File size: 6,527 Bytes
d90b3a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
/* coding=utf-8
 * Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include <torch/extension.h>

namespace fused_rope {

torch::Tensor fwd_cuda(const torch::Tensor& input,
                       const torch::Tensor& freqs,
                       const bool transpose_output);

torch::Tensor bwd_cuda(const torch::Tensor& output_grads,
                       const torch::Tensor& freqs,
                       const bool transpose_output);

torch::Tensor fwd_cached_cuda(const torch::Tensor& input,
                              const torch::Tensor& cos,
                              const torch::Tensor& sin,
                              const bool transpose_output);

torch::Tensor bwd_cached_cuda(const torch::Tensor& output_grads,
                              const torch::Tensor& cos,
                              const torch::Tensor& sin,
                              const bool transpose_output);

torch::Tensor fwd(const at::Tensor& input, const at::Tensor& freqs, const bool transpose_output)
{
    TORCH_CHECK(input.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(freqs.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(input.size(0) == freqs.size(0),
                "expected input and freqs tensor have the same sequence length");
    TORCH_CHECK(freqs.size(1) == 1 && freqs.size(2) == 1,
                "expected the second and third dims of the freqs tensor equal 1");
    TORCH_CHECK(input.size(3) >= freqs.size(3),
                "expected the last dim of the input tensor equals or is "
                "greater than the freqs tensor");
    TORCH_CHECK(freqs.scalar_type() == at::ScalarType::Float,
                "Dtype of the freqs tensor must be float");

    return fwd_cuda(input, freqs, transpose_output);
}

torch::Tensor bwd(const torch::Tensor& output_grads,
                  const at::Tensor& freqs,
                  const bool transpose_output)
{
    TORCH_CHECK(output_grads.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(freqs.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(output_grads.size(0) == freqs.size(0),
                "expected output_grads and freqs tensor have the same sequence length");
    TORCH_CHECK(freqs.size(1) == 1 && freqs.size(2) == 1,
                "expected the second and third dims of the freqs tensor equal 1");
    TORCH_CHECK(output_grads.size(3) >= freqs.size(3),
                "expected the last dim of the output_grads tensor equals or is "
                "greater than the freqs tensor");
    TORCH_CHECK(freqs.scalar_type() == at::ScalarType::Float,
                "Dtype of the freqs tensor must be float");

    return bwd_cuda(output_grads, freqs, transpose_output);
}

torch::Tensor fwd_cached(const at::Tensor& input,
                         const at::Tensor& cos,
                         const at::Tensor& sin,
                         const bool transpose_output)
{
    TORCH_CHECK(input.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(cos.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(sin.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(input.size(0) == cos.size(0),
                "expected input and cos tensor have the same sequence length");
    TORCH_CHECK(input.size(0) == sin.size(0),
                "expected input and sin tensor have the same sequence length");
    TORCH_CHECK(cos.size(1) == 1 && cos.size(2) == 1,
                "expected the second and third dims of the cos tensor equal 1");
    TORCH_CHECK(sin.size(1) == 1 && sin.size(2) == 1,
                "expected the second and third dims of the sin tensor equal 1");
    TORCH_CHECK(cos.size(3) == sin.size(3), "expected cos and sin tensor have the same last dim");
    TORCH_CHECK(input.size(3) >= cos.size(3),
                "expected the last dim of the input tensor equals or is "
                "greater than the cos tensor");
    TORCH_CHECK(cos.scalar_type() == sin.scalar_type(),
                "expected cos and sin tensor have the same dtype");

    return fwd_cached_cuda(input, cos, sin, transpose_output);
}

torch::Tensor bwd_cached(const torch::Tensor& output_grads,
                         const at::Tensor& cos,
                         const at::Tensor& sin,
                         const bool transpose_output)
{
    TORCH_CHECK(output_grads.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(cos.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(sin.dim() == 4, "expected 4D tensor");
    TORCH_CHECK(output_grads.size(0) == cos.size(0),
                "expected output_grads and cos tensor have the same sequence length");
    TORCH_CHECK(output_grads.size(0) == sin.size(0),
                "expected output_grads and sin tensor have the same sequence length");
    TORCH_CHECK(cos.size(1) == 1 && cos.size(2) == 1,
                "expected the second and third dims of the cos tensor equal 1");
    TORCH_CHECK(sin.size(1) == 1 && sin.size(2) == 1,
                "expected the second and third dims of the sin tensor equal 1");
    TORCH_CHECK(cos.size(3) == sin.size(3), "expected cos and sin tensor have the same last dim");
    TORCH_CHECK(output_grads.size(3) >= cos.size(3),
                "expected the last dim of the output_grads tensor equals or is "
                "greater than the cos tensor");
    TORCH_CHECK(cos.scalar_type() == sin.scalar_type(),
                "expected cos and sin tensor have the same dtype");

    return bwd_cached_cuda(output_grads, cos, sin, transpose_output);
}

}  // end namespace fused_rope

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
{
    m.def("forward", &fused_rope::fwd, "Fused Rotary Positional Embedding -- Forward.");
    m.def("backward", &fused_rope::bwd, "Fused Rotary Positional Embedding -- Backward.");
    m.def("forward_cached",
          &fused_rope::fwd_cached,
          "Fused Rotary Positional Embedding Cached -- Forward.");
    m.def("backward_cached",
          &fused_rope::bwd_cached,
          "Fused Rotary Positional Embedding Cached -- Backward.");
}