File size: 10,514 Bytes
b3fb4dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init

from typing import Optional,Tuple
import math
import logging

logger = logging.getLogger(__name__)


rwkv_emb_scale = 0.4 # try 0.4 for char-level english. try 1.0 for chinese.
rwkv_layer_decay = 1.0 # decay weights in higher layers. try 0.5 ~ 1.0.

class AttentionConfig:
  def __init__(self, ctx_len=100, **kwargs):
    self.ctx_len = ctx_len
    for k,v in kwargs.items():
        setattr(self, k, v)


########################################################################################################
# MHA_rotary: Multi-head Attention + Rotary Encoding + GeGLU FFN
########################################################################################################

class RotaryEmbedding(torch.nn.Module):
    def __init__(self, dim, base=10000):
        super().__init__()
        inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer('inv_freq', inv_freq)
        self.seq_len_cached = None
        self.cos_cached = None
        self.sin_cached = None

    def forward(self, x, seq_len=None):
        if seq_len != self.seq_len_cached:
            self.seq_len_cached = seq_len
            t = torch.arange(seq_len, device=x.device)
            freqs = torch.einsum('i,j->ij', t, self.inv_freq)
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
            self.cos_cached = emb.cos()
            self.sin_cached = emb.sin()
        return torch.stack([self.cos_cached, self.sin_cached])

class ContinuousRotaryEmbedding(torch.nn.Module):
    '''Continuous rotary position embedding'''
    def __init__(self, dim, sequence_scale):
        super().__init__()
        base=10000
        self.sequence_scale = sequence_scale
        self.register_buffer('inv_freq', 1. / (base ** (torch.arange(0, dim, 2))))
    
    def forward(self, t):
        t = (t + 0.5)* self.sequence_scale 
        freqs = torch.einsum('ij,k->ijk', t, self.inv_freq) # freqs: [B, L, dim//2]
        emb = torch.cat((freqs, freqs), dim=-1).unsqueeze(1) # emb: [B, 1, L, dim], 1 for broadcast in head_num dim
        return torch.stack([emb.cos(), emb.sin()])
    
def rotate_half(x):
    x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
    return torch.cat((-x2, x1), -1)

@torch.jit.script
def apply_rotary_pos_emb(q, k, cos, sin):
    cos, sin = cos[...,:q.shape[2],:], sin[...,:q.shape[2],:]
    return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)

class MHA_rotary(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.collect_attention_map = False
        self.attention_map = None
        assert args.encoder_dim % args.num_heads == 0
        self.num_heads = args.num_heads
        self.head_size = args.encoder_dim // args.num_heads

        if args.timeshift:
            self.time_shift = nn.ZeroPad2d((0,0,1,0))

        self.query = nn.Linear(args.encoder_dim, args.encoder_dim)
        self.key = nn.Linear(args.encoder_dim, args.encoder_dim)
        self.value = nn.Linear(args.encoder_dim, args.encoder_dim)

        # self.register_buffer("mask", torch.tril(torch.ones(config.ctx_len, config.ctx_len)))
        
        self.rotary_ndims = int(self.head_size * 0.5)
        
        self.rotary_emb = RotaryEmbedding(self.rotary_ndims)

        self.output = nn.Linear(args.encoder_dim, args.encoder_dim)

    def forward(self, x, RoPE, key_padding_mask=None):
        B, T, C = x.size()

        if hasattr(self, 'time_shift'):
            x = torch.cat([self.time_shift(x)[:, :-1, :C//2], x[:, :, C//2:]], dim = -1)

        q = self.query(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)       # (B, T, C) -> (B, nh, T, hs)
        k = self.key(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)         # (B, T, C) -> (B, nh, T, hs)
        v = self.value(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)       # (B, T, C) -> (B, nh, T, hs)

        q, query_pass = q[..., :self.rotary_ndims], q[..., self.rotary_ndims:]
        k, key_pass = k[..., :self.rotary_ndims], k[..., self.rotary_ndims:]
        
        # cos, sin = self.rotary_emb(q, seq_len=T)
        cos, sin = RoPE
        q, k = apply_rotary_pos_emb(q, k, cos, sin)                                     # rotary encoding
        q = torch.cat((q, query_pass), dim=-1)
        k = torch.cat((k, key_pass), dim=-1)  
        
        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))                 # self-attention: (B, nh, T, hs) * (B, nh, hs, T) -> (B, nh, T, T)
        if key_padding_mask is not None:
            key_padding_mask = key_padding_mask[:, None, None, :]           # (B, T) -> (B, 1, 1, T)
            att = att.masked_fill(key_padding_mask == 0, float('-inf'))
        att = F.softmax(att, dim = -1)                                                  # softmax

        x = att @ v                                                                     # (B, nh, T, T) * (B, nh, T, hs) -> (B, nh, T, hs)
        x = x.transpose(1, 2).contiguous().view(B, T, -1)                               # (B, nh, T, hs) -> (B, T, nh, hs) -> (B, T, C)

        x = self.output(x)

        if self.collect_attention_map:
            self.attention_map = att
        
        return x

class MHA_decoder(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.collect_attention_map = False
        self.attention_map = None
        assert args.encoder_dim % args.num_heads == 0
        self.num_heads = args.num_heads
        self.head_size = args.decoder_dim // args.num_heads

        if args.timeshift:
            self.time_shift = nn.ZeroPad2d((0,0,1,0))

        self.query = nn.Linear(args.decoder_dim, args.decoder_dim)
        self.key = nn.Linear(args.decoder_dim, args.decoder_dim)
        self.value = nn.Linear(args.decoder_dim, args.decoder_dim)

        # self.register_buffer("mask", torch.tril(torch.ones(config.ctx_len, config.ctx_len)))
        
        self.rotary_ndims = int(self.head_size * 0.5)
        
        self.rotary_emb = RotaryEmbedding(self.rotary_ndims)

        self.output = nn.Linear(args.decoder_dim, args.decoder_dim)

    def forward(self, x, memory,RoPE, key_padding_mask=None):
        B, T, C = x.size()
        _, L, M = memory.size()

        # print("x size: ", x.size(), 'memory size: ', memory.size())
        # print('B, T, C: ', B, T, C, 'L: ', L)

        q = self.query(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)       # (B, T, C) -> (B, nh, T, hs)
        k = self.key(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)         # (B, T, C) -> (B, nh, T, hs)
        v = self.value(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)       # (B, T, C) -> (B, nh, T, hs)

        q, query_pass = q[..., :self.rotary_ndims], q[..., self.rotary_ndims:]
        k, key_pass = k[..., :self.rotary_ndims], k[..., self.rotary_ndims:]
        
        # cos, sin = self.rotary_emb(q, seq_len=T)
        cos, sin = RoPE
        q, k = apply_rotary_pos_emb(q, k, cos, sin)                                     # rotary encoding
        q = torch.cat((q, query_pass), dim=-1)
        k = torch.cat((k, key_pass), dim=-1)  
        
        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))                 # self-attention: (B, nh, T, hs) * (B, nh, hs, T) -> (B, nh, T, T)
        if key_padding_mask is not None:
            key_padding_mask = key_padding_mask[:, None, None, :]           # (B, T) -> (B, 1, 1, T)
            att = att.masked_fill(key_padding_mask == 0, float('-inf'))
        att = F.softmax(att, dim = -1)                                                  # softmax

        x = att @ v  
        # print("after attention vals: ", x.shape)                                                                   # (B, nh, T, T) * (B, nh, T, hs) -> (B, nh, T, hs)
        x = x.transpose(1, 2).contiguous().view(B, T, -1)                               # (B, nh, T, hs) -> (B, T, nh, hs) -> (B, T, C)

        # x = self.output(x)

        # print("after linear: ", x.shape)                                                                   # (B, nh, T, T) * (B, nh, T, hs) -> (B, nh, T, hs)


        # cross attention:
        q = self.query(x).view(B, T, self.num_heads, self.head_size).transpose(1, 2)       # (B, T, C) -> (B, nh, T, hs)
        k = self.key(memory).view(B, L, self.num_heads, self.head_size).transpose(1, 2)         # (B, T, C) -> (B, nh, T, hs)
        v = self.value(memory).view(B, L, self.num_heads, self.head_size).transpose(1, 2)       # (B, T, C) -> (B, nh, T, hs)
        
        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))                 # self-attention: (B, nh, T, hs) * (B, nh, hs, T) -> (B, nh, T, T)
        # print("att size: ", att.size())
        if key_padding_mask is not None:
            key_padding_mask = key_padding_mask[:, None, None, :]           # (B, T) -> (B, 1, 1, T)
            att = att.masked_fill(key_padding_mask == 0, float('-inf'))
        att = F.softmax(att, dim = -1)                                                  # softmax

        x = att @ v                                                                     # (B, nh, T, T) * (B, nh, T, hs) -> (B, nh, T, hs)
        # print("x deocder size: ", x.size())
        x = x.transpose(1, 2).contiguous().view(B, T, -1)                               # (B, nh, T, hs) -> (B, T, nh, hs) -> (B, T, C)
        # print("x deocder size transposed: ", x.size())
        x = self.output(x)

        if self.collect_attention_map:
            self.attention_map = att

        return x

    class GeGLU(torch.nn.Module):
        def __init__(self, config, layer_id, time_shift = False):
            super().__init__()
            self.layer_id = layer_id

            if time_shift:
                self.time_shift = nn.ZeroPad2d((0,0,1,0))

            hidden_sz = 3 * config.n_ffn
            self.key = nn.Linear(config.n_embd, hidden_sz)
            self.value = nn.Linear(config.n_embd, hidden_sz)
            self.weight = nn.Linear(hidden_sz, config.n_embd)

        def forward(self, x):
            B, T, C = x.size()
            if hasattr(self, 'time_shift'):
                x = torch.cat([self.time_shift(x)[:, :-1, :C//2], x[:, :, C//2:]], dim = -1)
            
            k = self.key(x)
            v = self.value(x)        
            y = self.weight(F.gelu(k) * v)
            return y