Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
rope.h 10.5 KB
Newer Older
Xiaowei.zhang's avatar
Init the main branch for aiter  
Xiaowei.zhang committed
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 
// SPDX-License-Identifier: MIT
 
#pragma once

#include <torch/extension.h>

void rope_fwd_impl(
    torch::Tensor&       output,                    // [s, b, h, d]
    const torch::Tensor& input,                     // [s, b, h, d]
    const torch::Tensor& freqs,                     // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_bwd_impl(
    torch::Tensor&       input_grads,               // [s, b, h, d]
    const torch::Tensor& output_grads,              // [s, b, h, d]
    const torch::Tensor& freqs,                     // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_2c_fwd_impl(
    torch::Tensor&       output_x,                  // [s, b, h, d]
    torch::Tensor&       output_y,                  // [s, b, h, d]
    const torch::Tensor& input_x,                   // [s, b, h, d]
    const torch::Tensor& input_y,                   // [s, b, h, d]
    const torch::Tensor& freqs,                     // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_2c_bwd_impl(
    torch::Tensor&       input_grads_x,             // [s, b, h, d]
    torch::Tensor&       input_grads_y,             // [s, b, h, d]
    const torch::Tensor& output_grads_x,            // [s, b, h, d]
    const torch::Tensor& output_grads_y,            // [s, b, h, d]
    const torch::Tensor& freqs,                     // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_fwd_impl(
    torch::Tensor&       output,                    // [s, b, h, d]
    const torch::Tensor& input,                     // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_bwd_impl(
    torch::Tensor&       input_grads,               // [s, b, h, d]
    const torch::Tensor& output_grads,              // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_2c_fwd_impl(
    torch::Tensor&       output_x,                  // [s, b, h, d]
    torch::Tensor&       output_y,                  // [s, b, h, d]
    const torch::Tensor& input_x,                   // [s, b, h, d]
    const torch::Tensor& input_y,                   // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_2c_bwd_impl(
    torch::Tensor&       input_grads_x,             // [s, b, h, d]
    torch::Tensor&       input_grads_y,             // [s, b, h, d]
    const torch::Tensor& output_grads_x,            // [s, b, h, d]
    const torch::Tensor& output_grads_y,            // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_positions_fwd_impl(
    torch::Tensor&       output,                    // [s, b, h, d]
    const torch::Tensor& input,                     // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& positions,                 // [s, b]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_positions_2c_fwd_impl(
    torch::Tensor&       output_x,                  // [s, b, h, d]
    torch::Tensor&       output_y,                  // [s, b, h, d]
    const torch::Tensor& input_x,                   // [s, b, h, d]
    const torch::Tensor& input_y,                   // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& positions,                 // [s, b]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_positions_offsets_fwd_impl(
    torch::Tensor&       output,                    // [s, b, h, d]
    const torch::Tensor& input,                     // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& positions,                 // [s, b]
    const torch::Tensor& offsets,                   // [s, b]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_cached_positions_offsets_2c_fwd_impl(
    torch::Tensor&       output_x,                  // [s, b, h, d]
    torch::Tensor&       output_y,                  // [s, b, h, d]
    const torch::Tensor& input_x,                   // [s, b, h, d]
    const torch::Tensor& input_y,                   // [s, b, h, d]
    const torch::Tensor& cos,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& sin,                       // [s, 1, 1, d // 2] if reuse_freqs_front_part else [s, 1, 1, d]
    const torch::Tensor& positions,                 // [s, b]
    const torch::Tensor& offsets,                   // [s, b]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_thd_fwd_impl(
    torch::Tensor&       output,                    // [t, h, d]
    const torch::Tensor& input,                     // [t, h, d]
    const torch::Tensor& cu_seqlens,                // [b + 1]
    const torch::Tensor& freqs,                     // [max_s, 1, 1, d // 2] if reuse_freqs_front_part else [max_s, 1, 1, d], where max_s = cu_seqlens[-1]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_thd_bwd_impl(
    torch::Tensor&       input_grads,               // [t, h, d]
    const torch::Tensor& output_grads,              // [t, h, d]
    const torch::Tensor& cu_seqlens,                // [b + 1]
    const torch::Tensor& freqs,                     // [max_s, 1, 1, d // 2] if reuse_freqs_front_part else [max_s, 1, 1, d], where max_s = cu_seqlens[-1]
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_2d_fwd_impl(
    torch::Tensor&       output,                    // [b, s, h, d] where s = H * W
    const torch::Tensor& input,                     // [b, s, h, d] where s = H * W
    const torch::Tensor& cos_h,                     // [1, H', 1,  d // 4] if reuse_freqs_front_part else [1, H', 1,  d // 2], where H' >= H
    const torch::Tensor& sin_h,                     // [1, H', 1,  d // 4] if reuse_freqs_front_part else [1, H', 1,  d // 2], where H' >= H
    const torch::Tensor& cos_w,                     // [1, 1,  W', d // 4] if reuse_freqs_front_part else [1, 1,  W', d // 2], where W' >= W
    const torch::Tensor& sin_w,                     // [1, 1,  W', d // 4] if reuse_freqs_front_part else [1, 1,  W', d // 2], where W' >= W
    const int32_t        img_height,                // H
    const int32_t        img_width,                 // W
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);

void rope_2d_bwd_impl(
    torch::Tensor&       input_grads,               // [b, s, h, d] where s = H * W
    const torch::Tensor& output_grads,              // [b, s, h, d] where s = H * W
    const torch::Tensor& cos_h,                     // [1, H', 1,  d // 4] if reuse_freqs_front_part else [1, H', 1,  d // 2], where H' >= H
    const torch::Tensor& sin_h,                     // [1, H', 1,  d // 4] if reuse_freqs_front_part else [1, H', 1,  d // 2], where H' >= H
    const torch::Tensor& cos_w,                     // [1, 1,  W', d // 4] if reuse_freqs_front_part else [1, 1,  W', d // 2], where W' >= W
    const torch::Tensor& sin_w,                     // [1, 1,  W', d // 4] if reuse_freqs_front_part else [1, 1,  W', d // 2], where W' >= W
    const int32_t        img_height,                // H
    const int32_t        img_width,                 // W
    const int32_t        rotate_style,              // 0: NEOX style, 1: GPT-J style
    const bool           reuse_freqs_front_part,
    const bool           nope_first
);