的ID
|
||
hidden_act: str = 'silu', # 激活函数类型
|
||
hidden_size: int = 512, # 隐藏层维度
|
||
intermediate_size: int = None, # 前馈网络中间层维度
|
||
max_position_embeddings: int = 32768, # 最大位置编码长度
|
||
num_attention_heads: int = 8, # 注意力头数量
|
||
num_hidden_layers: int = 8, # Transformer层数
|
||
num_key_value_heads: int = 2, # KV注意力头数量（用于分组查询注意力）
|
||
vocab_size: int = 6400, # 词表大小
|
||
rms_norm_eps: float = 1e-05, # RMSNorm层的epsilon值
|
||
rope_theta: int = 1000000.0, # 旋转位置编码的theta参数
|
||
flash_attn: bool = True, # 是否使用Flash Attention优化
|
||
####################################################
|
||
# Here are the specific configurations of MOE
|
||
# When use_moe is false, the following is invalid
|
||
####################################################
|
||
use_moe: bool = False,
|
||
num_experts_per_tok: int = 2,
|
||
n_routed_experts: int = 4,
|
||
n_shared_experts: int = 1,
|
||
scoring_func: str = 'softmax',
|
||
aux_loss_alpha: float = 0.1,
|
||
seq_aux: bool = True,
|
||
norm_topk_prob: bool = True,
|
||
**kwargs
|
||
):
|
||
super().__init__(**kwargs)
|
||
self.dropout = dropout
|
||
self.bos_token_id = bos_token_id
|
||
self.eos_token_id = eos_token_id
|
||
self.hidden_act = hidden_act
|
||
self.hidden_size = hidden_size
|
||
self.intermediate_size = intermediate_size
|
||
self.max_position_embeddings = max_position_embeddings
|
||
self.num_attention_heads = num_attention_heads
|
||
self.num_hidden_layers = num_hidden_layers
|
||
self.num_key_value_heads = num_key_value_heads
|
||
self.vocab_size = vocab_size
|
||
self.rms_norm_eps = rms_norm_eps
|
||
self.rope_theta = rope_theta
|
||
self.flash_attn = flash_attn
|
||
####################################################
|
||
# Here are the specific configurations of MOE
|
||
# When use_moe is false, the following is invalid
|
||
####################################################
|
||
self.use_moe = use_moe
|
||
self.num_experts_per_tok = num_experts_per_tok # 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_shared_experts = n_shared_experts # 共享专家
|
||
self.scoring_func = scoring_func # 评分函数，默认为'softmax'
|
||
self.aux_loss_alpha = aux_loss_alpha # 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.norm_topk_prob = norm_topk_prob # 是否标准化top-k概率
|
||
|
||
|
||
# 📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘
|
||
# MiniMind Model
|
||
# 📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘📘
|
||
|
||
import math
|
||
import torch
|
||
from torch import nn
|
||
from transformers.activations import ACT2FN
|
||
from typing import Optional, Tuple, List, Union
|
||
import torch.nn.functional as F
|
||
from transformers import PreTrainedModel, GenerationMixin, PretrainedConfig
|
||
from transformers.modeling_outputs import CausalLMOutputWithPast
|
||
|
||
|
||
class RMSNorm(torch.nn.Module):
|
||
"""RMSNorm层实现
|
||
相比LayerNorm，RMSNorm去掉了均值归一化，只保留方差归一化，计算更简单且效果相当
|
||
"""
|
||
def __init__(self, dim: int, eps: float = 1e-5):
|
||
super().__init__()
|
||
self.eps = eps # 数值稳定性参数
|
||
self.weight = nn.Parameter(torch.ones(dim)) # 可学习的缩放参数
|
||
|
||
def _norm(self, x):
|
||
# 计算RMS（均方根）归一化
|
||
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
||
|
||
def forward(self, x):
|
||
# 应用缩放并保持输入数据类型
|
||
return self.weight * self._norm(x.float()).type_as(x)
|
||
|
||
|
||
def precompute_freqs_cis(dim: int, end: int = int(32 * 1024), theta: float = 1e6):
|
||
"""预计算旋转位置编码（RoPE）的三角函数值
|
||
|
||
Args:
|
||
dim: 隐藏层维度
|
||
end: 最大序列长度
|
||
theta: 位置编码的缩放因子
|
||
|
||
Returns:
|
||
freqs_cos: 余弦位置编码
|
||
freqs_sin: 正弦位置编码
|
||
"""
|
||
# 计算频率基础值
|
||
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
||
t = torch.arange(end, device=freqs.device)
|
||
# 计算外积得到位置编码矩阵
|
||
freqs = torch.outer(t, freqs).float()
|
||
# 生成最终的三角函数值
|
||
freqs_cos = torch.cat([torch.cos(freqs), torch.cos(freqs)], dim=-1)
|
||
freqs_sin = torch.cat([torch.sin(freqs), torch.sin(freqs)], dim=-1)
|
||
return freqs_cos, freqs_sin
|
||
|
||
|
||
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
||
"""应用旋转位置编码（RoPE）到查询和键向量
|
||
|
||
通过将位置信息编码到注意力的查询(Q)和键(K)向量中，使模型能够感知token的相对位置关系
|
||
"""
|
||
def rotate_half(x):
|
||
# 将向量的后半部分取负并与前半部分交换位置
|
||
return torch.cat((-x[..., x.shape[-1] // 2:], x[..., : x.shape[-1] // 2]), dim=-1)
|
||
|
||
# 应用旋转变换到查询和键向量
|
||
q_embed = (q * cos.unsqueeze(unsqueeze_dim)) + (rotate_half(q) * sin.unsqueeze(unsqueeze_dim))
|
||
k_embed = (k * cos.unsqueeze(unsqueeze_dim)) + (rotate_half(k) * sin.unsqueeze(unsqueeze_dim))
|
||
return q_embed, k_embed
|
||
|
||
|
||
def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
|
||
"""重复键值向量以匹配注意力头的数量
|
||
|
||
在分组查询注意力中，键值向量的头数少于查询向量，需要将键值向量重复多次
|
||
|
||
Args:
|
||
x: 输入张量，形状为[batch_size, seq_len, num_kv_heads, head_dim]
|
||
n_rep: 重复次数
|
||
"""
|
||
bs, slen, num_key_value_heads, head_dim = x.shape
|
||
if n_rep == 1:
|
||
return x
|
||
# 扩展并重塑张量以实现重复
|
||
return (
|
||
x[:, :, :, None, :]
|
||
.expand(bs, slen, num_key_value_heads, n_rep, head_dim)
|
||
.reshape(bs, slen, num_key_value_heads * n_rep, head_dim)
|
||
)
|
||
|
||
|
||
class Attention(nn.Module):
|
||
"""多头注意力机制实现
|
||
|
||
支持分组查询注意力（GQA）和Flash Attention优化
|
||
分组查询注意力通过减少键值头的数量来降低计算和内存开销
|
||
"""
|
||
def __init__(self, args: MiniMindConfig):
|
||
super().__init__()
|
||
# 设置键值头数量，支持分组查询注意力
|
||
self.num_key_value_heads = args.num_attention_heads if args.num_key_value_heads is None else args.num_key_value_heads
|
||
assert args.num_attention_heads % self.num_key_value_heads == 0
|
||
self.n_local_heads = args.num_attention_heads # 查询头总数
|
||
self.n_local_kv_heads = self.num_key_value_heads # 键值头数量
|
||
self.n_rep = self.n_local_heads // self.n_local_kv_heads # 每个键值头需要重复的次数
|
||
self.head_dim = args.hidden_size // args.num_attention_heads # 每个注意力头的维度
|
||
|
||
# 定义线性变换层
|
||
self.q_proj = nn.Linear(args.hidden_size, args.num_attention_heads * self.head_dim, bias=False) # 查询变换
|
||
self.k_proj = nn.Linear(args.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) # 键变换
|
||
self.v_proj = nn.Linear(args.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) # 值变换
|
||
self.o_proj = nn.Linear(args.num_attention_heads * self.head_dim, args.hidden_size, bias=False) # 输出变换
|
||
|
||
# Dropout层
|
||
self.attn_dropout = nn.Dropout(args.dropout) # 注意力权重的dropout
|
||
self.resid_dropout = nn.Dropout(args.dropout) # 残差连接的dropout
|
||
self.dropout = args.dropout
|
||
|
||
# 是否使用Flash Attention优化
|
||
self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn
|
||
|
||
def forward(self,
|
||
x: torch.Tensor,
|
||
position_embeddings: Tuple[torch.Tensor, torch.Tensor], # 位置编码的cos和sin值
|
||
past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # KV缓存
|
||
use_cache=False, # 是否使用KV缓存
|
||
attention_mask: Optional[torch.Tensor] = None): # 注意力掩码
|
||
"""前向传播函数
|
||
|
||
实现多头注意力的核心计算逻辑，包括：
|
||
1. 线性变换生成查询、键、值向量
|
||
2. 应用旋转位置编码
|
||
3. 处理KV缓存（用于加速自回归生成）
|
||
4. 执行注意力计算（支持Flash Attention优化）
|
||
"""
|
||
bsz, seq_len, _ = x.shape
|
||
# 生成查询、键、值向量
|
||
xq, xk, xv = self.q_proj(x), self.k_proj(x), self.v_proj(x)
|
||
# 重塑张量维度以适应多头注意力格式
|
||
xq = xq.view(bsz, seq_len, self.n_local_heads, self.head_dim)
|
||
xk = xk.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim)
|
||
xv = xv.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim)
|
||
|
||
# 应用旋转位置编码
|
||
cos, sin = position_embeddings
|
||
xq, xk = apply_rotary_pos_emb(xq, xk, cos[:seq_len], sin[:seq_len])
|
||
|
||
# 处理KV缓存，用于加速自回归生成
|
||
if past_key_value is not None:
|
||
xk = torch.cat([past_key_value[0], xk], dim=1) # 拼接历史键向量
|
||
xv = torch.cat([past_key_value[1], xv], dim=1) # 拼接历史值向量
|
||
past_kv = (xk, xv) if use_cache else None
|
||
|
||
# 调整维度顺序并重复键值向量以匹配查询头数量
|
||
xq, xk, xv = (
|
||
xq.transpose(1, 2), # [batch, num_heads, seq_len, head_dim]
|
||
repeat_kv(xk, self.n_rep).transpose(1, 2), # 重复键向量
|
||
repeat_kv(xv, self.n_rep).transpose(1, 2) # 重复值向量
|
||
)
|
||
|
||
if self.flash and seq_len != 1:
|
||
dropout_p = self.dropout if self.training else 0.0
|
||
attn_mask = None
|
||
if attention_mask is not None:
|
||
attn_mask = attention_mask.view(bsz, 1, 1, -1).expand(bsz, self.n_local_heads, seq_len, -1)
|
||
attn_mask = attn_mask.bool() if attention_mask is not None else None
|
||
|
||
output = F.scaled_dot_product_attention(xq, xk, xv, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=True)
|
||
else:
|
||
scores = (xq @ xk.transpose(-2, -1)) / math.sqrt(self.head_dim)
|
||
scores = scores + torch.triu(
|
||
torch.full((seq_len, seq_len), float("-inf"), device=scores.device),
|
||
diagonal=1
|
||
).unsqueeze(0).unsqueeze(0) # scores+mask
|
||
|
||
if attention_mask is not None:
|
||
extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
|
||
extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
|
||
scores = scores + extended_attention_mask
|
||
|
||
scores = F.softmax(scores.float(), dim=-1).type_as(xq)
|
||
scores = self.attn_dropout(scores)
|
||
output = scores @ xv
|
||
|
||
output = output.transpose(1, 2).reshape(bsz, seq_len, -1)
|
||
output = self.resid_dropout(self.o_proj(output))
|
||
return output, past_kv
|
||
|
||
|
||
class FeedForward(nn.Module):
|
||
"""前馈神经网络实现
|
||
|
||
使用SwiGLU激活函数的变体，包含门控机制，可以更好地控制信息流动
|
||
结构：输入 -> (门控路径 + 上投影路径) -> 下投影 -> 输出
|
||
"""
|
||
def __init__(self, config: MiniMindConfig):
|
||
super().__init__()
|
||
# 如果未指定中间层维度，则根据隐藏层维度计算
|
||
if config.intermediate_size is None:
|
||
intermediate_size = int(config.hidden_size * 8 / 3) # 设置为隐藏层维度的8/3倍
|
||
config.intermediate_size = 64 * ((intermediate_size + 64 - 1) // 64) # 调整为64的倍数
|
||
|
||
# 定义三个线性变换层
|
||
self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False) # 门控路径
|
||
self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) # 降维投影
|
||
self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False) # 升维投影
|
||
|
||
self.dropout = nn.Dropout(config.dropout) # dropout层
|
||
self.act_fn = ACT2FN[config.hidden_act] # 激活函数（默认为SiLU）
|
||
|
||
def forward(self, x):
|
||
# 实现SwiGLU的变体：激活(门控路径) * 上投影 -> 下投影 -> dropout
|
||
return self.dropout(self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)))
|
||
|
||
|
||
class MoEGate(nn.Module):
|
||
"""混合专家（MoE）的门控机制实现
|
||
|
||
负责为每个token选择最合适的专家组合，包括：
|
||
1. 计算每个token与专家的相关性分数
|
||
2. 选择top-k个专家处理每个token
|
||
3. 计算辅助损失以平衡专家的使用
|
||
"""
|
||
def __init__(self, config: MiniMindConfig):
|
||
super().__init__()
|
||
self.config = config
|
||
self.top_k = config.num_experts_per_tok # 每个token选择的专家数量
|
||
self.n_routed_experts = config.n_routed_experts # 可路由的专家总数
|
||
|
||
self.scoring_func = config.scoring_func # 专家选择的评分函数
|
||
self.alpha = config.aux_loss_alpha # 辅助损失的权重
|
||
self.seq_aux = config.seq_aux # 是否在序列级别计算辅助损失
|
||
|
||
self.norm_topk_prob = config.norm_topk_prob # 是否对top-k概率进行归一化
|
||
self.gating_dim = config.hidden_size # 用于门控的特征维度
|
||
self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim))) # 门控权重矩阵
|
||
self.reset_parameters() # 初始化参数
|
||
|
||
def reset_parameters(self) -> None:
|
||
import torch.nn.init as init
|
||
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
|
||
|
||
def forward(self, hidden_states):
|
||
bsz, seq_len, h = hidden_states.shape
|
||
hidden_states = hidden_states.view(-1, h)
|
||
logits = F.linear(hidden_states, self.weight, None)
|
||
if self.scoring_func == 'softmax':
|
||
scores = logits.softmax(dim=-1)
|
||
else:
|
||
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)
|
||
|
||
if self.top_k > 1 and self.norm_topk_prob:
|
||
denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
|
||
topk_weight = topk_weight / denominator
|
||
|
||
if self.training and self.alpha > 0.0:
|
||
scores_for_aux = scores
|
||
aux_topk = self.top_k
|
||
topk_idx_for_aux_loss = topk_idx.view(bsz, -1)
|
||
if self.seq_aux:
|
||
scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1)
|
||
ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device)
|
||
ce.scatter_add_(1, topk_idx_for_aux_loss,
|
||
torch.ones(bsz, seq_len * aux_topk, device=hidden_states.device)).div_(
|
||
seq_len * aux_topk / self.n_routed_experts)
|
||
aux_loss = (ce * scores_for_seq_aux.mean(dim=1)).sum(dim=1).mean() * self.alpha
|
||
else:
|
||
mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts)
|
||
ce = mask_ce.float().mean(0)
|
||
Pi = scores_for_aux.mean(0)
|
||
fi = ce * self.n_routed_experts
|
||
aux_loss = (Pi * fi).sum() * self.alpha
|
||
else:
|
||
aux_loss = 0
|
||
return topk_idx, topk_weight, aux_loss
|
||
|
||
|
||
class MOEFeedForward(nn.Module):
|
||
"""混合专家（MoE）前馈网络实现
|
||
|
||
包含多个专家网络和一个门控机制，每个token的处理会被动态路由到最合适的专家组合
|
||
可选地包含共享专家，它们会处理所有token
|
||
"""
|
||
def __init__(self, config: MiniMindConfig):
|
||
super().__init__()
|
||
self.config = config
|
||
# 初始化可路由的专家网络
|
||
self.experts = nn.ModuleList([
|
||
FeedForward(config)
|
||
for _ in range(config.n_routed_experts)
|
||
])
|
||
# 初始化门控机制
|
||
self.gate = MoEGate(config)
|
||
# 可选：初始化共享专家网络
|
||
if config.n_shared_experts > 0:
|
||
self.shared_experts = nn.ModuleList([
|
||
FeedForward(config)
|
||
for _ in range(config.n_shared_experts)
|
||
])
|
||
|
||
def forward(self, x):
|
||
identity = x
|
||
orig_shape = x.shape
|
||
bsz, seq_len, _ = x.shape
|
||
# 使用门控机制选择专家
|
||
topk_idx, topk_weight, aux_loss = self.gate(x)
|
||
x = x.view(-1, x.shape[-1])
|
||
flat_topk_idx = topk_idx.view(-1)
|
||
if self.training:
|
||
x = x.repeat_interleave(self.config.num_experts_per_tok, dim=0)
|
||
y = torch.empty_like(x, dtype=torch.float16)
|
||
for i, expert in enumerate(self.experts):
|
||
y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]).to(y.dtype) # 确保类型一致
|
||
y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
|
||
y = y.view(*orig_shape)
|
||
else:
|
||
y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
|
||
if self.config.n_shared_experts > 0:
|
||
for expert in self.shared_experts:
|
||
y = y + expert(identity)
|
||
self.aux_loss = aux_loss
|
||
return y
|
||
|
||
@torch.no_grad()
|
||
def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
|
||
expert_cache = torch.zeros_like(x)
|
||
idxs = flat_expert_indices.argsort()
|
||
tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)
|
||
token_idxs = idxs // self.config.num_experts_per_tok
|
||
# 当tokens_per_expert = [6, 15, 20, 26]，tokens_per_expert.shape[0]即为专家数量（此时为4）
|
||
# 且token_idxs = [3, 7, 19, 21, 24, 25, 4, 5, 6, 10, 11, 12...] 时
|
||
# 意味token_idxs[:6] -> [3, 7, 19, 21, 24, 25]这6个位置属于专家0处理的token（每个token有可能被多个专家处理，这取决于num_experts_per_tok）
|
||
# 接下来9个位置token_idxs[6:15] -> [4, 5, 6, 10, 11, 12...]属于专家1处理的token...依此类推
|
||
for i, end_idx in enumerate(tokens_per_expert):
|
||
start_idx = 0 if i == 0 else tokens_per_expert[i - 1]
|
||
if start_idx == end_idx:
|
||
continue
|
||
expert = self.experts[i]
|
||
exp_token_idx = token_idxs[start_idx:end_idx]
|
||
expert_tokens = x[exp_token_idx]
|
||
expert_out = expert(expert_tokens).to(expert_cache.dtype)
|
||
expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
|
||
expert_cache.scatter_add_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out)
|
||
|
||
return expert_cache
|
||
|
||
|
||
class MiniMindBlock(nn.Module):
|
||
def __init__(self, layer_id: int, config: MiniMindConfig):
|
||
super().__init__()
|
||
self.num_attention_heads = config.num_attention_heads
|
||
self.hidden_size = config.hidden_size
|
||
self.head_dim = config.hidden_size // config.num_attention_heads
|
||
self.self_attn = Attention(config)
|
||
|
||
self.layer_id = layer_id
|
||
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||
self.mlp = FeedForward(config) if not config.use_moe else MOEFeedForward(config)
|
||
|
||
def forward(self, hidden_states, position_embeddings, past_key_value=None, use_cache=False, attention_mask=None):
|
||
residual = hidden_states
|
||
hidden_states, present_key_value = self.self_attn(
|
||
self.input_layernorm(hidden_states), position_embeddings,
|
||
past_key_value, use_cache, attention_mask
|
||
)
|
||
hidden_states += residual
|
||
hidden_states = hidden_states + self.mlp(self.post_attention_layernorm(hidden_states))
|
||
return hidden_states, present_key_value
|
||
|
||
|
||
class MiniMindModel(nn.Module):
|
||
"""MiniMind模型的主体实现
|
||
|
||
一个基于Transformer的语言模型，支持以下特性：
|
||
1. 分组查询注意力（GQA）降低计算开销
|
||
2. 旋转位置编码（RoPE）处理位置信息
|
||
3. 混合专家机制（MoE）提升模型容量
|
||
4. RMSNorm归一化提升训练稳定性
|
||
"""
|
||
def __init__(self, config: MiniMindConfig):
|
||
super().__init__()
|
||
self.config = config
|
||
# 基础配置
|
||
self.vocab_size, self.num_hidden_layers = config.vocab_size, config.num_hidden_layers
|
||
# 词嵌入层
|
||
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
|
||
self.dropout = nn.Dropout(config.dropout)
|
||
# 初始化Transformer层
|
||
self.layers = nn.ModuleList([MiniMindBlock(l, config) for l in range(self.num_hidden_layers)])
|
||
# 最终的归一化层
|
||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||
|
||
# 预计算旋转位置编码的三角函数值
|
||
freqs_cos, freqs_sin = precompute_freqs_cis(dim=config.hidden_size // config.num_attention_heads,
|
||
end=config.max_position_embeddings, theta=config.rope_theta)
|
||
# 注册为非持久化缓冲区（不会被保存到模型文件中）
|
||
self.register_buffer("freqs_cos", freqs_cos, persistent=False)
|
||
self.register_buffer("freqs_sin", freqs_sin, persistent=False)
|
||
|
||
def forward(self,
|
||
input_ids: Optional[torch.Tensor] = None,
|
||
attention_mask: Optional[torch.Tensor] = None,
|
||
past_key_values: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
|
||
use_cache: bool = False,
|
||
**kwargs):
|
||
batch_size, seq_length = input_ids.shape
|
||
past_key_values = past_key_values or [None] * len(self.layers)
|
||
start_pos = past_key_values[0][0].shape[1] if past_key_values[0] is not None else 0
|
||
|
||
hidden_states = self.dropout(self.embed_tokens(input_ids))
|
||
|
||
position_embeddings = (
|
||
self.freqs_cos[start_pos:start_pos + seq_length],
|
||
self.freqs_sin[start_pos:start_pos + seq_length]
|
||
)
|
||
|
||
presents = []
|
||
for layer_idx, (layer, past_key_value) in enumerate(zip(self.layers, past_key_values)):
|
||
hidden_states, present = layer(
|
||
hidden_states,
|
||
position_embeddings,
|
||
past_key_value=past_key_value,
|
||
use_cache=use_cache,
|
||
attention_mask=attention_mask
|
||
)
|
||
presents.append(present)
|
||
|
||
hidden_states = self.norm(hidden_states)
|
||
|
||
aux_loss = sum(
|
||
layer.mlp.aux_loss
|
||
for layer in self.layers
|
||
if isinstance(layer.mlp, MOEFeedForward)
|
||
)
|
||
|
||
return hidden_states, presents, aux_loss
|
||
|
||
|
||
class MiniMindForCausalLM(PreTrainedModel, GenerationMixin):
|
||
config_class = MiniMindConfig
|
||
|
||
def __init__(self, config: MiniMindConfig = None):
|
||
self.config = config or MiniMindConfig()
|
||
super().__init__(self.config)
|
||
self.model = MiniMindModel(self.config)
|
||
self.lm_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=False)
|
||
self.model.embed_tokens.weight = self.lm_head.weight
|
||
self.OUT = CausalLMOutputWithPast()
|
||
|
||
def forward(self,
|
||
input_ids: Optional[torch.Tensor] = None,
|
||
attention_mask: Optional[torch.Tensor] = None,
|
||
past_key_values: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
|
||
use_cache: bool = False,
|
||
logits_to_keep: Union[int, torch.Tensor] = 0,
|
||
**args):
|
||
h, past_kvs, aux_loss = self.model(
|
||
input_ids=input_ids,
|
||
attention_mask=attention_mask,
|
||
past_key_values=past_key_values,
|
||
use_cache=use_cache,
|
||
**args
|
||
)
|
||
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
||
logits = self.lm_head(h[:, slice_indices, :])
|
||
self.OUT.__setitem__('last_hidden_state', h)
|
||
self.OUT.__setitem__('logits', logits)
|
||
self.OUT.__setitem__('aux_loss', aux_loss)
|
||
self.OUT.__setitem__('past_key_values', past_kvs)
|
||
return self.OUT
|