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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
| import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Tuple, Optional
class ScaledDotProductAttention(nn.Module):
"""Attention(Q,K,V) = softmax(QK^T / sqrt(dk)) * V"""
def __init__(self, dropout_p=0.1):
super().__init__()
self.dropout = nn.Dropout(dropout_p)
def forward(self, Q, K, V, mask=None):
d_k = Q.size(-1)
scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(d_k)
if mask is not None:
scores = scores.masked_fill(mask == 0, float('-inf'))
attn_weights = self.dropout(F.softmax(scores, dim=-1))
return torch.matmul(attn_weights, V), attn_weights
class MultiHeadAttention(nn.Module):
"""MultiHead(Q,K,V) = Concat(head_1,...,head_h) * W^O"""
def __init__(self, d_model=512, n_heads=8, dropout_p=0.1):
super().__init__()
assert d_model % n_heads == 0
self.d_model = d_model
self.n_heads = n_heads
self.d_k = d_model // n_heads
self.W_Q = nn.Linear(d_model, d_model)
self.W_K = nn.Linear(d_model, d_model)
self.W_V = nn.Linear(d_model, d_model)
self.W_O = nn.Linear(d_model, d_model)
self.attention = ScaledDotProductAttention(dropout_p)
self.dropout = nn.Dropout(dropout_p)
def forward(self, Q, K, V, mask=None):
batch_size = Q.size(0)
Q, K, V = self.W_Q(Q), self.W_K(K), self.W_V(V)
Q = Q.view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
K = K.view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
V = V.view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
if mask is not None:
mask = mask.unsqueeze(1).unsqueeze(1)
attn_out, attn_w = self.attention(Q, K, V, mask=mask)
attn_out = attn_out.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model)
return self.dropout(self.W_O(attn_out)), attn_w
class SelfAttention(nn.Module):
"""Q=K=V，序列内部相互关注"""
def __init__(self, d_model=512, n_heads=8, dropout_p=0.1):
super().__init__()
self.mha = MultiHeadAttention(d_model, n_heads, dropout_p)
def forward(self, x, mask=None):
return self.mha(x, x, x, mask=mask)
class CrossAttention(nn.Module):
"""Q来自decoder, K/V来自encoder"""
def __init__(self, d_model=512, n_heads=8, dropout_p=0.1):
super().__init__()
self.mha = MultiHeadAttention(d_model, n_heads, dropout_p)
def forward(self, query, key, value, mask=None):
return self.mha(query, key, value, mask=mask)
class PositionalEncoding(nn.Module):
"""PE(pos,2i)=sin(pos/10000^(2i/d)), PE(pos,2i+1)=cos(...)"""
def __init__(self, d_model, max_len=5000, dropout_p=0.1):
super().__init__()
self.dropout = nn.Dropout(dropout_p)
pe = torch.zeros(max_len, d_model)
pos = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(pos * div)
pe[:, 1::2] = torch.cos(pos * div)
self.register_buffer('pe', pe.unsqueeze(0))
def forward(self, x):
return self.dropout(x + self.pe[:, :x.size(1), :])
class GroupedQueryAttention(nn.Module):
"""GQA: 分组共享KV，GPT-3.5/Llama2使用"""
def __init__(self, d_model=512, n_heads=8, n_groups=4, dropout_p=0.1):
super().__init__()
assert n_heads % n_groups == 0
self.d_model, self.n_heads, self.n_groups = d_model, n_heads, n_groups
self.d_k = d_model // n_heads
self.W_Q = nn.Linear(d_model, d_model)
self.W_K = nn.Linear(d_model, n_groups * self.d_k)
self.W_V = nn.Linear(d_model, n_groups * self.d_k)
self.W_O = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout_p)
def forward(self, Q, K, V, mask=None):
bs = Q.size(0)
q = self.W_Q(Q).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
k = self.W_K(K).view(bs, -1, self.n_groups, self.d_k).transpose(1, 2)
v = self.W_V(V).view(bs, -1, self.n_groups, self.d_k).transpose(1, 2)
rf = self.n_heads // self.n_groups
k, v = k.repeat_interleave(rf, dim=1), v.repeat_interleave(rf, dim=1)
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
if mask is not None:
scores = scores.masked_fill(mask == 0, float('-inf'))
aw = F.softmax(scores, dim=-1)
out = torch.matmul(self.dropout(aw), v)
out = out.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
return self.W_O(out)
class LinearAttention(nn.Module):
"""复杂度O(n*d)，Attention=phi(Q)dot(phi(K)^T dot V)"""
def __init__(self, d_model=512, n_heads=8, dropout_p=0.1, nonlinearity='elu'):
super().__init__()
assert d_model % n_heads == 0
self.d_model = d_model
self.n_heads, self.d_k = n_heads, d_model // n_heads
self.W_Q = nn.Linear(d_model, d_model)
self.W_K = nn.Linear(d_model, d_model)
self.W_V = nn.Linear(d_model, d_model)
self.W_O = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout_p)
self.nonlinearity = nonlinearity
def _fm(self, x):
return F.elu(x) + 1 if self.nonlinearity == 'elu' else F.relu(x)
def forward(self, Q, K, V, mask=None):
bs = Q.size(0)
q = self._fm(self.W_Q(Q).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2))
k = self._fm(self.W_K(K).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2))
v = self.W_V(V).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
kv = torch.matmul(k.transpose(-2, -1), v)
out = torch.matmul(q, kv) / q.sum(dim=-1, keepdim=True).clamp(min=1e-8)
out = out.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
return self.W_O(out)
class RelativePositionAttention(nn.Module):
"""score(i,j) = Q_i dot K_j^T/sqrt(d) + B_r(i-j)"""
def __init__(self, d_model=512, n_heads=8, max_rel_pos=20, dropout_p=0.1):
super().__init__()
self.d_model, self.n_heads, self.max_rel_pos = d_model, n_heads, max_rel_pos
self.d_k = d_model // n_heads
self.W_Q = nn.Linear(d_model, d_model)
self.W_K = nn.Linear(d_model, d_model)
self.W_V = nn.Linear(d_model, d_model)
self.W_O = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout_p)
self.rel_pos_bias = nn.Parameter(torch.randn(2*max_rel_pos+1, n_heads) * 0.02)
def _rp_idx(self, seq_len):
return (torch.arange(seq_len).unsqueeze(1) - torch.arange(seq_len).unsqueeze(0) + self.max_rel_pos)
def forward(self, Q, K, V, mask=None):
bs, sl = Q.size(0), Q.size(1)
q = self.W_Q(Q).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
k = self.W_K(K).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
v = self.W_V(V).view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
rel_bias = self.rel_pos_bias[self._rp_idx(sl)].permute(2, 0, 1).unsqueeze(0).expand(bs, -1, -1, -1)
scores = scores + rel_bias
if mask is not None:
scores = scores.masked_fill(mask == 0, float('-inf'))
aw = F.softmax(scores, dim=-1)
out = torch.matmul(self.dropout(aw), v)
out = out.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
return self.W_O(out)
class CausalMultiHeadAttention(nn.Module):
"""带因果掩码的多头注意力，确保每个位置只能看到之前的token"""
def __init__(self, d_model=512, n_heads=8, dropout_p=0.1):
super().__init__()
self.d_model, self.n_heads, self.d_k = d_model, n_heads, d_model // n_heads
self.W_Q = nn.Linear(d_model, d_model)
self.W_K = nn.Linear(d_model, d_model)
self.W_V = nn.Linear(d_model, d_model)
self.W_O = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout_p)
def forward(self, x, mask=None):
bs, sl = x.size(0), x.size(1)
Q, K, V = self.W_Q(x), self.W_K(x), self.W_V(x)
Q = Q.view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
K = K.view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
V = V.view(bs, -1, self.n_heads, self.d_k).transpose(1, 2)
scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.d_k)
causal = torch.tril(torch.ones(sl, sl, device=x.device)).unsqueeze(0).unsqueeze(0)
scores = scores.masked_fill(causal == 0, float('-inf'))
if mask is not None:
scores = scores.masked_fill(mask == 0, float('-inf'))
aw = F.softmax(scores, dim=-1)
out = torch.matmul(self.dropout(aw), V)
out = out.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
return self.W_O(out), aw
def create_masks(seq_len_q, seq_len_k, batch_size, device='cpu'):
"""创建填充掩码和因果掩码"""
src_mask = torch.ones(batch_size, 1, 1, seq_len_k, device=device)
nopeak = torch.tril(torch.ones(1, seq_len_q, seq_len_q, device=device))
return src_mask, nopeak.expand(batch_size, -1, -1, -1)
class TransformerAttentionBlock(nn.Module):
"""完整Transformer块: MHA + Residual + LN + FFN + Residual + LN"""
def __init__(self, d_model=512, n_heads=8, dropout_p=0.1, ffn_dim=2048):
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout_p, batch_first=True)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.ffn = nn.Sequential(nn.Linear(d_model, ffn_dim), nn.ReLU(),
nn.Linear(ffn_dim, d_model), nn.Dropout(dropout_p))
self.dropout = nn.Dropout(dropout_p)
def forward(self, x, mask=None):
x = x + self.dropout(self.self_attn(x, x, x, key_padding_mask=mask)[0])
x = self.norm1(x)
x = self.norm2(x + self.ffn(x))
return x
def main():
batch, seq_len, d_model = 2, 10, 64
x = torch.randn(batch, seq_len, d_model)
print("=" * 60)
print("注意力机制完整实现 - 测试")
print("=" * 60)
attn = ScaledDotProductAttention()
Q, K, V = torch.randn(batch, 5, d_model), torch.randn(batch, 8, d_model), torch.randn(batch, 8, d_model)
out, w = attn(Q, K, V)
print(f"[1] Scaled Dot-Product: Q:{Q.shape} K:{K.shape} V:{V.shape} -> out:{out.shape} w:{w.shape} OK")
mha = MultiHeadAttention(d_model, 8)
out, w = mha(x, x, x)
print(f"[2] Multi-Head: {x.shape} -> out:{out.shape} w:{w.shape} OK")
sa = SelfAttention(d_model)
out, w = sa(x)
print(f"[3] Self-Attention: {x.shape} -> out:{out.shape} w:{w.shape} OK")
ca = CrossAttention(d_model)
q = torch.randn(batch, 6, d_model)
out, w = ca(q, x, x)
print(f"[4] Cross-Attention: q:{q.shape} k/v:{x.shape} -> out:{out.shape} w:{w.shape} OK")
pe = PositionalEncoding(d_model)
enc = pe(x)
print(f"[5] Positional Encoding: {x.shape} -> {enc.shape} OK")
gqa = GroupedQueryAttention(d_model, 8, 4)
out = gqa(x, x, x)
print(f"[6] GQA: {x.shape} -> {out.shape} OK")
la = LinearAttention(d_model)
out = la(x, x, x)
print(f"[7] Linear Attention: {x.shape} -> {out.shape} OK")
rpa = RelativePositionAttention(d_model, 8, 10)
out = rpa(x, x, x)
print(f"[8] Relative Position: {x.shape} -> {out.shape} OK")
ca_attn = CausalMultiHeadAttention(d_model)
out, w = ca_attn(x)
upper_sum = w[0, 0, :5, 5:].sum()
print(f"[9] Causal Attention: {x.shape} -> out:{out.shape} w:{w.shape} upper_tri:{upper_sum:.4f}(should=0) OK")
src_m, tgt_m = create_masks(seq_len, seq_len, batch)
print(f" Masks: src:{src_m.shape} tgt:{tgt_m.shape} OK")
block = TransformerAttentionBlock(d_model)
out = block(x)
print(f" Transformer Block: {x.shape} -> {out.shape} OK")
print("\n" + "=" * 60)
print("参数量对比 (d_model=512, n_heads=8)")
print("=" * 60)
for name, model_cls, kwargs in [
("Multi-Head", MultiHeadAttention, {"d_model": 512}),
("GQA (4 grp)", GroupedQueryAttention, {"d_model": 512, "n_groups": 4}),
("Linear Attn", LinearAttention, {"d_model": 512}),
("Rel Position", RelativePositionAttention, {"d_model": 512}),
("Causal MH", CausalMultiHeadAttention, {"d_model": 512}),
]:
m = model_cls(**kwargs)
p = sum(pp.numel() for pp in m.parameters())
print(f" {name:16s}: {p:>10,} params")
print("\n" + "=" * 60)
print("ALL TESTS PASSED!")
print("=" * 60)
if __name__ == "__main__":
main()
|
