File size: 5,181 Bytes
db6a3b7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import *
import torch
import torch.nn as nn
from ..basic import SparseTensor
from ..linear import SparseLinear
from ..nonlinearity import SparseGELU
from ..attention import SparseMultiHeadAttention, SerializeMode
from ...norm import LayerNorm32


class SparseFeedForwardNet(nn.Module):
    def __init__(self, channels: int, mlp_ratio: float = 4.0):
        super().__init__()
        self.mlp = nn.Sequential(
            SparseLinear(channels, int(channels * mlp_ratio)),
            SparseGELU(approximate="tanh"),
            SparseLinear(int(channels * mlp_ratio), channels),
        )

    def forward(self, x: SparseTensor) -> SparseTensor:
        return self.mlp(x)


class SparseTransformerBlock(nn.Module):
    """
    Sparse Transformer block (MSA + FFN).
    """
    def __init__(
        self,
        channels: int,
        num_heads: int,
        mlp_ratio: float = 4.0,
        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
        window_size: Optional[int] = None,
        shift_sequence: Optional[int] = None,
        shift_window: Optional[Tuple[int, int, int]] = None,
        serialize_mode: Optional[SerializeMode] = None,
        use_checkpoint: bool = False,
        use_rope: bool = False,
        qk_rms_norm: bool = False,
        qkv_bias: bool = True,
        ln_affine: bool = False,
    ):
        super().__init__()
        self.use_checkpoint = use_checkpoint
        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.attn = SparseMultiHeadAttention(
            channels,
            num_heads=num_heads,
            attn_mode=attn_mode,
            window_size=window_size,
            shift_sequence=shift_sequence,
            shift_window=shift_window,
            serialize_mode=serialize_mode,
            qkv_bias=qkv_bias,
            use_rope=use_rope,
            qk_rms_norm=qk_rms_norm,
        )
        self.mlp = SparseFeedForwardNet(
            channels,
            mlp_ratio=mlp_ratio,
        )

    def _forward(self, x: SparseTensor) -> SparseTensor:
        h = x.replace(self.norm1(x.feats))
        h = self.attn(h)
        x = x + h
        h = x.replace(self.norm2(x.feats))
        h = self.mlp(h)
        x = x + h
        return x

    def forward(self, x: SparseTensor) -> SparseTensor:
        if self.use_checkpoint:
            return torch.utils.checkpoint.checkpoint(self._forward, x, use_reentrant=False)
        else:
            return self._forward(x)


class SparseTransformerCrossBlock(nn.Module):
    """
    Sparse Transformer cross-attention block (MSA + MCA + FFN).
    """
    def __init__(
        self,
        channels: int,
        ctx_channels: int,
        num_heads: int,
        mlp_ratio: float = 4.0,
        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
        window_size: Optional[int] = None,
        shift_sequence: Optional[int] = None,
        shift_window: Optional[Tuple[int, int, int]] = None,
        serialize_mode: Optional[SerializeMode] = None,
        use_checkpoint: bool = False,
        use_rope: bool = False,
        qk_rms_norm: bool = False,
        qk_rms_norm_cross: bool = False,
        qkv_bias: bool = True,
        ln_affine: bool = False,
    ):
        super().__init__()
        self.use_checkpoint = use_checkpoint
        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.norm3 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.self_attn = SparseMultiHeadAttention(
            channels,
            num_heads=num_heads,
            type="self",
            attn_mode=attn_mode,
            window_size=window_size,
            shift_sequence=shift_sequence,
            shift_window=shift_window,
            serialize_mode=serialize_mode,
            qkv_bias=qkv_bias,
            use_rope=use_rope,
            qk_rms_norm=qk_rms_norm,
        )
        self.cross_attn = SparseMultiHeadAttention(
            channels,
            ctx_channels=ctx_channels,
            num_heads=num_heads,
            type="cross",
            attn_mode="full",
            qkv_bias=qkv_bias,
            qk_rms_norm=qk_rms_norm_cross,
        )
        self.mlp = SparseFeedForwardNet(
            channels,
            mlp_ratio=mlp_ratio,
        )

    def _forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor):
        h = x.replace(self.norm1(x.feats))
        h = self.self_attn(h)
        x = x + h
        h = x.replace(self.norm2(x.feats))
        h = self.cross_attn(h, context)
        x = x + h
        h = x.replace(self.norm3(x.feats))
        h = self.mlp(h)
        x = x + h
        return x

    def forward(self, x: SparseTensor, context: torch.Tensor):
        if self.use_checkpoint:
            return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
        else:
            return self._forward(x, context)