File size: 7,393 Bytes
c295391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# MIT License

# Copyright (c) Microsoft

# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:

# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.

# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

# Copyright (c) [2025] [Microsoft]
# SPDX-License-Identifier: MIT
from typing import *
import torch
import torch.nn as nn
import torch.nn.functional as F
from .full_attn import scaled_dot_product_attention


class MultiHeadRMSNorm(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: torch.Tensor) -> torch.Tensor:
        return (F.normalize(x.float(), dim = -1) * self.gamma * self.scale).to(x.dtype)


class RotaryPositionEmbedder(nn.Module):
    def __init__(self, hidden_size: int, in_channels: int = 3):
        super().__init__()
        assert hidden_size % 2 == 0, "Hidden size must be divisible by 2"
        self.hidden_size = hidden_size
        self.in_channels = in_channels
        self.freq_dim = hidden_size // in_channels // 2
        self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
        self.freqs = 1.0 / (10000 ** self.freqs)
        
    def _get_phases(self, indices: torch.Tensor) -> torch.Tensor:
        self.freqs = self.freqs.to(indices.device)
        phases = torch.outer(indices, self.freqs)
        phases = torch.polar(torch.ones_like(phases), phases)
        return phases
        
    def _rotary_embedding(self, x: torch.Tensor, phases: torch.Tensor) -> torch.Tensor:
        x_complex = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
        x_rotated = x_complex * phases
        x_embed = torch.view_as_real(x_rotated).reshape(*x_rotated.shape[:-1], -1).to(x.dtype)
        return x_embed
        
    def forward(self, q: torch.Tensor, k: torch.Tensor, indices: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            q (sp.SparseTensor): [..., N, D] tensor of queries
            k (sp.SparseTensor): [..., N, D] tensor of keys
            indices (torch.Tensor): [..., N, C] tensor of spatial positions
        """
        if indices is None:
            indices = torch.arange(q.shape[-2], device=q.device)
            if len(q.shape) > 2:
                indices = indices.unsqueeze(0).expand(q.shape[:-2] + (-1,))
        
        phases = self._get_phases(indices.reshape(-1)).reshape(*indices.shape[:-1], -1)
        if phases.shape[1] < self.hidden_size // 2:
            phases = torch.cat([phases, torch.polar(
                torch.ones(*phases.shape[:-1], self.hidden_size // 2 - phases.shape[1], device=phases.device),
                torch.zeros(*phases.shape[:-1], self.hidden_size // 2 - phases.shape[1], device=phases.device)
            )], dim=-1)
        q_embed = self._rotary_embedding(q, phases)
        k_embed = self._rotary_embedding(k, phases)
        return q_embed, k_embed
    

class MultiHeadAttention(nn.Module):
    def __init__(
        self,
        channels: int,
        num_heads: int,
        ctx_channels: Optional[int]=None,
        type: Literal["self", "cross"] = "self",
        attn_mode: Literal["full", "windowed"] = "full",
        window_size: Optional[int] = None,
        shift_window: Optional[Tuple[int, int, int]] = 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", "windowed"], f"Invalid attention mode: {attn_mode}"
        assert type == "self" or attn_mode == "full", "Cross-attention only supports full attention"
        
        if attn_mode == "windowed":
            raise NotImplementedError("Windowed attention is not yet implemented")
        
        self.channels = channels
        self.head_dim = channels // num_heads
        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_window = shift_window
        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 = MultiHeadRMSNorm(self.head_dim, num_heads)
            self.k_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
            
        self.to_out = nn.Linear(channels, channels)

        if use_rope:
            self.rope = RotaryPositionEmbedder(channels)
    
    def forward(self, x: torch.Tensor, context: Optional[torch.Tensor] = None, indices: Optional[torch.Tensor] = None) -> torch.Tensor:
        B, L, C = x.shape
        if self._type == "self":
            qkv = self.to_qkv(x)
            qkv = qkv.reshape(B, L, 3, self.num_heads, -1)
            if self.use_rope:
                q, k, v = qkv.unbind(dim=2)
                q, k = self.rope(q, k, indices)
                qkv = torch.stack([q, k, v], dim=2)
            if self.attn_mode == "full":
                if self.qk_rms_norm:
                    q, k, v = qkv.unbind(dim=2)
                    q = self.q_rms_norm(q)
                    k = self.k_rms_norm(k)
                    h = scaled_dot_product_attention(q, k, v)
                else:
                    h = scaled_dot_product_attention(qkv)
            elif self.attn_mode == "windowed":
                raise NotImplementedError("Windowed attention is not yet implemented")
        else:
            Lkv = context.shape[1]
            q = self.to_q(x)
            kv = self.to_kv(context)
            q = q.reshape(B, L, self.num_heads, -1)
            kv = kv.reshape(B, Lkv, 2, self.num_heads, -1)
            if self.qk_rms_norm:
                q = self.q_rms_norm(q)
                k, v = kv.unbind(dim=2)
                k = self.k_rms_norm(k)
                h = scaled_dot_product_attention(q, k, v)
            else:
                h = scaled_dot_product_attention(q, kv)
        h = h.reshape(B, L, -1)
        h = self.to_out(h)
        return h