File size: 8,404 Bytes
16370b2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#include <torch/all.h>
#include <torch/python.h>
#include <c10/cuda/CUDAGuard.h>

// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_256.cpp
void vecquant8matmul_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros,
  torch::Tensor g_idx
);

void vecquant8matmul(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros,
  torch::Tensor g_idx
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_cuda(vec, mat, mul, scales, zeros, g_idx);
}

void vecquant8matmul_batched_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_cuda(vec, mat, mul, scales, zeros);
}

void vecquant8matmul_batched_column_compression_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_column_compression(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
}

void vecquant4matmul_batched_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant4matmul_batched(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant4matmul_batched_cuda(vec, mat, mul, scales, zeros);
}

void vecquant4matmul_batched_column_compression_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant4matmul_batched_column_compression(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant4matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
}

void vecquant8matmul_batched_old_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_old(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
}


void vecquant4matmul_batched_old_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant4matmul_batched_old(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant4matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
}

void vecquant8matmul_batched_column_compression_old_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_column_compression_old(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
}

void vecquant4matmul_batched_column_compression_old_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant4matmul_batched_column_compression_old(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant4matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
}



void vecquant8matmul_batched_faster_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_faster(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_faster_cuda(vec, mat, mul, scales, zeros);
}


void vecquant8matmul_batched_faster_old_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_faster_old(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_faster_old_cuda(vec, mat, mul, scales, zeros);
}

void vecquant8matmul_batched_column_compression_faster_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_column_compression_faster(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_column_compression_faster_cuda(vec, mat, mul, scales, zeros);
}


void vecquant8matmul_batched_column_compression_faster_old_cuda(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
);

void vecquant8matmul_batched_column_compression_faster_old(
  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
  torch::Tensor scales, torch::Tensor zeros
) {
  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
  vecquant8matmul_batched_column_compression_faster_old_cuda(vec, mat, mul, scales, zeros);
}



PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("vecquant8matmul", &vecquant8matmul, "Vector 8-bit Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant8matmul_batched", &vecquant8matmul_batched, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant8matmul_batched_old", &vecquant8matmul_batched_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant8matmul_batched_faster", &vecquant8matmul_batched_faster, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant8matmul_batched_faster_old", &vecquant8matmul_batched_faster_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant4matmul_batched_old", &vecquant4matmul_batched_old, "Vector 4-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant8matmul_batched_column_compression", &vecquant8matmul_batched_column_compression, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
  m.def("vecquant8matmul_batched_column_compression_old", &vecquant8matmul_batched_column_compression_old, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
  m.def("vecquant8matmul_batched_column_compression_faster", &vecquant8matmul_batched_column_compression_faster, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
  m.def("vecquant8matmul_batched_column_compression_faster_old", &vecquant8matmul_batched_column_compression_faster_old, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
  m.def("vecquant4matmul_batched_column_compression_old", &vecquant4matmul_batched_column_compression_old, "Vector old 4-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
  m.def("vecquant4matmul_batched", &vecquant4matmul_batched, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
  m.def("vecquant4matmul_batched_column_compression", &vecquant4matmul_batched_column_compression, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
}