File size: 6,530 Bytes
f08eddf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import importlib.metadata
import math

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

try:
    import flash_attn
    from flash_attn.flash_attn_interface import _flash_attn_forward
    from flash_attn.flash_attn_interface import flash_attn_varlen_func
except ImportError:
    flash_attn = None
    flash_attn_varlen_func = None
    _flash_attn_forward = None


MEMORY_LAYOUT = {
    "flash": (
        lambda x: x.view(x.shape[0] * x.shape[1], *x.shape[2:]),
        lambda x: x,
    ),
    "torch": (
        lambda x: x.transpose(1, 2),
        lambda x: x.transpose(1, 2),
    ),
    "vanilla": (
        lambda x: x.transpose(1, 2),
        lambda x: x.transpose(1, 2),
    ),
}


def get_cu_seqlens(text_mask, img_len):
    """Calculate cu_seqlens_q, cu_seqlens_kv using text_mask and img_len

    Args:
        text_mask (torch.Tensor): the mask of text
        img_len (int): the length of image

    Returns:
        torch.Tensor: the calculated cu_seqlens for flash attention
    """
    batch_size = text_mask.shape[0]
    text_len = text_mask.sum(dim=1)
    max_len = text_mask.shape[1] + img_len

    cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device="cuda")

    for i in range(batch_size):
        s = text_len[i] + img_len
        s1 = i * max_len + s
        s2 = (i + 1) * max_len
        cu_seqlens[2 * i + 1] = s1
        cu_seqlens[2 * i + 2] = s2

    return cu_seqlens


def attention(
    q,
    k,
    v,
    mode="flash",
    drop_rate=0,
    attn_mask=None,
    causal=False,
    cu_seqlens_q=None,
    cu_seqlens_kv=None,
    max_seqlen_q=None,
    max_seqlen_kv=None,
    batch_size=1,
):
    """
    Perform QKV self attention.

    Args:
        q (torch.Tensor): Query tensor with shape [b, s, a, d], where a is the number of heads.
        k (torch.Tensor): Key tensor with shape [b, s1, a, d]
        v (torch.Tensor): Value tensor with shape [b, s1, a, d]
        mode (str): Attention mode. Choose from 'self_flash', 'cross_flash', 'torch', and 'vanilla'.
        drop_rate (float): Dropout rate in attention map. (default: 0)
        attn_mask (torch.Tensor): Attention mask with shape [b, s1] (cross_attn), or [b, a, s, s1] (torch or vanilla).
            (default: None)
        causal (bool): Whether to use causal attention. (default: False)
        cu_seqlens_q (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
            used to index into q.
        cu_seqlens_kv (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
            used to index into kv.
        max_seqlen_q (int): The maximum sequence length in the batch of q.
        max_seqlen_kv (int): The maximum sequence length in the batch of k and v.

    Returns:
        torch.Tensor: Output tensor after self attention with shape [b, s, ad]
    """
    pre_attn_layout, post_attn_layout = MEMORY_LAYOUT[mode]
    q = pre_attn_layout(q)
    k = pre_attn_layout(k)
    v = pre_attn_layout(v)

    if mode == "torch":
        if attn_mask is not None and attn_mask.dtype != torch.bool:
            attn_mask = attn_mask.to(q.dtype)
        x = F.scaled_dot_product_attention(
            q, k, v, attn_mask=attn_mask, dropout_p=drop_rate, is_causal=causal
        )
    elif mode == "flash":
        x = flash_attn_varlen_func(
            q,
            k,
            v,
            cu_seqlens_q,
            cu_seqlens_kv,
            max_seqlen_q,
            max_seqlen_kv,
        )
        # x with shape [(bxs), a, d]
        x = x.view(
            batch_size, max_seqlen_q, x.shape[-2], x.shape[-1]
        )  # reshape x to [b, s, a, d]
    elif mode == "vanilla":
        scale_factor = 1 / math.sqrt(q.size(-1))

        b, a, s, _ = q.shape
        s1 = k.size(2)
        attn_bias = torch.zeros(b, a, s, s1, dtype=q.dtype, device=q.device)
        if causal:
            # Only applied to self attention
            assert (
                attn_mask is None
            ), "Causal mask and attn_mask cannot be used together"
            temp_mask = torch.ones(b, a, s, s, dtype=torch.bool, device=q.device).tril(
                diagonal=0
            )
            attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
            attn_bias.to(q.dtype)

        if attn_mask is not None:
            if attn_mask.dtype == torch.bool:
                attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
            else:
                attn_bias += attn_mask

        # TODO: Maybe force q and k to be float32 to avoid numerical overflow
        attn = (q @ k.transpose(-2, -1)) * scale_factor
        attn += attn_bias
        attn = attn.softmax(dim=-1)
        attn = torch.dropout(attn, p=drop_rate, train=True)
        x = attn @ v
    else:
        raise NotImplementedError(f"Unsupported attention mode: {mode}")

    x = post_attn_layout(x)
    b, s, a, d = x.shape
    out = x.reshape(b, s, -1)
    return out


def parallel_attention(
    hybrid_seq_parallel_attn,
    q,
    k,
    v,
    img_q_len,
    img_kv_len,
    cu_seqlens_q,
    cu_seqlens_kv
):
    attn1 = hybrid_seq_parallel_attn(
        None,
        q[:, :img_q_len, :, :],
        k[:, :img_kv_len, :, :],
        v[:, :img_kv_len, :, :],
        dropout_p=0.0,
        causal=False,
        joint_tensor_query=q[:,img_q_len:cu_seqlens_q[1]],
        joint_tensor_key=k[:,img_kv_len:cu_seqlens_kv[1]],
        joint_tensor_value=v[:,img_kv_len:cu_seqlens_kv[1]],
        joint_strategy="rear",
    )
    if flash_attn.__version__ >= '2.7.0':
        attn2, *_ = _flash_attn_forward(
            q[:,cu_seqlens_q[1]:],
            k[:,cu_seqlens_kv[1]:],
            v[:,cu_seqlens_kv[1]:],
            dropout_p=0.0,
            softmax_scale=q.shape[-1] ** (-0.5),
            causal=False,
            window_size_left=-1,
            window_size_right=-1,
            softcap=0.0,
            alibi_slopes=None,
            return_softmax=False,
        )
    else:
        attn2, *_ = _flash_attn_forward(
            q[:,cu_seqlens_q[1]:],
            k[:,cu_seqlens_kv[1]:],
            v[:,cu_seqlens_kv[1]:],
            dropout_p=0.0,
            softmax_scale=q.shape[-1] ** (-0.5),
            causal=False,
            window_size=(-1, -1),
            softcap=0.0,
            alibi_slopes=None,
            return_softmax=False,
        )
    attn = torch.cat([attn1, attn2], dim=1)
    b, s, a, d = attn.shape
    attn = attn.reshape(b, s, -1)

    return attn