File size: 4,477 Bytes
4196369
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import torch
from torch import nn, cat, is_tensor
import torch.nn.functional as F
from torch.nn import Module, ModuleList
from torch.utils._pytree import tree_map

from einops.layers.torch import Rearrange

from rotary_embedding_torch import RotaryEmbedding

# functions

def exists(v):
    return v is not None

# classes

class ImplicitMLPAttention(Module):
    def __init__(
        self,
        dim,
        mlp_hiddens: tuple[int, ...],
        *,
        activation = nn.SiLU(),
        heads = 8,
        talking_heads = True,
        prenorm = True,
        keys_rmsnorm = True # https://openreview.net/forum?id=HkztQWZfl2
    ):
        super().__init__()
        assert isinstance(mlp_hiddens, tuple) and len(mlp_hiddens) >= 2
        dim_mlp_in, *dim_mlp_inner, dim_mlp_out = mlp_hiddens

        self.norm = nn.RMSNorm(dim) if prenorm else nn.Identity()

        dim_query_inner = dim_mlp_in * heads
        self.to_queries = nn.Linear(dim, dim_query_inner, bias = False)

        # keys and values

        self.rotary_embed = RotaryEmbedding(min(mlp_hiddens)) # just use the minimum dimension, the rest is partially rotaried

        # each key value forms an implicit weight (memory) of (dim_key, dim_values)
        # chaining them would then be the implicit MLP from TTT / Titans

        self.keys = ModuleList([])
        self.key_norms = ModuleList([])

        self.values = ModuleList([])

        for dim_in, dim_out in zip(mlp_hiddens[:-1], mlp_hiddens[1:]):

            dim_keys_inner = dim_in * heads
            dim_values_inner = dim_out * heads

            keys = nn.Linear(dim, dim_keys_inner, bias = False)
            key_norms = nn.RMSNorm(dim_in) if keys_rmsnorm else nn.Identity()

            values = nn.Linear(dim, dim_values_inner, bias = False)

            self.keys.append(keys)
            self.key_norms.append(key_norms)

            self.values.append(values)

        self.activation = activation

        # talking head - Shazeer et al.

        self.talking_heads = nn.Identity()

        if talking_heads and len(dim_mlp_inner) > 0:
            self.talking_heads = nn.Conv2d(heads, heads, 1, bias = False)
            nn.init.dirac_(self.talking_heads.weight)

        # split merging of heads

        self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
        self.merge_heads = Rearrange('b h n d -> b n (h d)')

        self.to_out = nn.Linear(dim_mlp_out * heads, dim, bias = False)

    def forward(
        self,
        tokens,
        cache = None,
        return_kv_cache = False
    ):
        batch, seq_len, device = *tokens.shape[:2], tokens.device

        tokens = self.norm(tokens)

        queries = self.to_queries(tokens)

        keys = [fn(tokens) for fn in self.keys]
        values = [fn(tokens) for fn in self.values]

        # split heads for input as well as all keys, values that form the implicit weights

        queries, keys, values = tree_map(self.split_heads, (queries, keys, values))

        # maybe norm all keys

        keys = [norm(k) for norm, k in zip(self.key_norms, keys)]

        # cache

        if exists(cache):
            cache_keys, cache_values = cache

            keys = [cat(args, dim = -2) for args in zip(cache_keys, keys)]
            values = [cat(args, dim = -2) for args in zip(cache_values, values)]

        # attend

        def attend(q, k, v):
            q, k = self.rotary_embed.rotate_queries_with_cached_keys(q, k)

            return F.scaled_dot_product_attention(q, k, v, is_causal = True)

        # implicit memory mlp

        out = queries

        for i, (key, value) in enumerate(zip(keys, values), start = 1):
            is_last = i == len(keys)

            out = attend(out, key, value)

            if not is_last:
                out = self.talking_heads(out)
                out = self.activation(out)

        # merge heads

        out = self.merge_heads(out)
        out = self.to_out(out)

        if not return_kv_cache:
            return out

        return out, (keys, values)

# 3 layers implicit MLP attention - 64 -> 128 -> 128 -> 64 w/ relu

if __name__ == '__main__':

    implicit_mlp_attn = ImplicitMLPAttention(
        512,
        (64, 128, 128, 64),
        activation = nn.ReLU()
    )

    tokens = torch.randn(1, 1024, 512)

    out, cache = implicit_mlp_attn(tokens)
    out, cache = implicit_mlp_attn(tokens, cache = cache)

    assert out.shape == tokens.shape