File size: 6,103 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
from typing import *
import torch
import torch.nn as nn
import torch.nn.functional as F
from .. import SparseTensor
from .full_attn import sparse_scaled_dot_product_attention
from .serialized_attn import SerializeMode, sparse_serialized_scaled_dot_product_self_attention
from .windowed_attn import sparse_windowed_scaled_dot_product_self_attention
from ...attention import RotaryPositionEmbedder


class SparseMultiHeadRMSNorm(nn.Module):
    def __init__(self, dim: int, heads: int):
        super().__init__()
        self.scale = dim ** 0.5
        self.gamma = nn.Parameter(torch.ones(heads, dim))

    def forward(self, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
        x_type = x.dtype
        x = x.float()
        if isinstance(x, SparseTensor):
            x = x.replace(F.normalize(x.feats, dim=-1))
        else:
            x = F.normalize(x, dim=-1)            
        return (x * self.gamma * self.scale).to(x_type)


class SparseMultiHeadAttention(nn.Module):
    def __init__(
        self,
        channels: int,
        num_heads: int,
        ctx_channels: Optional[int] = None,
        type: Literal["self", "cross"] = "self",
        attn_mode: Literal["full", "serialized", "windowed"] = "full",
        window_size: Optional[int] = None,
        shift_sequence: Optional[int] = None,
        shift_window: Optional[Tuple[int, int, int]] = None,
        serialize_mode: Optional[SerializeMode] = None,
        qkv_bias: bool = True,
        use_rope: bool = False,
        qk_rms_norm: bool = False,
    ):
        super().__init__()
        assert channels % num_heads == 0
        assert type in ["self", "cross"], f"Invalid attention type: {type}"
        assert attn_mode in ["full", "serialized", "windowed"], f"Invalid attention mode: {attn_mode}"
        assert type == "self" or attn_mode == "full", "Cross-attention only supports full attention"
        assert type == "self" or use_rope is False, "Rotary position embeddings only supported for self-attention"
        self.channels = channels
        self.ctx_channels = ctx_channels if ctx_channels is not None else channels
        self.num_heads = num_heads
        self._type = type
        self.attn_mode = attn_mode
        self.window_size = window_size
        self.shift_sequence = shift_sequence
        self.shift_window = shift_window
        self.serialize_mode = serialize_mode
        self.use_rope = use_rope
        self.qk_rms_norm = qk_rms_norm

        if self._type == "self":
            self.to_qkv = nn.Linear(channels, channels * 3, bias=qkv_bias)
        else:
            self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
            self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
        
        if self.qk_rms_norm:
            self.q_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
            self.k_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
            
        self.to_out = nn.Linear(channels, channels)

        if use_rope:
            self.rope = RotaryPositionEmbedder(channels)

    @staticmethod
    def _linear(module: nn.Linear, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
        if isinstance(x, SparseTensor):
            return x.replace(module(x.feats))
        else:
            return module(x)

    @staticmethod
    def _reshape_chs(x: Union[SparseTensor, torch.Tensor], shape: Tuple[int, ...]) -> Union[SparseTensor, torch.Tensor]:
        if isinstance(x, SparseTensor):
            return x.reshape(*shape)
        else:
            return x.reshape(*x.shape[:2], *shape)

    def _fused_pre(self, x: Union[SparseTensor, torch.Tensor], num_fused: int) -> Union[SparseTensor, torch.Tensor]:
        if isinstance(x, SparseTensor):
            x_feats = x.feats.unsqueeze(0)
        else:
            x_feats = x
        x_feats = x_feats.reshape(*x_feats.shape[:2], num_fused, self.num_heads, -1)
        return x.replace(x_feats.squeeze(0)) if isinstance(x, SparseTensor) else x_feats

    def _rope(self, qkv: SparseTensor) -> SparseTensor:
        q, k, v = qkv.feats.unbind(dim=1)   # [T, H, C]
        q, k = self.rope(q, k, qkv.coords[:, 1:])
        qkv = qkv.replace(torch.stack([q, k, v], dim=1)) 
        return qkv
    
    def forward(self, x: Union[SparseTensor, torch.Tensor], context: Optional[Union[SparseTensor, torch.Tensor]] = None) -> Union[SparseTensor, torch.Tensor]:
        if self._type == "self":
            qkv = self._linear(self.to_qkv, x)
            qkv = self._fused_pre(qkv, num_fused=3)
            if self.use_rope:
                qkv = self._rope(qkv)
            if self.qk_rms_norm:
                q, k, v = qkv.unbind(dim=1)
                q = self.q_rms_norm(q)
                k = self.k_rms_norm(k)
                qkv = qkv.replace(torch.stack([q.feats, k.feats, v.feats], dim=1))
            if self.attn_mode == "full":
                h = sparse_scaled_dot_product_attention(qkv)
            elif self.attn_mode == "serialized":
                h = sparse_serialized_scaled_dot_product_self_attention(
                    qkv, self.window_size, serialize_mode=self.serialize_mode, shift_sequence=self.shift_sequence, shift_window=self.shift_window
                )
            elif self.attn_mode == "windowed":
                h = sparse_windowed_scaled_dot_product_self_attention(
                    qkv, self.window_size, shift_window=self.shift_window
                )
        else:
            q = self._linear(self.to_q, x)
            q = self._reshape_chs(q, (self.num_heads, -1))
            kv = self._linear(self.to_kv, context)
            kv = self._fused_pre(kv, num_fused=2)
            if self.qk_rms_norm:
                q = self.q_rms_norm(q)
                k, v = kv.unbind(dim=1)
                k = self.k_rms_norm(k)
                kv = kv.replace(torch.stack([k.feats, v.feats], dim=1))
            h = sparse_scaled_dot_product_attention(q, kv)
        h = self._reshape_chs(h, (-1,))
        h = self._linear(self.to_out, h)
        return h