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
|
"""Transformer decoder module."""
from typing import Optional
from torch import Tensor, nn
from .attention import Attention
from .embedding.rotary import RotaryEmbedding
class Decoder(nn.Module):
def __init__(
self,
dim: int,
ff_mult: int,
heads: int,
dim_head: int,
depth: int,
dropout_rate: float = 0.0,
) -> None:
super().__init__()
self.norm = nn.LayerNorm(dim)
self.layers = nn.ModuleList(
[
nn.ModuleList(
[
Attention(
dim=dim,
heads=heads,
causal=True,
dim_head=dim_head,
ff_mult=ff_mult,
dropout_rate=dropout_rate,
use_flash=True,
norm_context=False,
rotary_emb=RotaryEmbedding(dim_head),
),
Attention(
dim=dim,
heads=heads,
causal=False,
dim_head=dim_head,
ff_mult=ff_mult,
dropout_rate=dropout_rate,
use_flash=True,
norm_context=False,
),
]
)
for _ in range(depth)
]
)
def self_attn(self, x: Tensor, mask: Tensor) -> Tensor:
for self_attn, _ in self.layers:
x = x + self_attn(x, mask=mask)
return self.norm(x)
def forward(
self,
x: Tensor,
context: Tensor,
mask: Optional[Tensor] = None,
) -> Tensor:
"""Applies decoder block on input signals."""
for self_attn, cross_attn in self.layers:
x = x + self_attn(x, mask=mask)
x = x + cross_attn(x, context=context)
return self.norm(x)
|
