forked from juncongmoo/pyllama
-
Notifications
You must be signed in to change notification settings - Fork 0
/
theHack_2.py
144 lines (124 loc) · 5.65 KB
/
theHack_2.py
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
import os
if os.path.exists("theHack"):
os.chdir("theHack")
else:
print("Please run theHack_1.py first")
exit()
import torch
import torch.nn as nn
def precompute_freqs_cis(dim: int = 128, end: int = 128, theta: float = 10000.0, device = torch.device('cpu')):
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
t = torch.arange(end, device=device) # type: ignore
freqs = torch.outer(t, freqs).float() # type: ignore
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
print('freqs_cis ', freqs_cis.shape)
return freqs_cis
def precompute_mask(seqlen = 128, device = torch.device('cpu')):
mask = torch.full((1, 1, seqlen, seqlen), float("-inf"), device=device)
mask = torch.triu(mask, diagonal= 1)
print('mask ', mask.shape)
return mask
freqs_cis = precompute_freqs_cis()
mask = precompute_mask()
from theHack.BadTransformerLLM import myBadTransfomerBlock, myBadTransformerUnit
def hackTheTransformer(id = 0, epochs = 4096, device = 'cuda:0'):
print('hackTheTransformer id ', id)
theTransformer = torch.load('theTransformerLayer%d.pth' % id, map_location=device)
try:
theBadTransformer = torch.load('BadTransformer/theBadTransformerBlock%d.pth' % id, map_location=device)
print('Loaded BadTransformer/theBadTransformerBlock%d.pth' % id)
except Exception as e:
print(e)
print('BadTransformer/theBadTransformerBlock%d.pth not found' % id)
theBadTransformer = myBadTransfomerBlock().to(device)
myFreqs = freqs_cis.clone().to(device)
myMask = mask.clone().to(device)
if not os.path.exists('BadTransformer'):
os.mkdir('BadTransformer')
myOptimizer = torch.optim.Adam(theBadTransformer.parameters(), lr=0.0001)
myLoss = torch.nn.L1Loss()
for _ in range(epochs):
dummyInput = torch.rand(1, 128, 4096, device=device)
# respondFromTheTransformer = theTransformerA(dummyInput, myFreqs, myMask)
# respondFromTheTransformer = theTransformerB(respondFromTheTransformer, myFreqs, myMask)
# respondFromTheTransformer = theTransformerC(respondFromTheTransformer, myFreqs, myMask)
# respondFromTheTransformer = theTransformerD(respondFromTheTransformer, myFreqs, myMask)
respondFromTheTransformer = theTransformer(dummyInput, myFreqs, myMask)
respondFromTheBadTransformer = theBadTransformer(dummyInput)
loss = myLoss(respondFromTheTransformer, respondFromTheBadTransformer)
myOptimizer.zero_grad()
loss.backward()
myOptimizer.step()
print('HackTheTransformer %d loss %.6f' % (id, loss.item()))
torch.save(theBadTransformer, 'BadTransformer/theBadTransformerBlock%d.pth' % id)
open('BadTransformer/theBadTransformerBlock%d.loss' % id, 'w').write(str(loss.item()))
print('HackTheTransformer %d done' % id)
def hackMultiTransformer(num = 2, id = 0, epochs = 4096, device = 'cuda:0'):
print('hackMultiTransformer num %d id %d' % (num, id))
TransID = id * num
Transformers = []
for i in range(num):
Transformers.append(torch.load('theTransformerLayer%d.pth' % (TransID + i), map_location=device))
print('Load theTransformerLayer%d.pth' % (TransID + i))
try:
theBadTransformer = torch.load('BadTransformer/theBadTransformerUnit%d.pth' % id, map_location=device)
print('Loaded BadTransformer/theBadTransformerUnit%d.pth' % id)
except Exception as e:
print(e)
print('No BadTransformer/theBadTransformerUnit%d.pth' % id)
theBadTransformer = myBadTransformerUnit().to(device)
myFreqs = freqs_cis.clone().to(device)
myMask = mask.clone().to(device)
if not os.path.exists('BadTransformer'):
os.mkdir('BadTransformer')
myOptimizer = torch.optim.Adam(theBadTransformer.parameters(), lr=0.0001)
myLoss = torch.nn.L1Loss()
for _ in range(epochs):
dummyInput = torch.rand(1, 128, 4096, device=device)
respondFromTheTransformer = dummyInput
for i in Transformers:
respondFromTheTransformer = i(respondFromTheTransformer, myFreqs, myMask)
respondFromTheBadTransformer = theBadTransformer(dummyInput)
loss = myLoss(respondFromTheTransformer, respondFromTheBadTransformer)
myOptimizer.zero_grad()
loss.backward()
myOptimizer.step()
print('HackMultiTransformer %d loss %.6f' % (id, loss.item()))
torch.save(theBadTransformer, 'BadTransformer/theBadTransformerUnit%d.pth' % id)
open('BadTransformer/theBadTransformerUnit%d.loss' % id, 'w').write(str(loss.item()))
print('HackMultiTransformer %d done' % id)
#hackTheTransformer(0)
#hackTheTransformer(1, 8192)
#hackMultiTransformer(2, 0, 32768)
#hackMultiTransformer(2, 1, 32768)
hackMultiTransformer(2, 2, 32768)
# import threading
# threadlist = []
# for i in range(8):
# threadlist.append(threading.Thread(target=hackTheTransformer, args=(i, 32768, 'cuda:%d' % i)))
# for i in threadlist:
# i.start()
# for i in threadlist:
# i.join()
# threadlist = []
# for i in range(8, 16):
# threadlist.append(threading.Thread(target=hackTheTransformer, args=(i, 32768, 'cuda:%d' % (i % 8))))
# for i in threadlist:
# i.start()
# for i in threadlist:
# i.join()
# threadlist = []
# for i in range(16, 24):
# threadlist.append(threading.Thread(target=hackTheTransformer, args=(i, 32768, 'cuda:%d' % (i % 8))))
# for i in threadlist:
# i.start()
# for i in threadlist:
# i.join()
# threadlist = []
# for i in range(24, 32):
# threadlist.append(threading.Thread(target=hackTheTransformer, args=(i, 32768, 'cuda:%d' % (i % 8))))
# for i in threadlist:
# i.start()
# for i in threadlist:
# i.join()
# print('All done')