File size: 6,086 Bytes
612b16b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import torch
from torch import nn
from typing import Literal

class LogitMixingAttention (nn.Module) :
  def __init__ (self,embed_dim,num_head,mode:Literal['scaled','kfactor']=None,factor_init=1e-4,causal_mask=False) :
    super().__init__()
    self.mode = mode
    if self.mode == None :
      raise RuntimeError(f"the mode must be specified kfactor/scaled")

    self.causal_mask = causal_mask
    self.lg = nn.Linear(embed_dim,embed_dim,bias=False)
    self.lo = nn.Linear(embed_dim,embed_dim,bias=False)
    self.factor = nn.Parameter(
        data=torch.normal(mean=0,std=factor_init,size=(1,1,embed_dim))
    )
    self.num_head = num_head
    self.dim_k = embed_dim//num_head

  def split_head (self,x : torch.Tensor) :
    b,s,d = x.shape
    x = torch.reshape(x,(b,s,self.num_head,self.dim_k))
    return x.permute(0,2,1,3)

  def scaled_attention (self,q,k,v):
    b,s,d = q.shape
    if self.causal_mask :
      mask = 1 - torch.tril(torch.ones((s,s)),diagonal=0).to(q.device)
      mask = mask * -1e9
      mask = mask.unsqueeze(0)
    else :
      mask = None
    lm = q + (self.factor * (k + v))
    lm = self.lg(lm)
    lm = self.split_head(lm)
    k = self.split_head(k)
    v = self.split_head(v)
    score = torch.matmul(lm,k.transpose(-1,-2)) / self.dim_k ** 0.5
    if mask is not None :
      score = score + mask
    attn = nn.functional.softmax(score,dim=-1)
    attn = torch.matmul(attn,v)
    attn = attn.permute(0,2,1,3)
    attn = torch.reshape(attn,(b,s,d))
    out = self.lo(attn)
    return out

  def kfactor_attention (self,q,k,v) :
    b,s,d = q.shape
    score = q + (self.factor* k ) / self.dim_k ** 0.5
    score = self.split_head(score)
    v = self.lg(v)
    v = self.split_head(v)
    score = nn.functional.softmax(score,dim=-1)
    attn = (score * v) + v
    attn = attn.permute(0,2,1,3)
    attn = torch.reshape(attn,(b,s,d))
    out = self.lo(attn)
    return out

  def forward(self,q,k,v) :
    if self.mode == 'scaled' :
      return self.scaled_attention(q,k,v)
    elif self.mode == 'kfactor' :
      return self.kfactor_attention(q,k,v)
    else :
      raise RuntimeError(f"the mode must be specified kfactor/scaled")



class LCM (nn.Module) :
  def __init__ (self,embed_dim,drop_rate) :
    super().__init__()
    self.step1 = nn.Linear(embed_dim,embed_dim,bias=False)
    self.step2 = nn.Linear(embed_dim,embed_dim,bias=False)
    self.gelu1 = nn.GELU(approximate='tanh')
    self.gelu2 = nn.GELU(approximate='tanh')
    self.mg = nn.Linear(embed_dim,embed_dim,bias=False)
    self.tanh = nn.Tanh()
    self.norm = nn.RMSNorm(embed_dim)
    self.drop = nn.Dropout(drop_rate)

  def forward(self,x):
    z = self.norm(x)
    step1 = self.step1(z)
    step1 = self.gelu1(step1)
    step2 = self.step2(z)
    step2 = self.gelu2(step2)
    mx = step1 + step2
    mx = self.drop(mx)
    mx = self.mg(mx)
    mx = self.tanh(mx)
    return x + mx

class GlobalRouterBlock (nn.Module) :
  def __init__ (self,embed_dim,hidden_dim,num_expert) :
    super().__init__()
    self.Linear1 = nn.Linear(embed_dim,hidden_dim)
    self.linear2 = nn.Linear(hidden_dim,num_expert)

  def forward(self,x) :
    x = x[:,-1,:]
    x = self.Linear1(x)
    x = self.linear2(x)
    return x


class TransformersBlock (nn.Module) :
  def __init__ (self,embed_dim,drop_rate) :
    super().__init__()
    self.attention = LogitMixingAttention(embed_dim=embed_dim,num_head=embed_dim//64,mode='kfactor')
    self.norm = nn.RMSNorm(embed_dim)
    self.dropout = nn.Dropout(drop_rate)
    self.lcm = LCM(embed_dim=embed_dim,drop_rate=drop_rate)

  def forward (self,x) :
    z = self.norm(x)
    attn = self.attention(z,z,z)
    attn = self.dropout(attn)
    x = x + attn
    x = self.lcm(x)
    return x

class LCTLM(nn.Module) :
  def __init__ (self,embed_dim,drop_rate) :
    super().__init__()
    self.block1 = TransformersBlock(embed_dim,drop_rate)
    self.block2 = TransformersBlock(embed_dim,drop_rate)
    self.block3 = TransformersBlock(embed_dim,drop_rate)
    self.block4 = TransformersBlock(embed_dim,drop_rate)

  def forward(self,x,idx) :
    if idx == 0 :
      x = self.block1(x)
    elif idx == 1 :
      x = self.block2(x)
    elif idx == 2 :
      x = self.block3(x)
    else :
      x = self.block4(x)

    return x

class LCTLM2 (nn.Module) :
  def __init__ (self,vocab_size = 30001,embed_dim=640,drop_rate=0.1,maxpos=250,temperature=1.2) :
    super().__init__()
    self.temperature = temperature
    self.embedding = nn.Embedding(vocab_size,embed_dim)
    self.pos_embedding = nn.Embedding(maxpos,embed_dim)
    self.lctlm1 = LCTLM(embed_dim,drop_rate)
    self.lctlm2 = LCTLM(embed_dim,drop_rate)
    self.lctlm3 = LCTLM(embed_dim,drop_rate)
    self.lctlm4 = LCTLM(embed_dim,drop_rate)
    self.lctlm5 = LCTLM(embed_dim,drop_rate)
    self.lctlm6 = LCTLM(embed_dim,drop_rate)
    self.lctlm7 = LCTLM(embed_dim,drop_rate)
    self.lctlm8 = LCTLM(embed_dim,drop_rate)
    self.global_ffn = nn.Sequential(
        nn.Linear(embed_dim,embed_dim*4,bias=False),
        nn.GELU(approximate='tanh'),
        nn.Linear(embed_dim*4,embed_dim,bias=False)
    )
    self.routers = GlobalRouterBlock(embed_dim=embed_dim,hidden_dim=128,num_expert=8*4)
    self.fn = nn.Linear(embed_dim,vocab_size,bias=False)
    self.scale = embed_dim ** 0.5


  def forward(self,x) :
    b,s = x.shape
    x = self.embedding(x)
    x =  x * self.scale
    pos = torch.arange(s,device=x.device)
    pos = self.pos_embedding(pos)
    pos = pos.unsqueeze(0)
    x = x + pos
    r = self.routers(x)
    r = r / self.temperature
    _,idx = torch.topk(r,k=8)
    idx = idx[0]//8
    x = self.lctlm1(x,idx[0])
    x = self.lctlm2(x,idx[1])
    x = self.lctlm3(x,idx[2])
    x = self.lctlm4(x,idx[3])
    x = self.lctlm5(x,idx[4])
    x = self.lctlm6(x,idx[5])
    x = self.lctlm7(x,idx[6])
    x = self.lctlm8(x,idx[7])

    x = self.global_ffn(x)
    x = self.fn(x)
    return x