File size: 8,047 Bytes
46df05d
 
eec6c0e
 
 
 
 
46df05d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e55e319
 
 
 
 
 
8771224
e55e319
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8771224
e55e319
 
 
 
 
 
8771224
 
 
 
 
 
e55e319
 
 
 
 
 
 
 
46df05d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
# Copyright (c) 2023, Tri Dao.

""""
The implementation was adopted from
https://github.com/Dao-AILab/flash-attention/blob/43950dda456e095969d842fca7a73c5bfe3cecd0
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributed import ProcessGroup


try:
    from flash_attn.ops.activations import swiglu
except ImportError:
    swiglu = None

try:
    from flash_attn.ops.fused_dense import ColumnParallelLinear, RowParallelLinear
except ImportError:
    ColumnParallelLinear, RowParallelLinear = None, None

try:
    from flash_attn.ops.fused_dense import FusedMLP, ParallelFusedMLP
except ImportError:
    FusedMLP, ParallelFusedMLP = None, None


class GLUMLP(nn.Module):
    def __init__(
            self,
            in_features,
            hidden_features,
            activation,
            use_flash_attn,
            return_residual=False,
            hidden_dropout_prob=0.1
    ):
        super().__init__()
        self.hidden_features = hidden_features
        self.gated_layers = nn.Linear(
            in_features, hidden_features * 2, bias=False
        )
        if activation == 'relu':
            self.act = nn.ReLU()
        elif activation == 'gelu':
            self.act = nn.GELU()
        else:
            raise ValueError(
                f"activation {activation} not supported"
            )
        self.wo = nn.Linear(hidden_features, in_features)
        self.dropout = nn.Dropout(hidden_dropout_prob)
        self.return_residual = return_residual
        self.use_flash_attn = use_flash_attn
        #self.layernorm = nn.LayerNorm(in_features, eps=layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        residual_connection = hidden_states
        # compute the activation
        hidden_states = self.gated_layers(hidden_states)
        if self.use_flash_attn:
            gated = hidden_states[:, : self.hidden_features]
            non_gated = hidden_states[:, self.hidden_features :]
        else:
            gated = hidden_states[:, :, : self.hidden_features]
            non_gated = hidden_states[:, :, self.hidden_features :]
        hidden_states = self.act(gated) * non_gated
        hidden_states = self.dropout(hidden_states)
        # multiply by the second matrix
        hidden_states = self.wo(hidden_states)
        # add the residual connection and post-LN
        # hidden_states = self.layernorm(hidden_states + residual_connection)
        return hidden_states if not self.return_residual else (hidden_states, residual_connection)

class Mlp(nn.Module):
    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        activation=F.gelu,
        bias1=True,
        bias2=True,
        return_residual=False,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        out_features = out_features if out_features is not None else in_features
        hidden_features = hidden_features if hidden_features is not None else in_features * 4
        self.return_residual = return_residual
        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias1, **factory_kwargs)
        self.activation = activation
        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)

    def forward(self, x):
        y = self.fc1(x)
        y = self.activation(y)
        y = self.fc2(y)
        return y if not self.return_residual else (y, x)


class ParallelMLP(nn.Module):
    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        activation=F.gelu,
        process_group: ProcessGroup = None,
        sequence_parallel=True,
        bias1=True,
        bias2=True,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        assert ColumnParallelLinear is not None, "Need to install fused_dense"
        assert RowParallelLinear is not None, "Need to install fused_dense"
        out_features = out_features if out_features is not None else in_features
        hidden_features = hidden_features if hidden_features is not None else in_features * 4
        self.fc1 = ColumnParallelLinear(
            in_features,
            hidden_features,
            process_group,
            bias=bias1,
            sequence_parallel=sequence_parallel,
            **factory_kwargs,
        )
        self.activation = activation
        self.fc2 = RowParallelLinear(
            hidden_features,
            out_features,
            process_group,
            bias=bias2,
            sequence_parallel=sequence_parallel,
            **factory_kwargs,
        )

    def forward(self, x):
        y = self.fc1(x)
        y = self.activation(y)
        y = self.fc2(y)
        return y


class GatedMlp(nn.Module):
    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        activation=F.sigmoid,
        bias1=True,
        bias2=True,
        multiple_of=128,
        return_residual=False,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        out_features = out_features if out_features is not None else in_features
        hidden_features = (
            hidden_features if hidden_features is not None else int(8 * in_features / 3)
        )
        hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
        self.return_residual = return_residual
        self.fc1 = nn.Linear(in_features, 2 * hidden_features, bias=bias1, **factory_kwargs)
        self.activation = activation
        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)

    def forward(self, x):
        y = self.fc1(x)
        if self.activation == F.sigmoid:  # Special case for GLU
            y = F.glu(y, dim=-1)
        elif self.activation == F.silu and swiglu is not None:  # Special case for SwiGLU
            y, gate = y.chunk(2, dim=-1)
            y = swiglu(gate, y)
        else:
            y, gate = y.chunk(2, dim=-1)
            y = y * self.activation(gate)
        y = self.fc2(y)
        return y if not self.return_residual else (y, x)


class ParallelGatedMlp(nn.Module):
    """Parallel GatedMlp"""

    def __init__(
        self,
        in_features,
        process_group,
        hidden_features=None,
        out_features=None,
        activation=F.sigmoid,
        bias1=True,
        bias2=True,
        multiple_of=128,
        sequence_parallel=True,
        device=None,
        dtype=None,
    ):
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        out_features = out_features if out_features is not None else in_features
        hidden_features = (
            hidden_features if hidden_features is not None else int(8 * in_features / 3)
        )
        hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
        if ColumnParallelLinear is None or RowParallelLinear is None:
            raise ImportError("fused_dense is not installed")
        self.fc1 = ColumnParallelLinear(
            in_features,
            2 * hidden_features,
            process_group,
            bias=bias1,
            sequence_parallel=sequence_parallel,
            **factory_kwargs,
        )
        self.activation = activation
        self.fc2 = RowParallelLinear(
            hidden_features,
            out_features,
            process_group,
            bias=bias2,
            sequence_parallel=sequence_parallel,
            **factory_kwargs,
        )

    def forward(self, x):
        y = self.fc1(x)
        if self.activation == F.sigmoid:  # Special case for GLU
            y = F.glu(y, dim=-1)
        else:
            y, gate = y.chunk(2, dim=-1)
            y = y * self.activation(gate)
        y = self.fc2(y)
        return y