File size: 8,581 Bytes
ad0ea6a
6cb2310
ad0ea6a
6cb2310
 
 
 
 
 
 
ad0ea6a
6cb2310
 
 
ad0ea6a
6cb2310
 
 
 
 
 
 
 
ad0ea6a
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
c56818b
6cb2310
 
ad0ea6a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6cb2310
 
 
 
 
ad0ea6a
 
 
 
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ad0ea6a
 
 
6cb2310
 
ad0ea6a
 
 
 
 
 
6cb2310
 
ad0ea6a
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
 
 
ad0ea6a
 
 
6cb2310
 
ad0ea6a
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c56818b
6cb2310
 
ad0ea6a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6cb2310
 
 
 
 
ad0ea6a
 
 
 
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
ad0ea6a
 
 
 
 
 
 
6cb2310
 
ad0ea6a
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
 
 
ad0ea6a
 
 
6cb2310
 
ad0ea6a
 
 
6cb2310
 
 
 
 
 
 
 
 
 
 
 
 
 
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
232
233
234
"""
Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments
"""

import logging
import math
from typing import Optional, Tuple

import torch
import transformers.models.llama.modeling_llama
from torch import nn

try:
    import xformers.ops
except ImportError:
    logging.error("xformers not found! Please install it before trying to use it.")


def hijack_llama_attention():
    transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward


def hijack_llama_sdp_attention():
    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
        sdp_attention_forward
    )


def xformers_forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    # pylint: disable=duplicate-code
    bsz, q_len, _ = hidden_states.size()

    query_states = (
        self.q_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    key_states = (
        self.k_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    value_states = (
        self.v_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )

    kv_seq_len = key_states.shape[-2]
    if past_key_value is not None:
        kv_seq_len += past_key_value[0].shape[-2]
    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
    (
        query_states,
        key_states,
    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
        query_states, key_states, cos, sin, position_ids
    )
    # [bsz, nh, t, hd]

    if past_key_value is not None:
        # reuse k, v, self_attention
        key_states = torch.cat([past_key_value[0], key_states], dim=2)
        value_states = torch.cat([past_key_value[1], value_states], dim=2)

    past_key_value = (key_states, value_states) if use_cache else None

    # We only apply xformers optimizations if we don't need to output the whole attention matrix
    if not output_attentions:
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
        # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
        if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
            # input and output should be of form (bsz, q_len, num_heads, head_dim)
            attn_output = xformers.ops.memory_efficient_attention(
                query_states, key_states, value_states, attn_bias=None
            )
        else:
            # input and output should be of form (bsz, q_len, num_heads, head_dim)
            attn_output = xformers.ops.memory_efficient_attention(
                query_states,
                key_states,
                value_states,
                attn_bias=xformers.ops.LowerTriangularMask(),
            )
        attn_weights = None
    else:
        attn_weights = torch.matmul(
            query_states, key_states.transpose(2, 3)
        ) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask
            attn_weights = torch.max(
                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
            )

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(
            attn_weights, dim=-1, dtype=torch.float32
        ).to(query_states.dtype)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2)

    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
    attn_output = self.o_proj(attn_output)
    return attn_output, attn_weights, past_key_value


def sdp_attention_forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    # pylint: disable=duplicate-code
    bsz, q_len, _ = hidden_states.size()

    query_states = (
        self.q_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    key_states = (
        self.k_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    value_states = (
        self.v_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )

    kv_seq_len = key_states.shape[-2]
    if past_key_value is not None:
        kv_seq_len += past_key_value[0].shape[-2]
    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
    (
        query_states,
        key_states,
    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
        query_states, key_states, cos, sin, position_ids
    )
    # [bsz, nh, t, hd]

    if past_key_value is not None:
        # reuse k, v, self_attention
        key_states = torch.cat([past_key_value[0], key_states], dim=2)
        value_states = torch.cat([past_key_value[1], value_states], dim=2)

    past_key_value = (key_states, value_states) if use_cache else None

    # We only apply sdp attention if we don't need to output the whole attention matrix
    if not output_attentions:
        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            is_causal=False,
        )
        attn_weights = None
    else:
        attn_weights = torch.matmul(
            query_states, key_states.transpose(2, 3)
        ) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask
            attn_weights = torch.max(
                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
            )

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(
            attn_weights, dim=-1, dtype=torch.float32
        ).to(query_states.dtype)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

    attn_output = attn_output.transpose(1, 2)
    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

    attn_output = self.o_proj(attn_output)

    return attn_output, attn_weights, past_key_value