File size: 6,081 Bytes
9ae1b1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch


class SVDLinearAddition(torch.nn.Module):
    def __init__(self, linear, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        w, b = linear.weight, linear.bias

        self.bias = b

        with torch.no_grad():
            U, S, Vt = torch.linalg.svd(w, full_matrices=False)
            self.U = torch.nn.Parameter(U)
            self.S = torch.nn.Parameter(S)
            self.Vt = torch.nn.Parameter(Vt)
            self.epsilon = torch.nn.Parameter(torch.zeros_like(S))

    def forward(self, x):
        w = self.U @ torch.diag(self.S + self.epsilon) @ self.Vt
        x = torch.nn.functional.linear(x, w, self.bias)
        return x


class SVDLinearAdditionQKV(torch.nn.Module):
    def __init__(self, linear, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        w, b = linear.weight, linear.bias

        self.bias = b

        with torch.no_grad():
            c = w.shape[0] // 3
            q = w[:c]
            k = w[c : 2 * c]
            v = w[2 * c :]
            U_q, S_q, Vt_q = torch.linalg.svd(q, full_matrices=False)
            self.U_q = torch.nn.Parameter(U_q)
            self.S_q = torch.nn.Parameter(S_q)
            self.Vt_q = torch.nn.Parameter(Vt_q)
            self.epsilon_q = torch.nn.Parameter(torch.zeros_like(S_q))
            U_k, S_k, Vt_k = torch.linalg.svd(k, full_matrices=False)
            self.U_k = torch.nn.Parameter(U_k)
            self.S_k = torch.nn.Parameter(S_k)
            self.Vt_k = torch.nn.Parameter(Vt_k)
            self.epsilon_k = torch.nn.Parameter(torch.zeros_like(S_k))
            U_v, S_v, Vt_v = torch.linalg.svd(v, full_matrices=False)
            self.U_v = torch.nn.Parameter(U_v)
            self.S_v = torch.nn.Parameter(S_v)
            self.Vt_v = torch.nn.Parameter(Vt_v)
            self.epsilon_v = torch.nn.Parameter(torch.zeros_like(S_v))

    def forward(self, x):
        w = torch.concatenate(
            (
                self.U_q @ torch.diag(self.S_q + self.epsilon_q) @ self.Vt_q,
                self.U_k @ torch.diag(self.S_k + self.epsilon_k) @ self.Vt_k,
                self.U_v @ torch.diag(self.S_v + self.epsilon_v) @ self.Vt_v,
            )
        )
        x = torch.nn.functional.linear(x, w, self.bias)
        return x


def replace_linear_addition_noqk(module, name):
    for attr_str in dir(module):
        target_attr = getattr(module, attr_str)
        if type(target_attr) == torch.nn.Linear:
            if 'qkv' in attr_str:
                print('replaced: ', name, attr_str)
                svd_linear_qkv = SVDLinearAdditionQKV(target_attr)
                svd_linear_qkv.U_q.requires_grad = False
                svd_linear_qkv.U_k.requires_grad = False
                svd_linear_qkv.U_v.requires_grad = False
                svd_linear_qkv.S_q.requires_grad = False
                svd_linear_qkv.S_k.requires_grad = False
                svd_linear_qkv.S_v.requires_grad = False
                svd_linear_qkv.Vt_q.requires_grad = False
                svd_linear_qkv.Vt_k.requires_grad = False
                svd_linear_qkv.Vt_v.requires_grad = False
                svd_linear_qkv.epsilon_q.requires_grad = False
                svd_linear_qkv.epsilon_k.requires_grad = False
                svd_linear_qkv.epsilon_v.requires_grad = True
                svd_linear_qkv.bias.requires_grad = False
                setattr(module, attr_str, svd_linear_qkv)
            else:
                print('replaced: ', name, attr_str)
                svd_linear = SVDLinearAddition(target_attr)
                svd_linear.U.requires_grad = False
                svd_linear.S.requires_grad = False
                svd_linear.Vt.requires_grad = False
                svd_linear.bias.requires_grad = False
                svd_linear.epsilon.requires_grad = True
                setattr(module, attr_str, svd_linear)

    for name, immediate_child_module in module.named_children():
        replace_linear_addition_noqk(immediate_child_module, name)


def replace_back(module, name):
    for attr_str in dir(module):
        target_attr = getattr(module, attr_str)

        if type(target_attr) == SVDLinearAddition:
            print('replaced back: ', name, attr_str)
            with torch.no_grad():
                linear = torch.nn.Linear(
                    target_attr.Vt.shape[1],
                    target_attr.U.shape[0],
                    device=target_attr.U.device,
                )
                linear.weight.add_(
                    target_attr.U
                    @ torch.diag(target_attr.S + target_attr.epsilon)
                    @ target_attr.Vt
                    - linear.weight
                )
                linear.bias.add_(target_attr.bias - linear.bias)

            setattr(module, attr_str, linear)

        elif type(target_attr) == SVDLinearAdditionQKV:
            print('replaced back: ', name, attr_str)
            with torch.no_grad():
                w = torch.concatenate(
                    (
                        target_attr.U_q
                        @ torch.diag(target_attr.S_q + target_attr.epsilon_q)
                        @ target_attr.Vt_q,
                        target_attr.U_k
                        @ torch.diag(target_attr.S_k + target_attr.epsilon_k)
                        @ target_attr.Vt_k,
                        target_attr.U_v
                        @ torch.diag(target_attr.S_v + target_attr.epsilon_v)
                        @ target_attr.Vt_v,
                    )
                )
                linear = torch.nn.Linear(
                    w.shape[1], w.shape[0], device=target_attr.U_q.device
                )
                linear.weight.add_(w - linear.weight)
                linear.bias.add_(target_attr.bias - linear.bias)

            setattr(module, attr_str, linear)

    # iterate through immediate child modules. Note, the recursion is done by our code no need to use named_modules()
    for name, immediate_child_module in module.named_children():
        replace_back(immediate_child_module, name)