File size: 6,621 Bytes
72c0672
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates.

from dataclasses import dataclass, field
from enum import Enum
from typing import Optional, Tuple

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

from lingua.transformer import InitStdFactor, RMSNorm
from lingua.probe import log_stats

from apps.fastRNN.component.rnn_common import conv1d, scan


@dataclass
class BaseMinLSTMArgs:
    dim: int = 512
    n_layers: int = 8
    n_heads: int = 1

    multiple_of: int = 256
    ffn_dim_multiplier: Optional[float] = None

    conv_size: Optional[int] = None

    norm_eps: float = 1e-5

    init_base_std: Optional[float] = None
    init_std_factor: str = "disabled"


class LSTM(nn.Module):
    def __init__(
        self,
        dim: int,
        hidden_dim: int,  # h_t dim (state expansion)
        n_heads: int,
        multiple_of: int,
        ffn_dim_multiplier: Optional[float],
        conv_size: Optional[int] = None,
    ):
        super().__init__()

        hidden_dim = int(2 * hidden_dim / 3)
        if ffn_dim_multiplier is not None:
            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
        assert (
            hidden_dim % n_heads == 0
        ), f"Hidden dim must be divisible by n_heads: {hidden_dim} % {n_heads} != 0"

        self.dim = dim
        self.hidden_dim = hidden_dim

        self.n_heads = n_heads
        self.head_dim = hidden_dim // n_heads

        self.conv_size = conv_size
        if conv_size is not None:
            assert ((self.hidden_dim) % 8 == 0) and (
                conv_size in [2, 3, 4]
            ), f"Causal conv1d only supports conv_size in [2, 3, 4] and hidden_dim % 8 == 0, got {self.hidden_dim} and {conv_size}"
            self.conv_dim = 2 * self.hidden_dim
            self.conv_weight = nn.Parameter(torch.empty((self.conv_dim, conv_size)))

        self.w = nn.Linear(
            dim,
            hidden_dim,
            bias=False,
        )

        self.wfi = nn.Linear(
            dim,
            2 * hidden_dim,
            bias=False,
        )

        self.wh_tilde = nn.Linear(
            dim,
            hidden_dim,
            bias=False,
        )

        self.wo = nn.Linear(
            hidden_dim,
            dim,
            bias=False,
        )

    def forward(
        self, x: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
    ) -> torch.Tensor:
        bsz, seq_len, _ = x.shape

        w0 = self.w(x.view_as(x))

        fi = self.wfi(x.view_as(x)).transpose(1, 2)
        h_tilde = self.wh_tilde(x.view_as(x)).transpose(1, 2)

        if self.conv_size is not None:
            conv1d_w = log_stats(self.conv_weight, "conv1d.w")
            fi = conv1d(
                x=fi,
                conv_weight=conv1d_w,
                tok_idx=tok_idx,
                cu_seqlens=cu_seqlens,
                impl=impl,
                cache=self.cache.conv_cache if hasattr(self, "cache") else None,
            )

        fi = fi.reshape(bsz * self.n_heads, 2 * self.head_dim, seq_len)
        h_tilde = h_tilde.reshape(bsz * self.n_heads, self.head_dim, seq_len)

        f, i = fi.chunk(2, dim=1)
        f, i = F.sigmoid(f), F.sigmoid(i)
        denom = 1 / (f + i + 1e-4)

        h = scan(
            a=(f * denom),
            b=(h_tilde * i * denom),
            cu_seqlens=cu_seqlens,
            impl=impl,
            cache=self.cache.state_cache if hasattr(self, "cache") else None,
        )

        h = h.view(bsz, self.hidden_dim, seq_len).transpose(1, 2)
        h = log_stats(h, "hidden_state")

        h = h * F.silu(w0)

        out = self.wo(h)

        return out

    def reset_parameters(self, init_std, factor):
        in_init_std = init_std or (self.dim ** (-0.5))
        out_init_std = init_std or (self.hidden_dim ** (-0.5))
        out_init_std = out_init_std / factor

        for w in [self.w, self.wfi, self.wh_tilde]:
            nn.init.trunc_normal_(
                w.weight, std=in_init_std, a=-3 * in_init_std, b=3 * in_init_std
            )

        nn.init.trunc_normal_(
            self.wo.weight, std=out_init_std, a=-3 * in_init_std, b=3 * in_init_std
        )

        if self.conv_size is not None:
            conv_std = init_std or (self.conv_size ** (-0.5))
            nn.init.trunc_normal_(
                self.conv_weight,
                mean=0.0,
                std=conv_std,
                a=-3 * conv_std,
                b=3 * conv_std,
            )


class LSTMBlock(nn.Module):
    def __init__(self, args: BaseMinLSTMArgs):
        super().__init__()

        self.lstm_norm = RMSNorm(args.dim, eps=args.norm_eps)
        self.lstm = LSTM(
            dim=args.dim,
            hidden_dim=3 * args.dim,
            n_heads=args.n_heads,
            multiple_of=args.multiple_of,
            ffn_dim_multiplier=args.ffn_dim_multiplier,
            conv_size=args.conv_size,
        )

    def forward(
        self, x: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
    ) -> torch.Tensor:
        x = x + self.lstm(self.lstm_norm(x), tok_idx=tok_idx, cu_seqlens=cu_seqlens, impl=impl)
        return x

    def init_weights(self, init_std: Optional[float], factor: InitStdFactor):
        self.lstm.reset_parameters(init_std, factor)
        self.lstm_norm.reset_parameters()


class BaseMinLSTM(nn.Module):
    def __init__(self, args: BaseMinLSTMArgs):
        super().__init__()

        self.dim = args.dim
        self.init_base_std = args.init_base_std
        self.init_std_factor = InitStdFactor(args.init_std_factor)

        self.layers = nn.ModuleList()
        for _ in range(args.n_layers):
            self.layers.append(LSTMBlock(args))

    def forward(
        self, x: torch.Tensor, tok_idx: torch.Tensor, cu_seqlens: torch.Tensor, impl: str = "parallel"
    ) -> torch.Tensor:
        for layer in self.layers:
            x = layer(x, tok_idx=tok_idx, cu_seqlens=cu_seqlens, impl=impl)
        return x

    def reset_parameters(self):
        pass

    def init_weights(self):
        self.reset_parameters()
        for depth, layer in enumerate(self.layers):
            factor = {
                InitStdFactor.CURRENT_DEPTH: (2 * (depth + 1)) ** 0.5,
                InitStdFactor.GLOBAL_DEPTH: (2 * (len(self.layers) + 1)) ** 0.5,
                InitStdFactor.DIM_RATIO: self.dim / 4096,
                InitStdFactor.DISABLED: 1.0,
            }[self.init_std_factor]

            layer.init_weights(self.init_base_std, factor)