-
Notifications
You must be signed in to change notification settings - Fork 10
/
gau.py
134 lines (118 loc) · 4.52 KB
/
gau.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.activations import ACT2FN
def rope(x, dim):
"""RoPE position embedding."""
shape = x.shape
if isinstance(dim, int):
dim = [dim]
spatial_shape = [shape[i] for i in dim]
total_len = 1
for i in spatial_shape:
total_len *= i
position = torch.reshape(
torch.arange(total_len, dtype=x.dtype,
device=x.device), spatial_shape
)
for i in range(dim[-1] + 1, len(shape) - 1, 1):
position = position.unsqueeze(-1)
half_size = shape[-1] // 2
freq_seq = -torch.arange(half_size, dtype=x.dtype, device=x.device) / float(
half_size
)
inv_freq = 10000 ** freq_seq
sinusoid = torch.einsum("...,d->...d", position, inv_freq)
sin = sinusoid.sin()
cos = sinusoid.cos()
x1, x2 = torch.chunk(x, 2, dim=-1)
return torch.cat([x1 * cos - x2 * sin, x2 * cos + x1 * sin], dim=-1)
class ScaleNorm(nn.Module):
def __init__(self, eps=1e-5):
super().__init__()
self.eps = eps
self.scala = nn.Parameter(torch.ones(1))
def forward(self, x):
mean_square = (x ** 2).mean(dim=-1, keepdim=True)
x = x * torch.rsqrt(mean_square + self.eps) * self.scala
return x
class GAU(nn.Module):
"""GAU block.
Input shape: batch size x sequence length x model size
"""
def __init__(
self,
hidden_size=768,
expansion_factor=2,
s=128,
norm_type="layer_norm",
eps=1e-5,
hidden_act="silu",
max_position_embeddings=512,
):
super().__init__()
self.s = s
self.e = int(hidden_size * expansion_factor)
self.uv = nn.Linear(hidden_size, 2 * self.e + self.s)
self.weight = nn.Parameter(torch.randn(2, self.s))
self.bias = nn.Parameter(torch.zeros(2, self.s))
self.o = nn.Linear(self.e, hidden_size)
self.LayerNorm = (
nn.LayerNorm(hidden_size, eps=eps)
if norm_type == "layer_norm"
else ScaleNorm(eps=eps)
)
self.w = nn.Parameter(torch.randn(2 * max_position_embeddings - 1))
self.a = nn.Parameter(torch.randn(1, self.s))
self.b = nn.Parameter(torch.randn(1, self.s))
self.act_fn = ACT2FN[hidden_act]
self.max_position_embeddings = max_position_embeddings
nn.init.normal_(self.weight, std=0.02)
nn.init.normal_(self.w, std=0.02)
nn.init.normal_(self.a, std=0.02)
nn.init.normal_(self.b, std=0.02)
def rel_pos_bias(self, seq_len):
"""Relative position bias."""
if seq_len <= 512:
# Construct Toeplitz matrix directly when the sequence length is less than 512
t = F.pad(self.w[: 2 * seq_len - 1], [0, seq_len]).repeat(seq_len)
t = t[..., :-seq_len].reshape(-1, seq_len, 3 * seq_len - 2)
r = (2 * seq_len - 1) // 2
t = t[..., r:-r]
else:
# Construct Toeplitz matrix using RoPE when the sequence length is over 512.
a = rope(self.a.repeat(seq_len, 1), dim=0)
b = rope(self.b.repeat(seq_len, 1), dim=0)
t = torch.einsum("mk,nk ->mn", a, b)
return t
def forward(self, x, attention_mask=None, output_attentions=False, causal=False):
seq_len = x.shape[1]
shortcut, x = x, self.LayerNorm(x)
uv = self.uv(x)
u, v, base = torch.split(self.act_fn(
uv), [self.e, self.e, self.s], dim=-1)
# Generate Query (q) and Key (k) from base.
base = torch.einsum("...r,hr->...hr", base, self.weight) + self.bias
base = rope(base, dim=1)
q, k = torch.unbind(base, dim=-2)
# Calculate the quadratic attention.
qk = torch.einsum("bnd,bmd->bnm", q, k)
bias = self.rel_pos_bias(self.max_position_embeddings)[
:, :seq_len, :seq_len]
kernel = torch.square(torch.relu(
qk / self.max_position_embeddings + bias))
# attention_mask
if attention_mask is not None:
assert attention_mask.ndim == 2
attn_mask = (
attention_mask[:, None, :] * attention_mask[:, :, None]
).type_as(x)
kernel *= attn_mask
if causal:
causal_mask = torch.tril(torch.ones(seq_len, seq_len), diagonal=0)
kernel *= causal_mask
x = u * torch.einsum("bnm,bme->bne", kernel, v)
x = self.o(x)
if output_attentions:
return x + shortcut, kernel
return (x + shortcut,)