yangapku commited on
Commit
319ed0f
1 Parent(s): 687862a

update kvcache

Browse files
configuration_qwen.py CHANGED
@@ -35,6 +35,8 @@ class QWenConfig(PretrainedConfig):
35
  intermediate_size=22016,
36
  no_bias=True,
37
  tie_word_embeddings=False,
 
 
38
  **kwargs,
39
  ):
40
  self.vocab_size = vocab_size
@@ -59,6 +61,8 @@ class QWenConfig(PretrainedConfig):
59
  self.use_logn_attn = use_logn_attn
60
  self.use_flash_attn = use_flash_attn
61
  self.no_bias = no_bias
 
 
62
  super().__init__(
63
  tie_word_embeddings=tie_word_embeddings,
64
  **kwargs
 
35
  intermediate_size=22016,
36
  no_bias=True,
37
  tie_word_embeddings=False,
38
+ use_cache_quantization=False,
39
+ use_cache_kernel=False,
40
  **kwargs,
41
  ):
42
  self.vocab_size = vocab_size
 
61
  self.use_logn_attn = use_logn_attn
62
  self.use_flash_attn = use_flash_attn
63
  self.no_bias = no_bias
64
+ self.use_cache_quantization=use_cache_quantization
65
+ self.use_cache_kernel=use_cache_kernel
66
  super().__init__(
67
  tie_word_embeddings=tie_word_embeddings,
68
  **kwargs
kernels/cache_autogptq_cuda_256.cpp ADDED
@@ -0,0 +1,198 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #include <torch/all.h>
2
+ #include <torch/python.h>
3
+ #include <c10/cuda/CUDAGuard.h>
4
+
5
+ // adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_256.cpp
6
+ void vecquant8matmul_cuda(
7
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
8
+ torch::Tensor scales, torch::Tensor zeros,
9
+ torch::Tensor g_idx
10
+ );
11
+
12
+ void vecquant8matmul(
13
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
14
+ torch::Tensor scales, torch::Tensor zeros,
15
+ torch::Tensor g_idx
16
+ ) {
17
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
18
+ vecquant8matmul_cuda(vec, mat, mul, scales, zeros, g_idx);
19
+ }
20
+
21
+ void vecquant8matmul_batched_cuda(
22
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
23
+ torch::Tensor scales, torch::Tensor zeros
24
+ );
25
+
26
+ void vecquant8matmul_batched(
27
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
28
+ torch::Tensor scales, torch::Tensor zeros
29
+ ) {
30
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
31
+ vecquant8matmul_batched_cuda(vec, mat, mul, scales, zeros);
32
+ }
33
+
34
+ void vecquant8matmul_batched_column_compression_cuda(
35
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
36
+ torch::Tensor scales, torch::Tensor zeros
37
+ );
38
+
39
+ void vecquant8matmul_batched_column_compression(
40
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
41
+ torch::Tensor scales, torch::Tensor zeros
42
+ ) {
43
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
44
+ vecquant8matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
45
+ }
46
+
47
+ void vecquant4matmul_batched_cuda(
48
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
49
+ torch::Tensor scales, torch::Tensor zeros
50
+ );
51
+
52
+ void vecquant4matmul_batched(
53
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
54
+ torch::Tensor scales, torch::Tensor zeros
55
+ ) {
56
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
57
+ vecquant4matmul_batched_cuda(vec, mat, mul, scales, zeros);
58
+ }
59
+
60
+ void vecquant4matmul_batched_column_compression_cuda(
61
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
62
+ torch::Tensor scales, torch::Tensor zeros
63
+ );
64
+
65
+ void vecquant4matmul_batched_column_compression(
66
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
67
+ torch::Tensor scales, torch::Tensor zeros
68
+ ) {
69
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
70
+ vecquant4matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
71
+ }
72
+
73
+ void vecquant8matmul_batched_old_cuda(
74
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
75
+ torch::Tensor scales, torch::Tensor zeros
76
+ );
77
+
78
+ void vecquant8matmul_batched_old(
79
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
80
+ torch::Tensor scales, torch::Tensor zeros
81
+ ) {
82
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
83
+ vecquant8matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
84
+ }
85
+
86
+
87
+ void vecquant4matmul_batched_old_cuda(
88
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
89
+ torch::Tensor scales, torch::Tensor zeros
90
+ );
91
+
92
+ void vecquant4matmul_batched_old(
93
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
94
+ torch::Tensor scales, torch::Tensor zeros
95
+ ) {
96
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
97
+ vecquant4matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
98
+ }
99
+
100
+ void vecquant8matmul_batched_column_compression_old_cuda(
101
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
102
+ torch::Tensor scales, torch::Tensor zeros
103
+ );
104
+
105
+ void vecquant8matmul_batched_column_compression_old(
106
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
107
+ torch::Tensor scales, torch::Tensor zeros
108
+ ) {
109
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
110
+ vecquant8matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
111
+ }
112
+
113
+ void vecquant4matmul_batched_column_compression_old_cuda(
114
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
115
+ torch::Tensor scales, torch::Tensor zeros
116
+ );
117
+
118
+ void vecquant4matmul_batched_column_compression_old(
119
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
120
+ torch::Tensor scales, torch::Tensor zeros
121
+ ) {
122
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
123
+ vecquant4matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
124
+ }
125
+
126
+
127
+
128
+ void vecquant8matmul_batched_faster_cuda(
129
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
130
+ torch::Tensor scales, torch::Tensor zeros
131
+ );
132
+
133
+ void vecquant8matmul_batched_faster(
134
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
135
+ torch::Tensor scales, torch::Tensor zeros
136
+ ) {
137
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
138
+ vecquant8matmul_batched_faster_cuda(vec, mat, mul, scales, zeros);
139
+ }
140
+
141
+
142
+ void vecquant8matmul_batched_faster_old_cuda(
143
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
144
+ torch::Tensor scales, torch::Tensor zeros
145
+ );
146
+
147
+ void vecquant8matmul_batched_faster_old(
148
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
149
+ torch::Tensor scales, torch::Tensor zeros
150
+ ) {
151
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
152
+ vecquant8matmul_batched_faster_old_cuda(vec, mat, mul, scales, zeros);
153
+ }
154
+
155
+ void vecquant8matmul_batched_column_compression_faster_cuda(
156
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
157
+ torch::Tensor scales, torch::Tensor zeros
158
+ );
159
+
160
+ void vecquant8matmul_batched_column_compression_faster(
161
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
162
+ torch::Tensor scales, torch::Tensor zeros
163
+ ) {
164
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
165
+ vecquant8matmul_batched_column_compression_faster_cuda(vec, mat, mul, scales, zeros);
166
+ }
167
+
168
+
169
+ void vecquant8matmul_batched_column_compression_faster_old_cuda(
170
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
171
+ torch::Tensor scales, torch::Tensor zeros
172
+ );
173
+
174
+ void vecquant8matmul_batched_column_compression_faster_old(
175
+ torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
176
+ torch::Tensor scales, torch::Tensor zeros
177
+ ) {
178
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
179
+ vecquant8matmul_batched_column_compression_faster_old_cuda(vec, mat, mul, scales, zeros);
180
+ }
181
+
182
+
183
+
184
+ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
185
+ m.def("vecquant8matmul", &vecquant8matmul, "Vector 8-bit Quantized Matrix Multiplication (CUDA) (desc_act)");
186
+ m.def("vecquant8matmul_batched", &vecquant8matmul_batched, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
187
+ m.def("vecquant8matmul_batched_old", &vecquant8matmul_batched_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
188
+ m.def("vecquant8matmul_batched_faster", &vecquant8matmul_batched_faster, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
189
+ m.def("vecquant8matmul_batched_faster_old", &vecquant8matmul_batched_faster_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
190
+ m.def("vecquant4matmul_batched_old", &vecquant4matmul_batched_old, "Vector 4-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
191
+ 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)");
192
+ 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)");
193
+ 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)");
194
+ 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)");
195
+ 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)");
196
+ m.def("vecquant4matmul_batched", &vecquant4matmul_batched, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
197
+ 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)");
198
+ }
kernels/cache_autogptq_cuda_kernel_256.cu ADDED
@@ -0,0 +1,1708 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #define _CRT_SECURE_NO_WARNINGS
2
+ #include <torch/all.h>
3
+ #include <torch/python.h>
4
+ #include <cuda.h>
5
+ #include <cuda_runtime.h>
6
+ #include <cuda_fp16.h>
7
+ #include <stdint.h>
8
+
9
+ #if (defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 700) || defined(USE_ROCM)
10
+ // adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_kernel_256.cu
11
+ __device__ __forceinline__ void atomicAdd(c10::Half* address, c10::Half val) {
12
+ unsigned int *address_as_ui = reinterpret_cast<unsigned int *>(reinterpret_cast<char *>(address) - (reinterpret_cast<size_t>(address) & 2));
13
+ unsigned int old = *address_as_ui;
14
+ unsigned int assumed;
15
+
16
+ do {
17
+ assumed = old;
18
+ unsigned short hsum = reinterpret_cast<size_t>(address) & 2 ? (old >> 16) : (old & 0xffff);
19
+ hsum += val;
20
+ old = reinterpret_cast<size_t>(address) & 2
21
+ ? (old & 0xffff) | (hsum << 16)
22
+ : (old & 0xffff0000) | hsum;
23
+ old = atomicCAS(address_as_ui, assumed, old);
24
+
25
+ // Note: uses integer comparison to avoid hang in case of NaN (since NaN != NaN)
26
+ } while (assumed != old);
27
+ }
28
+ __device__ __forceinline__ void atomicAdd(__half* address, c10::Half val) {
29
+ unsigned int * address_as_ui = (unsigned int *) ((char *)address - ((size_t)address & 2));
30
+ unsigned int old = *address_as_ui;
31
+ unsigned int assumed;
32
+
33
+ do {
34
+ assumed = old;
35
+ __half_raw hsum;
36
+ hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
37
+ half tmpres = __hadd(hsum, val);
38
+ hsum = __half_raw(tmpres);
39
+ old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16) : (old & 0xffff0000) | hsum.x;
40
+ old = atomicCAS(address_as_ui, assumed, old);
41
+ } while (assumed != old);
42
+ }
43
+ #endif
44
+
45
+ template <typename scalar_t>
46
+ __global__ void VecQuant8MatMulKernel(
47
+ const scalar_t* __restrict__ vec,
48
+ const int* __restrict__ mat,
49
+ scalar_t* __restrict__ mul,
50
+ const scalar_t* __restrict__ scales,
51
+ const int* __restrict__ zeros,
52
+ const int* __restrict__ g_idx,
53
+ int batch,
54
+ int vec_height,
55
+ int height,
56
+ int width,
57
+ int zero_width
58
+ );
59
+
60
+ template <typename scalar_t>
61
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel(
62
+ const scalar_t* __restrict__ vec,
63
+ const int* __restrict__ mat,
64
+ scalar_t* __restrict__ mul,
65
+ const scalar_t* __restrict__ scales,
66
+ const int* __restrict__ zeros,
67
+ int batch,
68
+ int heads,
69
+ int vec_row,
70
+ int height,
71
+ int width
72
+ );
73
+
74
+ template <typename scalar_t>
75
+ __global__ void VecQuant4BatchMatMulColumnCompressionKernel(
76
+ const scalar_t* __restrict__ vec,
77
+ const int* __restrict__ mat,
78
+ scalar_t* __restrict__ mul,
79
+ const scalar_t* __restrict__ scales,
80
+ const int* __restrict__ zeros,
81
+ int batch,
82
+ int heads,
83
+ int vec_row,
84
+ int height,
85
+ int width
86
+ );
87
+
88
+ template <typename scalar_t>
89
+ __global__ void VecQuant8BatchMatMulKernel(
90
+ const scalar_t* __restrict__ vec,
91
+ const int* __restrict__ mat,
92
+ scalar_t* __restrict__ mul,
93
+ const scalar_t* __restrict__ scales,
94
+ const int* __restrict__ zeros,
95
+ int batch,
96
+ int heads,
97
+ int vec_row,
98
+ int vec_height,
99
+ int height,
100
+ int width,
101
+ int zero_width
102
+ );
103
+
104
+ template <typename scalar_t>
105
+ __global__ void VecQuant4BatchMatMulKernel(
106
+ const scalar_t* __restrict__ vec,
107
+ const int* __restrict__ mat,
108
+ scalar_t* __restrict__ mul,
109
+ const scalar_t* __restrict__ scales,
110
+ const int* __restrict__ zeros,
111
+ int batch,
112
+ int heads,
113
+ int vec_row,
114
+ int vec_height,
115
+ int height,
116
+ int width,
117
+ int zero_width
118
+ );
119
+
120
+
121
+
122
+ template <typename scalar_t>
123
+ __global__ void VecQuant8BatchMatMulKernel_old(
124
+ const scalar_t* __restrict__ vec,
125
+ const uint8_t* __restrict__ mat,
126
+ scalar_t* __restrict__ mul,
127
+ const scalar_t* __restrict__ scales,
128
+ const scalar_t* __restrict__ zeros,
129
+ int batch,
130
+ int heads,
131
+ int vec_row,
132
+ int vec_height,
133
+ int height,
134
+ int width,
135
+ int zero_width
136
+ );
137
+
138
+ __global__ void VecQuant8BatchMatMulKernel_faster(
139
+ const half* __restrict__ vec,
140
+ const uint8_t* __restrict__ mat,
141
+ half* __restrict__ mul,
142
+ const half* __restrict__ scales,
143
+ const half* __restrict__ zeros,
144
+ int batch,
145
+ int heads,
146
+ int vec_row,
147
+ int vec_height,
148
+ int height,
149
+ int width,
150
+ int zero_width
151
+ );
152
+
153
+
154
+
155
+ __global__ void VecQuant8BatchMatMulKernel_faster_old(
156
+ const half* __restrict__ vec,
157
+ const uint8_t* __restrict__ mat,
158
+ half* __restrict__ mul,
159
+ const half* __restrict__ scales,
160
+ const half* __restrict__ zeros,
161
+ int batch,
162
+ int heads,
163
+ int vec_row,
164
+ int vec_height,
165
+ int height,
166
+ int width
167
+ );
168
+
169
+
170
+ template <typename scalar_t>
171
+ __global__ void VecQuant4BatchMatMulKernel_old(
172
+ const scalar_t* __restrict__ vec,
173
+ const uint8_t* __restrict__ mat,
174
+ scalar_t* __restrict__ mul,
175
+ const scalar_t* __restrict__ scales,
176
+ const scalar_t* __restrict__ zeros,
177
+ int batch,
178
+ int heads,
179
+ int vec_row,
180
+ int vec_height,
181
+ int height,
182
+ int width,
183
+ int zero_width
184
+ );
185
+
186
+
187
+ template <typename scalar_t>
188
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
189
+ const scalar_t* __restrict__ vec,
190
+ const uint8_t* __restrict__ mat,
191
+ scalar_t* __restrict__ mul,
192
+ const scalar_t* __restrict__ scales,
193
+ const scalar_t* __restrict__ zeros,
194
+ int batch,
195
+ int heads,
196
+ int vec_row,
197
+ int height,
198
+ int width
199
+ );
200
+
201
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
202
+ const half* __restrict__ vec,
203
+ const uint8_t* __restrict__ mat,
204
+ half* __restrict__ mul,
205
+ const half* __restrict__ scales,
206
+ const half* __restrict__ zeros,
207
+ int batch,
208
+ int heads,
209
+ int vec_row,
210
+ int height,
211
+ int width
212
+ );
213
+
214
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
215
+ const half* __restrict__ vec,
216
+ const uint8_t* __restrict__ mat,
217
+ half* __restrict__ mul,
218
+ const half* __restrict__ scales,
219
+ const half* __restrict__ zeros,
220
+ int batch,
221
+ int heads,
222
+ int vec_row,
223
+ int height,
224
+ int width
225
+ );
226
+
227
+
228
+ template <typename scalar_t>
229
+ __global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
230
+ const scalar_t* __restrict__ vec,
231
+ const uint8_t* __restrict__ mat,
232
+ scalar_t* __restrict__ mul,
233
+ const scalar_t* __restrict__ scales,
234
+ const scalar_t* __restrict__ zeros,
235
+ int batch,
236
+ int heads,
237
+ int vec_row,
238
+ int height,
239
+ int width
240
+ );
241
+
242
+
243
+ __global__ void VecQuant8BatchMatMulKernel_faster(
244
+ const half* __restrict__ vec,
245
+ const uint8_t* __restrict__ mat,
246
+ half* __restrict__ mul,
247
+ const half* __restrict__ scales,
248
+ const half* __restrict__ zeros,
249
+ int batch,
250
+ int heads,
251
+ int vec_row,
252
+ int vec_height,
253
+ int height,
254
+ int width
255
+ );
256
+
257
+
258
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
259
+ const half* __restrict__ vec,
260
+ const uint8_t* __restrict__ mat,
261
+ half* __restrict__ mul,
262
+ const half* __restrict__ scales,
263
+ const half* __restrict__ zeros,
264
+ int batch,
265
+ int heads,
266
+ int vec_row,
267
+ int height,
268
+ int width
269
+ );
270
+
271
+ const int BLOCKWIDTH = 128;
272
+ const int BLOCKHEIGHT8 = 32;
273
+ const int BLOCKHEIGHT4 = 16;
274
+ const int BLOCKHEIGHT_OLD4 = 128;
275
+ //const int BLOCKHEIGHT_OLD8 = 128;
276
+
277
+ __device__ inline unsigned int as_unsigned(int i) {
278
+ return *reinterpret_cast<unsigned int*>(&i);
279
+ }
280
+
281
+ __device__ inline int as_int(int i) {
282
+ return *reinterpret_cast<int*>(&i);
283
+ }
284
+
285
+ void vecquant8matmul_batched_column_compression_cuda(
286
+ torch::Tensor vec,
287
+ torch::Tensor mat,
288
+ torch::Tensor mul,
289
+ torch::Tensor scales,
290
+ torch::Tensor zeros
291
+ ) {
292
+ int batch = vec.size(0);
293
+ int heads = vec.size(1);
294
+ int vec_row = vec.size(2);
295
+ int height = vec.size(3);
296
+ int width = mat.size(3) * 4;
297
+
298
+ dim3 blocks(
299
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
300
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
301
+ );
302
+ dim3 threads(BLOCKWIDTH);
303
+
304
+ AT_DISPATCH_FLOATING_TYPES(
305
+ vec.type(), "vecquant8matmul_batched_cuda", ([&] {
306
+ VecQuant8BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
307
+ vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
308
+ scales.data<scalar_t>(), zeros.data<int>(),
309
+ batch, heads, vec_row, height, width
310
+ );
311
+ })
312
+ );
313
+
314
+ }
315
+
316
+ template <typename scalar_t>
317
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel(
318
+ const scalar_t* __restrict__ vec,
319
+ const int* __restrict__ mat,
320
+ scalar_t* __restrict__ mul,
321
+ const scalar_t* __restrict__ scales,
322
+ const int* __restrict__ zeros,
323
+ int batch,
324
+ int heads,
325
+ int vec_row,
326
+ int height,
327
+ int width
328
+ ) {
329
+ int weight_total = batch * heads * height * width / 4;
330
+ int input_total = batch * heads * vec_row * height;
331
+ int out_total = batch * heads * vec_row * width;
332
+ int tid = threadIdx.x;
333
+ // h is index of height with step being BLOCKWIDTH
334
+ int h = BLOCKWIDTH * blockIdx.x;
335
+ // w is index of width with step being 1
336
+ int w = BLOCKWIDTH * blockIdx.y + tid;
337
+ if (w >= width && tid >= height) {
338
+ return;
339
+ }
340
+
341
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
342
+ int k;
343
+ scalar_t w_tmp;
344
+
345
+ float weight[BLOCKWIDTH];
346
+
347
+ for (int b = 0; b < batch; ++b){
348
+ for (int head = 0; head < heads; ++head){
349
+ int batch_shift = b * heads + head;
350
+ for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
351
+ int i_w = (w / 4);
352
+ int w_bit = (w % 4) * 8;
353
+
354
+ int w_index = (batch_shift * height + h + k) * width / 4 + i_w;
355
+ if (w_index >= weight_total || w >= width) {
356
+ weight[k] = 0;
357
+ } else {
358
+ scalar_t scale = scales[batch_shift * height + h + k];
359
+ scalar_t zero = zeros[batch_shift * height + h + k];
360
+ w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xFF);
361
+ weight[k] = scale * (w_tmp - zero);
362
+ }
363
+ }
364
+
365
+ scalar_t res;
366
+ for (int vr = 0; vr < vec_row; ++vr){
367
+ res = 0;
368
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
369
+ if (vec_index < input_total) {
370
+ blockvec[tid] = vec[vec_index];
371
+ } else {
372
+ blockvec[tid] = 0;
373
+ }
374
+
375
+ __syncthreads();
376
+ for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
377
+ // res is the dot product of BLOCKWIDTH elements (part of width)
378
+ res += weight[k] * blockvec[k];
379
+ }
380
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
381
+ int out_index = (batch_shift * vec_row + vr) * width + w;
382
+ if (out_index < out_total) {
383
+ atomicAdd(&mul[out_index], res);
384
+ }
385
+ __syncthreads();
386
+ }
387
+ }
388
+ }
389
+ }
390
+
391
+ void vecquant8matmul_batched_cuda(
392
+ torch::Tensor vec,
393
+ torch::Tensor mat,
394
+ torch::Tensor mul,
395
+ torch::Tensor scales,
396
+ torch::Tensor zeros
397
+ ) {
398
+ int batch = vec.size(0);
399
+ int heads = vec.size(1);
400
+ int vec_row = vec.size(2);
401
+ int vec_height = vec.size(3);
402
+ int height = mat.size(2);
403
+ int width = mat.size(3);
404
+ int zero_width = zeros.size(2);
405
+
406
+ dim3 blocks(
407
+ (height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
408
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
409
+ );
410
+ dim3 threads(BLOCKWIDTH);
411
+
412
+ AT_DISPATCH_FLOATING_TYPES(
413
+ vec.type(), "vecquant8matmul_batched_cuda", ([&] {
414
+ VecQuant8BatchMatMulKernel<<<blocks, threads>>>(
415
+ vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
416
+ scales.data<scalar_t>(), zeros.data<int>(),
417
+ batch, heads, vec_row, vec_height, height, width, zero_width
418
+ );
419
+ })
420
+ );
421
+
422
+ }
423
+
424
+ template <typename scalar_t>
425
+ __global__ void VecQuant8BatchMatMulKernel(
426
+ const scalar_t* __restrict__ vec,
427
+ const int* __restrict__ mat,
428
+ scalar_t* __restrict__ mul,
429
+ const scalar_t* __restrict__ scales,
430
+ const int* __restrict__ zeros,
431
+ int batch,
432
+ int heads,
433
+ int vec_row,
434
+ int vec_height,
435
+ int height,
436
+ int width,
437
+ int zero_width
438
+ ) {
439
+ int weight_total = batch * heads * height * width;
440
+ int input_total = batch * heads * vec_row * vec_height;
441
+ int out_total = batch * heads * vec_row * width;
442
+ int tid = threadIdx.x;
443
+ // h is index of height with step being BLOCKHEIGHT8
444
+ int h = BLOCKHEIGHT8 * blockIdx.x;
445
+ // w is index of width with step being 1
446
+ int w = BLOCKWIDTH * blockIdx.y + tid;
447
+ if (w >= width && tid >= vec_height) {
448
+ return;
449
+ }
450
+
451
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
452
+ // i is index of mat of block first row
453
+ int i = width * h + w;
454
+ // if (i >= width * height) {
455
+ // return;
456
+ // }
457
+ int k;
458
+ scalar_t w_tmp;
459
+
460
+ int z_w = w / 4;
461
+ int z_mod = (w % 4) * 8;
462
+
463
+ float weight[BLOCKWIDTH];
464
+
465
+ for (int b = 0; b < batch; ++b){
466
+ for (int head = 0; head < heads; ++head){
467
+ int batch_shift = b * heads + head;
468
+ for (k = 0; k < BLOCKWIDTH && h * 4 + k < vec_height; ++k){
469
+ int k_w = (k / 4);
470
+ int k_bit = (k % 4) * 8;
471
+
472
+ int w_index = batch_shift * height * width + i + (k_w * width);
473
+ if (w_index >= weight_total || w >= width) {
474
+ weight[k] = 0;
475
+ } else {
476
+ scalar_t scale = scales[batch_shift * width + w];
477
+ scalar_t zero;
478
+ if (zero_width == width) {
479
+ zero = zeros[batch_shift * width + w];
480
+ } else {
481
+ zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
482
+ }
483
+ w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xFF);
484
+ weight[k] = scale * (w_tmp - zero);
485
+ }
486
+ }
487
+
488
+ scalar_t res;
489
+ for (int vr = 0; vr < vec_row; ++vr){
490
+ res = 0;
491
+ int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
492
+ if (vec_index < input_total) {
493
+ blockvec[tid] = vec[vec_index];
494
+ } else {
495
+ blockvec[tid] = 0;
496
+ }
497
+
498
+ __syncthreads();
499
+ for (k = 0; k < BLOCKWIDTH && h * 4 + k < vec_height; ++k){
500
+ // res is the dot product of BLOCKWIDTH elements (part of width)
501
+ res += weight[k] * blockvec[k];
502
+ }
503
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
504
+ int out_index = (batch_shift * vec_row + vr) * width + w;
505
+ if (out_index < out_total) {
506
+ atomicAdd(&mul[out_index], res);
507
+ }
508
+ __syncthreads();
509
+ }
510
+ }
511
+ }
512
+ }
513
+
514
+
515
+ void vecquant8matmul_cuda(
516
+ torch::Tensor vec,
517
+ torch::Tensor mat,
518
+ torch::Tensor mul,
519
+ torch::Tensor scales,
520
+ torch::Tensor zeros,
521
+ torch::Tensor g_idx
522
+ ) {
523
+ int batch = vec.size(0);
524
+ int vec_height = vec.size(1);
525
+ int height = mat.size(0);
526
+ int width = mat.size(1);
527
+ int zero_width = zeros.size(1);
528
+
529
+ dim3 blocks(
530
+ (height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
531
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
532
+ );
533
+ dim3 threads(BLOCKWIDTH);
534
+
535
+ AT_DISPATCH_FLOATING_TYPES(
536
+ vec.type(), "vecquant8matmul_cuda", ([&] {
537
+ VecQuant8MatMulKernel<<<blocks, threads>>>(
538
+ vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
539
+ scales.data<scalar_t>(), zeros.data<int>(), g_idx.data<int>(),
540
+ batch, vec_height, height, width, zero_width
541
+ );
542
+ })
543
+ );
544
+ }
545
+
546
+ template <typename scalar_t>
547
+ __global__ void VecQuant8MatMulKernel(
548
+ const scalar_t* __restrict__ vec,
549
+ const int* __restrict__ mat,
550
+ scalar_t* __restrict__ mul,
551
+ const scalar_t* __restrict__ scales,
552
+ const int* __restrict__ zeros,
553
+ const int* __restrict__ g_idx,
554
+ int batch,
555
+ int vec_height,
556
+ int height,
557
+ int width,
558
+ int zero_width
559
+ ) {
560
+ int h = BLOCKHEIGHT8 * blockIdx.x;
561
+ int w = BLOCKWIDTH * blockIdx.y + threadIdx.x;
562
+
563
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
564
+ int i = width * h + w;
565
+ int g_h = h * 4;
566
+ int k;
567
+ unsigned int g;
568
+ scalar_t w_tmp;
569
+
570
+ int z_w = w / 4;
571
+ int z_mod = (w % 4) * 8;
572
+
573
+ float weight[BLOCKWIDTH];
574
+
575
+ for (k = 0; k < BLOCKWIDTH; ++k){
576
+ int k_w = (k / 4);
577
+ int k_bit = (k % 4) * 8;
578
+
579
+ g = as_int(g_idx[g_h + k]);
580
+ scalar_t scale = scales[g * width + w];
581
+ scalar_t zero = scalar_t(((as_unsigned(zeros[g * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
582
+
583
+ w_tmp = ((as_unsigned(mat[i + (k_w * width)]) >> k_bit) & 0xFF);
584
+
585
+ weight[k] = scale * (w_tmp - zero);
586
+ }
587
+
588
+
589
+ scalar_t res;
590
+ for (int b = 0; b < batch; ++b){
591
+ res = 0;
592
+ blockvec[threadIdx.x] = vec[b * vec_height + blockIdx.x * BLOCKWIDTH + threadIdx.x];
593
+ __syncthreads();
594
+ for (k = 0; k < BLOCKWIDTH; ++k){
595
+ res += weight[k] * blockvec[k];
596
+ }
597
+ atomicAdd(&mul[b * width + w], res);
598
+ __syncthreads();
599
+ }
600
+ }
601
+
602
+
603
+
604
+ void vecquant4matmul_batched_cuda(
605
+ torch::Tensor vec,
606
+ torch::Tensor mat,
607
+ torch::Tensor mul,
608
+ torch::Tensor scales,
609
+ torch::Tensor zeros
610
+ ) {
611
+ int batch = vec.size(0);
612
+ int heads = vec.size(1);
613
+ int vec_row = vec.size(2);
614
+ int vec_height = vec.size(3);
615
+ int height = mat.size(2);
616
+ int width = mat.size(3);
617
+ int zero_width = zeros.size(2);
618
+
619
+ dim3 blocks(
620
+ (height + BLOCKHEIGHT4 - 1) / BLOCKHEIGHT4,
621
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
622
+ );
623
+ dim3 threads(BLOCKWIDTH);
624
+
625
+ AT_DISPATCH_FLOATING_TYPES(
626
+ vec.type(), "vecquant4matmul_batched_cuda", ([&] {
627
+ VecQuant4BatchMatMulKernel<<<blocks, threads>>>(
628
+ vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
629
+ scales.data<scalar_t>(), zeros.data<int>(),
630
+ batch, heads, vec_row, vec_height, height, width, zero_width
631
+ );
632
+ })
633
+ );
634
+
635
+ }
636
+
637
+ template <typename scalar_t>
638
+ __global__ void VecQuant4BatchMatMulKernel(
639
+ const scalar_t* __restrict__ vec,
640
+ const int* __restrict__ mat,
641
+ scalar_t* __restrict__ mul,
642
+ const scalar_t* __restrict__ scales,
643
+ const int* __restrict__ zeros,
644
+ int batch,
645
+ int heads,
646
+ int vec_row,
647
+ int vec_height,
648
+ int height,
649
+ int width,
650
+ int zero_width
651
+ ) {
652
+ int weight_total = batch * heads * height * width;
653
+ int input_total = batch * heads * vec_row * vec_height;
654
+ int out_total = batch * heads * vec_row * width;
655
+ int tid = threadIdx.x;
656
+ // h is index of height with step being BLOCKHEIGHT4
657
+ int h = BLOCKHEIGHT4 * blockIdx.x;
658
+ // w is index of width with step being 1
659
+ int w = BLOCKWIDTH * blockIdx.y + tid;
660
+ if (w >= width && tid >= vec_height) {
661
+ return;
662
+ }
663
+
664
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
665
+ // i is index of mat of block first row
666
+ int i = width * h + w;
667
+ int k;
668
+ scalar_t w_tmp;
669
+
670
+ int z_w = w / 8;
671
+ int z_mod = (w % 8) * 4;
672
+
673
+ float weight[BLOCKWIDTH];
674
+
675
+ for (int b = 0; b < batch; ++b){
676
+ for (int head = 0; head < heads; ++head){
677
+ int batch_shift = b * heads + head;
678
+ for (k = 0; k < BLOCKWIDTH && h * 8 + k < vec_height; ++k){
679
+ int k_w = (k / 8);
680
+ int k_bit = (k % 8) * 4;
681
+
682
+ int w_index = batch_shift * height * width + i + (k_w * width);
683
+ if (w_index >= weight_total || w >= width) {
684
+ weight[k] = 0;
685
+ } else {
686
+ scalar_t scale = scales[batch_shift * width + w];
687
+ scalar_t zero;
688
+ if (zero_width == width) {
689
+ zero = zeros[batch_shift * width + w];
690
+ } else {
691
+ zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xF));
692
+ }
693
+ w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
694
+ weight[k] = scale * (w_tmp - zero);
695
+ }
696
+ }
697
+
698
+ scalar_t res;
699
+ for (int vr = 0; vr < vec_row; ++vr){
700
+ res = 0;
701
+ int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
702
+ if (vec_index < input_total) {
703
+ blockvec[tid] = vec[vec_index];
704
+ } else {
705
+ blockvec[tid] = 0;
706
+ }
707
+
708
+ __syncthreads();
709
+ for (k = 0; k < BLOCKWIDTH && h * 8 + k < vec_height; ++k){
710
+ // res is the dot product of BLOCKWIDTH elements (part of width)
711
+ res += weight[k] * blockvec[k];
712
+ }
713
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
714
+ int out_index = (batch_shift * vec_row + vr) * width + w;
715
+ if (out_index < out_total) {
716
+ atomicAdd(&mul[out_index], res);
717
+ }
718
+ __syncthreads();
719
+ }
720
+ }
721
+ }
722
+ }
723
+
724
+
725
+
726
+ void vecquant4matmul_batched_column_compression_cuda(
727
+ torch::Tensor vec,
728
+ torch::Tensor mat,
729
+ torch::Tensor mul,
730
+ torch::Tensor scales,
731
+ torch::Tensor zeros
732
+ ) {
733
+ int batch = vec.size(0);
734
+ int heads = vec.size(1);
735
+ int vec_row = vec.size(2);
736
+ int height = vec.size(3);
737
+ int width = mat.size(3) * 8;
738
+
739
+ dim3 blocks(
740
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
741
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
742
+ );
743
+ dim3 threads(BLOCKWIDTH);
744
+
745
+ AT_DISPATCH_FLOATING_TYPES(
746
+ vec.type(), "vecquant4matmul_batched_cuda", ([&] {
747
+ VecQuant4BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
748
+ vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
749
+ scales.data<scalar_t>(), zeros.data<int>(),
750
+ batch, heads, vec_row, height, width
751
+ );
752
+ })
753
+ );
754
+
755
+ }
756
+
757
+ template <typename scalar_t>
758
+ __global__ void VecQuant4BatchMatMulColumnCompressionKernel(
759
+ const scalar_t* __restrict__ vec,
760
+ const int* __restrict__ mat,
761
+ scalar_t* __restrict__ mul,
762
+ const scalar_t* __restrict__ scales,
763
+ const int* __restrict__ zeros,
764
+ int batch,
765
+ int heads,
766
+ int vec_row,
767
+ int height,
768
+ int width
769
+ ) {
770
+ int weight_total = batch * heads * height * width / 8;
771
+ int input_total = batch * heads * vec_row * height;
772
+ int out_total = batch * heads * vec_row * width;
773
+ int tid = threadIdx.x;
774
+ // h is index of height with step being BLOCKWIDTH
775
+ int h = BLOCKWIDTH * blockIdx.x;
776
+ // w is index of width with step being 1
777
+ int w = BLOCKWIDTH * blockIdx.y + tid;
778
+ if (w >= width && tid >= height) {
779
+ return;
780
+ }
781
+
782
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
783
+ int k;
784
+ scalar_t w_tmp;
785
+
786
+ float weight[BLOCKWIDTH];
787
+
788
+ for (int b = 0; b < batch; ++b){
789
+ for (int head = 0; head < heads; ++head){
790
+ int batch_shift = b * heads + head;
791
+ for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
792
+ int i_w = (w / 8);
793
+ int w_bit = (w % 8) * 4;
794
+
795
+ int w_index = (batch_shift * height + h + k) * width / 8 + i_w;
796
+ if (w_index >= weight_total || w >= width) {
797
+ weight[k] = 0;
798
+ } else {
799
+ scalar_t scale = scales[batch_shift * height + h + k];
800
+ scalar_t zero = zeros[batch_shift * height + h + k];
801
+ w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xF);
802
+ weight[k] = scale * (w_tmp - zero);
803
+ }
804
+ }
805
+
806
+ scalar_t res;
807
+ for (int vr = 0; vr < vec_row; ++vr){
808
+ res = 0;
809
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
810
+ if (vec_index < input_total) {
811
+ blockvec[tid] = vec[vec_index];
812
+ } else {
813
+ blockvec[tid] = 0;
814
+ }
815
+
816
+ __syncthreads();
817
+ for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
818
+ // res is the dot product of BLOCKWIDTH elements (part of width)
819
+ res += weight[k] * blockvec[k];
820
+ }
821
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
822
+ int out_index = (batch_shift * vec_row + vr) * width + w;
823
+ if (out_index < out_total) {
824
+ atomicAdd(&mul[out_index], res);
825
+ }
826
+ __syncthreads();
827
+ }
828
+ }
829
+ }
830
+ }
831
+
832
+
833
+ void vecquant8matmul_batched_old_cuda(
834
+ torch::Tensor vec,
835
+ torch::Tensor mat,
836
+ torch::Tensor mul,
837
+ torch::Tensor scales,
838
+ torch::Tensor zeros
839
+ ) {
840
+ int batch = vec.size(0);
841
+ int heads = vec.size(1);
842
+ int vec_row = vec.size(2);
843
+ int vec_height = vec.size(3);
844
+ int height = mat.size(2);
845
+ int width = mat.size(3);
846
+ int zero_width = zeros.size(2);
847
+
848
+ dim3 blocks(
849
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
850
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
851
+ );
852
+ dim3 threads(BLOCKWIDTH);
853
+
854
+ AT_DISPATCH_FLOATING_TYPES(
855
+ vec.type(), "vecquant8matmul_batched_old_cuda", ([&] {
856
+ VecQuant8BatchMatMulKernel_old<<<blocks, threads>>>(
857
+ vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
858
+ scales.data<scalar_t>(), zeros.data<scalar_t>(),
859
+ batch, heads, vec_row, vec_height, height, width, zero_width
860
+ );
861
+ })
862
+ );
863
+ }
864
+
865
+
866
+ template <typename scalar_t>
867
+ __global__ void VecQuant8BatchMatMulKernel_old(
868
+ const scalar_t* __restrict__ vec,
869
+ const uint8_t* __restrict__ mat,
870
+ scalar_t* __restrict__ mul,
871
+ const scalar_t* __restrict__ scales,
872
+ const scalar_t* __restrict__ zeros,
873
+ int batch,
874
+ int heads,
875
+ int vec_row,
876
+ int vec_height,
877
+ int height,
878
+ int width,
879
+ int zero_width
880
+ ) {
881
+ int weight_total = batch * heads * height * width;
882
+ int input_total = batch * heads * vec_row * vec_height;
883
+ int out_total = batch * heads * vec_row * width;
884
+ int tid = threadIdx.x;
885
+ // h is index of height with step being BLOCKHEIGHT8
886
+ int h = BLOCKWIDTH * blockIdx.x;
887
+ // w is index of width with step being 1
888
+ int w = BLOCKWIDTH * blockIdx.y + tid;
889
+ if (w >= width && tid >= vec_height) {
890
+ return;
891
+ }
892
+
893
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
894
+ // i is index of mat of block first row
895
+ int i = width * h + w;
896
+ int k;
897
+ scalar_t w_tmp;
898
+
899
+ float weight[BLOCKWIDTH];
900
+ for (int b = 0; b < batch; ++b){
901
+ for (int head = 0; head < heads; ++head){
902
+ int batch_shift = b * heads + head;
903
+ for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
904
+ int k_w = k;
905
+ int w_index = batch_shift * height * width + i + (k_w * width);
906
+ if (w_index >= weight_total || w >= width) {
907
+ weight[k] = 0;
908
+ } else {
909
+ scalar_t scale = scales[batch_shift * width + w];
910
+ scalar_t zero = zeros[batch_shift * width + w];
911
+ w_tmp = as_unsigned(mat[w_index]);
912
+ weight[k] = scale * (w_tmp - zero);
913
+ }
914
+ }
915
+
916
+ scalar_t res;
917
+ for (int vr = 0; vr < vec_row; ++vr){
918
+ res = 0;
919
+ int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
920
+ if (vec_index < input_total) {
921
+ blockvec[tid] = vec[vec_index];
922
+ } else {
923
+ blockvec[tid] = 0;
924
+ }
925
+
926
+ __syncthreads();
927
+ for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
928
+ // res is the dot product of BLOCKWIDTH elements (part of width)
929
+ res += weight[k] * blockvec[k];
930
+ }
931
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
932
+ int out_index = (batch_shift * vec_row + vr) * width + w;
933
+ if (out_index < out_total) {
934
+ atomicAdd(&mul[out_index], res);
935
+ }
936
+ __syncthreads();
937
+ }
938
+ }
939
+ }
940
+ }
941
+
942
+
943
+
944
+ void vecquant8matmul_batched_faster_cuda(
945
+ torch::Tensor vec,
946
+ torch::Tensor mat,
947
+ torch::Tensor mul,
948
+ torch::Tensor scales,
949
+ torch::Tensor zeros
950
+ ) {
951
+ int batch = vec.size(0);
952
+ int heads = vec.size(1);
953
+ int vec_row = vec.size(2);
954
+ int vec_height = vec.size(3);
955
+ int height = mat.size(2);
956
+ int width = mat.size(3);
957
+ int zero_width = zeros.size(2);
958
+
959
+ dim3 blocks(
960
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
961
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
962
+ );
963
+ dim3 threads(BLOCKWIDTH);
964
+
965
+ VecQuant8BatchMatMulKernel_faster<<<blocks, threads>>>(
966
+ (half*) vec.data_ptr(),
967
+ (uint8_t*) mat.data_ptr(),
968
+ (half*) mul.data_ptr(),
969
+ (half*) scales.data_ptr(),
970
+ (half*) zeros.data_ptr(),
971
+ batch, heads, vec_row, vec_height, height, width, zero_width
972
+ );
973
+ }
974
+
975
+
976
+
977
+ __global__ void VecQuant8BatchMatMulKernel_faster(
978
+ const half* __restrict__ vec,
979
+ const uint8_t* __restrict__ mat,
980
+ half* __restrict__ mul,
981
+ const half* __restrict__ scales,
982
+ const half* __restrict__ zeros,
983
+ int batch,
984
+ int heads,
985
+ int vec_row,
986
+ int vec_height,
987
+ int height,
988
+ int width,
989
+ int zero_width
990
+ ) {
991
+ //int weight_total = batch * heads * height * width;
992
+ int input_total = batch * heads * vec_row * vec_height;
993
+ int out_total = batch * heads * vec_row * width;
994
+ int tid = threadIdx.x;
995
+ int h = BLOCKWIDTH * blockIdx.x;
996
+ int w = BLOCKWIDTH * blockIdx.y + tid;
997
+ if (w >= width && tid >= height) {
998
+ return;
999
+ }
1000
+
1001
+ __shared__ float blockvec[BLOCKWIDTH];
1002
+ int i = width * h + w;
1003
+ int k;
1004
+ float w_tmp;
1005
+
1006
+ float weight[BLOCKWIDTH];
1007
+ for (int b = 0; b < batch; ++b){
1008
+ for (int head = 0; head < heads; ++head){
1009
+ int batch_shift = b * heads + head;
1010
+ for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
1011
+ int k_w = k;
1012
+ int w_index = batch_shift * height * width + i + (k_w * width);
1013
+ float scale = __half2float(scales[batch_shift * width + w]);
1014
+ float zero = __half2float(zeros[batch_shift * width + w]);
1015
+ w_tmp = as_unsigned(mat[w_index]);
1016
+ weight[k] = scale *(w_tmp-zero);
1017
+ }
1018
+
1019
+ float res;
1020
+ for (int vr = 0; vr < vec_row; ++vr){
1021
+ res = 0;
1022
+ int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
1023
+ if (vec_index < input_total) {
1024
+ blockvec[tid] = __half2float(vec[vec_index]);
1025
+ } else {
1026
+ blockvec[tid] = 0;
1027
+ }
1028
+ __syncthreads();
1029
+ for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
1030
+ float temp_res = weight[k]*blockvec[k];
1031
+ res += temp_res;
1032
+ }
1033
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1034
+ if (out_index < out_total) {
1035
+ atomicAdd(&mul[out_index], __float2half(res));
1036
+ }
1037
+ __syncthreads();
1038
+ }
1039
+ }
1040
+ }
1041
+ }
1042
+
1043
+
1044
+
1045
+
1046
+ void vecquant8matmul_batched_column_compression_faster_cuda(
1047
+ torch::Tensor vec,
1048
+ torch::Tensor mat,
1049
+ torch::Tensor mul,
1050
+ torch::Tensor scales,
1051
+ torch::Tensor zeros
1052
+ ) {
1053
+ int batch = vec.size(0);
1054
+ int heads = vec.size(1);
1055
+ int vec_row = vec.size(2);
1056
+ int height = vec.size(3);
1057
+ int width = mat.size(3);
1058
+
1059
+ dim3 blocks(
1060
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
1061
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
1062
+ );
1063
+ dim3 threads(BLOCKWIDTH);
1064
+
1065
+ VecQuant8BatchMatMulColumnCompressionKernel_faster<<<blocks, threads>>>(
1066
+ (half*) vec.data_ptr(),
1067
+ (uint8_t*) mat.data_ptr(),
1068
+ (half*) mul.data_ptr(),
1069
+ (half*) scales.data_ptr(),
1070
+ (half*) zeros.data_ptr(),
1071
+ batch, heads, vec_row, height, width
1072
+ );
1073
+
1074
+ }
1075
+
1076
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
1077
+ const half* __restrict__ vec,
1078
+ const uint8_t* __restrict__ mat,
1079
+ half* __restrict__ mul,
1080
+ const half* __restrict__ scales,
1081
+ const half* __restrict__ zeros,
1082
+ int batch,
1083
+ int heads,
1084
+ int vec_row,
1085
+ int height,
1086
+ int width
1087
+ ) {
1088
+ //int weight_total = batch * heads * height * width;
1089
+ int input_total = batch * heads * vec_row * height;
1090
+ int out_total = batch * heads * vec_row * width;
1091
+ int tid = threadIdx.x;
1092
+ int h = BLOCKWIDTH * blockIdx.x;
1093
+ int w = BLOCKWIDTH * blockIdx.y + tid;
1094
+ if (w >= width && tid >= height) {
1095
+ return;
1096
+ }
1097
+
1098
+ __shared__ float blockvec[BLOCKWIDTH];
1099
+ int k;
1100
+ float w_tmp;
1101
+ float weight[BLOCKWIDTH];
1102
+
1103
+ for (int b = 0; b < batch; ++b){
1104
+ for (int head = 0; head < heads; ++head){
1105
+ int batch_shift = b * heads + head;
1106
+ for (k = 0; k < BLOCKWIDTH; ++k){
1107
+ int w_index = (batch_shift * height + h + k) * width + w;
1108
+ float scale = __half2float(scales[batch_shift * height + h + k]);
1109
+ float zero = __half2float(zeros[batch_shift * height + h + k]);
1110
+ w_tmp = mat[w_index];
1111
+ weight[k] = scale * (w_tmp-zero);
1112
+ }
1113
+
1114
+ float res;
1115
+ for (int vr = 0; vr < vec_row; ++vr){
1116
+ res = 0;
1117
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
1118
+ if (vec_index < input_total) {
1119
+ blockvec[tid] = __half2float(vec[vec_index]);
1120
+ } else {
1121
+ blockvec[tid] = 0;
1122
+ }
1123
+ __syncthreads();
1124
+ for (k = 0; k < BLOCKWIDTH; ++k){
1125
+ res += weight[k]*blockvec[k];
1126
+ }
1127
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1128
+ if (out_index < out_total) {
1129
+ atomicAdd(&mul[out_index], __float2half(res));
1130
+ }
1131
+ __syncthreads();
1132
+ }
1133
+ }
1134
+ }
1135
+ }
1136
+
1137
+
1138
+
1139
+ void vecquant8matmul_batched_column_compression_old_cuda(
1140
+ torch::Tensor vec,
1141
+ torch::Tensor mat,
1142
+ torch::Tensor mul,
1143
+ torch::Tensor scales,
1144
+ torch::Tensor zeros
1145
+ ) {
1146
+ int batch = vec.size(0);
1147
+ int heads = vec.size(1);
1148
+ int vec_row = vec.size(2);
1149
+ int height = vec.size(3);
1150
+ int width = mat.size(3);
1151
+
1152
+ dim3 blocks(
1153
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
1154
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
1155
+ );
1156
+ dim3 threads(BLOCKWIDTH);
1157
+
1158
+ AT_DISPATCH_FLOATING_TYPES(
1159
+ vec.type(), "vecquant8matmul_batched_column_compression_old_cuda", ([&] {
1160
+ VecQuant8BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
1161
+ vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
1162
+ scales.data<scalar_t>(), zeros.data<scalar_t>(),
1163
+ batch, heads, vec_row, height, width
1164
+ );
1165
+ })
1166
+ );
1167
+
1168
+ }
1169
+
1170
+ template <typename scalar_t>
1171
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
1172
+ const scalar_t* __restrict__ vec,
1173
+ const uint8_t* __restrict__ mat,
1174
+ scalar_t* __restrict__ mul,
1175
+ const scalar_t* __restrict__ scales,
1176
+ const scalar_t* __restrict__ zeros,
1177
+ int batch,
1178
+ int heads,
1179
+ int vec_row,
1180
+ int height,
1181
+ int width
1182
+ ) {
1183
+ int weight_total = batch * heads * height * width;
1184
+ int input_total = batch * heads * vec_row * height;
1185
+ int out_total = batch * heads * vec_row * width;
1186
+ int tid = threadIdx.x;
1187
+ // h is index of height with step being BLOCKWIDTH
1188
+ int h = BLOCKWIDTH * blockIdx.x;
1189
+ // w is index of width with step being 1
1190
+ int w = BLOCKWIDTH * blockIdx.y + tid;
1191
+ if (w >= width && tid >= height) {
1192
+ return;
1193
+ }
1194
+
1195
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
1196
+ int k;
1197
+ scalar_t w_tmp;
1198
+
1199
+ float weight[BLOCKWIDTH];
1200
+
1201
+ for (int b = 0; b < batch; ++b){
1202
+ for (int head = 0; head < heads; ++head){
1203
+ int batch_shift = b * heads + head;
1204
+ for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
1205
+ int w_index = (batch_shift * height + h + k) * width + w;
1206
+ if (w_index >= weight_total || w >= width) {
1207
+ weight[k] = 0;
1208
+ } else {
1209
+ scalar_t scale = scales[batch_shift * height + h + k];
1210
+ scalar_t zero = zeros[batch_shift * height + h + k];
1211
+ w_tmp = mat[w_index];
1212
+ weight[k] = scale * (w_tmp - zero);
1213
+ }
1214
+ }
1215
+
1216
+ scalar_t res;
1217
+ for (int vr = 0; vr < vec_row; ++vr){
1218
+ res = 0;
1219
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
1220
+ if (vec_index < input_total) {
1221
+ blockvec[tid] = vec[vec_index];
1222
+ } else {
1223
+ blockvec[tid] = 0;
1224
+ }
1225
+
1226
+ __syncthreads();
1227
+ for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
1228
+ // res is the dot product of BLOCKWIDTH elements (part of width)
1229
+ res += weight[k] * blockvec[k];
1230
+ }
1231
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
1232
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1233
+ if (out_index < out_total) {
1234
+ atomicAdd(&mul[out_index], res);
1235
+ }
1236
+ __syncthreads();
1237
+ }
1238
+ }
1239
+ }
1240
+ }
1241
+
1242
+
1243
+ void vecquant4matmul_batched_old_cuda(
1244
+ torch::Tensor vec,
1245
+ torch::Tensor mat,
1246
+ torch::Tensor mul,
1247
+ torch::Tensor scales,
1248
+ torch::Tensor zeros
1249
+ ) {
1250
+ int batch = vec.size(0);
1251
+ int heads = vec.size(1);
1252
+ int vec_row = vec.size(2);
1253
+ int vec_height = vec.size(3);
1254
+ int height = mat.size(2);
1255
+ int width = mat.size(3);
1256
+ int zero_width = zeros.size(2);
1257
+
1258
+ dim3 blocks(
1259
+ (height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
1260
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
1261
+ );
1262
+ dim3 threads(BLOCKWIDTH);
1263
+
1264
+ AT_DISPATCH_FLOATING_TYPES(
1265
+ vec.type(), "vecquant4matmul_batched_old_cuda", ([&] {
1266
+ VecQuant4BatchMatMulKernel_old<<<blocks, threads>>>(
1267
+ vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
1268
+ scales.data<scalar_t>(), zeros.data<scalar_t>(),
1269
+ batch, heads, vec_row, vec_height, height, width, zero_width
1270
+ );
1271
+ })
1272
+ );
1273
+
1274
+ }
1275
+
1276
+ template <typename scalar_t>
1277
+ __global__ void VecQuant4BatchMatMulKernel_old(
1278
+ const scalar_t* __restrict__ vec,
1279
+ const uint8_t* __restrict__ mat,
1280
+ scalar_t* __restrict__ mul,
1281
+ const scalar_t* __restrict__ scales,
1282
+ const scalar_t* __restrict__ zeros,
1283
+ int batch,
1284
+ int heads,
1285
+ int vec_row,
1286
+ int vec_height,
1287
+ int height,
1288
+ int width,
1289
+ int zero_width
1290
+ ) {
1291
+ int weight_total = batch * heads * height * width;
1292
+ int input_total = batch * heads * vec_row * vec_height;
1293
+ int out_total = batch * heads * vec_row * width;
1294
+ int tid = threadIdx.x;
1295
+ // h is index of height with step being BLOCKHEIGHT_OLD4
1296
+ int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
1297
+ // w is index of width with step being 1
1298
+ int w = BLOCKWIDTH * blockIdx.y + tid;
1299
+ if (w >= width && tid >= vec_height) {
1300
+ return;
1301
+ }
1302
+
1303
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
1304
+ // i is index of mat of block first row
1305
+ int i = width * h + w;
1306
+ int k;
1307
+ scalar_t w_tmp;
1308
+
1309
+ float weight[BLOCKWIDTH];
1310
+ for (int b = 0; b < batch; ++b){
1311
+ for (int head = 0; head < heads; ++head){
1312
+ int batch_shift = b * heads + head;
1313
+ for (k = 0; k < BLOCKWIDTH && h*2 + k < vec_height; ++k){
1314
+ int k_w = (k / 2);
1315
+ int k_bit = (k % 2) * 4;
1316
+ int w_index = batch_shift * height * width + i + (k_w * width);
1317
+ if (w_index >= weight_total || w >= width) {
1318
+ weight[k] = 0;
1319
+ } else {
1320
+ scalar_t scale = scales[batch_shift * width + w];
1321
+ scalar_t zero = zeros[batch_shift * width + w];
1322
+ w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
1323
+ weight[k] = scale * (w_tmp - zero);
1324
+ }
1325
+ }
1326
+
1327
+ scalar_t res;
1328
+ for (int vr = 0; vr < vec_row; ++vr){
1329
+ res = 0;
1330
+ int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
1331
+ if (vec_index < input_total) {
1332
+ blockvec[tid] = vec[vec_index];
1333
+ } else {
1334
+ blockvec[tid] = 0;
1335
+ }
1336
+
1337
+ __syncthreads();
1338
+ for (k = 0; k < BLOCKWIDTH && h*2 + k < vec_height; ++k){
1339
+ // res is the dot product of BLOCKWIDTH elements (part of width)
1340
+ res += weight[k] * blockvec[k];
1341
+ }
1342
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
1343
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1344
+ if (out_index < out_total) {
1345
+ atomicAdd(&mul[out_index], res);
1346
+ }
1347
+ __syncthreads();
1348
+ }
1349
+ }
1350
+ }
1351
+ }
1352
+
1353
+
1354
+
1355
+
1356
+
1357
+ void vecquant4matmul_batched_column_compression_old_cuda(
1358
+ torch::Tensor vec,
1359
+ torch::Tensor mat,
1360
+ torch::Tensor mul,
1361
+ torch::Tensor scales,
1362
+ torch::Tensor zeros
1363
+ ) {
1364
+ int batch = vec.size(0);
1365
+ int heads = vec.size(1);
1366
+ int vec_row = vec.size(2);
1367
+ int height = vec.size(3);
1368
+ int width = mat.size(3);
1369
+
1370
+ dim3 blocks(
1371
+ (height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
1372
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
1373
+ );
1374
+ dim3 threads(BLOCKWIDTH);
1375
+
1376
+ AT_DISPATCH_FLOATING_TYPES(
1377
+ vec.type(), "vecquant4matmul_batched_column_compression_old_cuda", ([&] {
1378
+ VecQuant4BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
1379
+ vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
1380
+ scales.data<scalar_t>(), zeros.data<scalar_t>(),
1381
+ batch, heads, vec_row, height, width
1382
+ );
1383
+ })
1384
+ );
1385
+
1386
+ }
1387
+
1388
+ template <typename scalar_t>
1389
+ __global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
1390
+ const scalar_t* __restrict__ vec,
1391
+ const uint8_t* __restrict__ mat,
1392
+ scalar_t* __restrict__ mul,
1393
+ const scalar_t* __restrict__ scales,
1394
+ const scalar_t* __restrict__ zeros,
1395
+ int batch,
1396
+ int heads,
1397
+ int vec_row,
1398
+ int height,
1399
+ int width
1400
+ ) {
1401
+ int weight_total = batch * heads * height * width;
1402
+ int input_total = batch * heads * vec_row * height;
1403
+ int out_total = batch * heads * vec_row * width;
1404
+ int tid = threadIdx.x;
1405
+ // h is index of height with step being BLOCKWIDTH
1406
+ int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
1407
+ // w is index of width with step being 1
1408
+ int w = BLOCKWIDTH * blockIdx.y + tid;
1409
+ if (w >= width && tid >= height) {
1410
+ return;
1411
+ }
1412
+
1413
+ __shared__ scalar_t blockvec[BLOCKWIDTH];
1414
+ int k;
1415
+ scalar_t w_tmp;
1416
+
1417
+ float weight[BLOCKWIDTH];
1418
+
1419
+ for (int b = 0; b < batch; ++b){
1420
+ for (int head = 0; head < heads; ++head){
1421
+ int batch_shift = b * heads + head;
1422
+ for (k = 0; k < BLOCKWIDTH && h*2 + k < height; ++k){
1423
+ int k_w = (k / 2);
1424
+ int k_bit = (k % 2) * 4;
1425
+ int w_index = (batch_shift * height + h + k) * width + k_w;
1426
+ if (w_index >= weight_total || w >= width) {
1427
+ weight[k] = 0;
1428
+ } else {
1429
+ scalar_t scale = scales[batch_shift * height + h + k];
1430
+ scalar_t zero = zeros[batch_shift * height + h + k];
1431
+ w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
1432
+ weight[k] = scale * (w_tmp - zero);
1433
+ }
1434
+ }
1435
+
1436
+ scalar_t res;
1437
+ for (int vr = 0; vr < vec_row; ++vr){
1438
+ res = 0;
1439
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
1440
+ if (vec_index < input_total) {
1441
+ blockvec[tid] = vec[vec_index];
1442
+ } else {
1443
+ blockvec[tid] = 0;
1444
+ }
1445
+
1446
+ __syncthreads();
1447
+ for (k = 0; k < BLOCKWIDTH && h*2 + k < height; ++k){
1448
+ // res is the dot product of BLOCKWIDTH elements (part of width)
1449
+ res += weight[k] * blockvec[k];
1450
+ }
1451
+ // add res to the final result, final matrix shape: (batch, vec_row, width)
1452
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1453
+ if (out_index < out_total) {
1454
+ atomicAdd(&mul[out_index], res);
1455
+ }
1456
+ __syncthreads();
1457
+ }
1458
+ }
1459
+ }
1460
+ }
1461
+
1462
+
1463
+
1464
+
1465
+
1466
+ void vecquant8matmul_batched_faster_old_cuda(
1467
+ torch::Tensor vec,
1468
+ torch::Tensor mat,
1469
+ torch::Tensor mul,
1470
+ torch::Tensor scales,
1471
+ torch::Tensor zeros
1472
+ ) {
1473
+ int batch = vec.size(0);
1474
+ int heads = vec.size(1);
1475
+ int vec_row = vec.size(2);
1476
+ int vec_height = vec.size(3);
1477
+ int height = mat.size(2);
1478
+ int width = mat.size(3);
1479
+
1480
+ dim3 blocks(
1481
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
1482
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
1483
+ );
1484
+ dim3 threads(BLOCKWIDTH);
1485
+
1486
+ VecQuant8BatchMatMulKernel_faster_old<<<blocks, threads>>>(
1487
+ (half*) vec.data_ptr(),
1488
+ (uint8_t*) mat.data_ptr(),
1489
+ (half*) mul.data_ptr(),
1490
+ (half*) scales.data_ptr(),
1491
+ (half*) zeros.data_ptr(),
1492
+ batch, heads, vec_row, vec_height, height, width
1493
+ );
1494
+ }
1495
+
1496
+
1497
+ __global__ void VecQuant8BatchMatMulKernel_faster_old(
1498
+ const half* __restrict__ vec,
1499
+ const uint8_t* __restrict__ mat,
1500
+ half* __restrict__ mul,
1501
+ const half* __restrict__ scales,
1502
+ const half* __restrict__ zeros,
1503
+ int batch,
1504
+ int heads,
1505
+ int vec_row,
1506
+ int vec_height,
1507
+ int height,
1508
+ int width
1509
+ ) {
1510
+ int weight_total = batch * heads * height * width;
1511
+ int input_total = batch * heads * vec_row * vec_height;
1512
+ int out_total = batch * heads * vec_row * width;
1513
+ int tid = threadIdx.x;
1514
+ const int BLOCKWIDTH_half = BLOCKWIDTH/2;
1515
+
1516
+ int h = BLOCKWIDTH * blockIdx.x; //head_dim, dim=-1
1517
+ int w = BLOCKWIDTH * blockIdx.y + tid; //seq-len, +0-256 ,dim=-2
1518
+ /*
1519
+ if (w >= width && tid >= vec_height) {
1520
+ return;
1521
+ }
1522
+ */
1523
+ __shared__ half blockvec[BLOCKWIDTH]; //256
1524
+ int i = width * h + w;
1525
+ int k;
1526
+
1527
+ half w_tmp1 = __float2half(0);
1528
+ half w_tmp2 = __float2half(0);
1529
+
1530
+ half2 weight[BLOCKWIDTH_half];
1531
+ for (int b = 0; b < batch; ++b){
1532
+ for (int head = 0; head < heads; ++head){
1533
+ int batch_shift = b * heads + head;
1534
+ //int zero_index = batch_shift;
1535
+ for (k = 0; k < BLOCKWIDTH_half; ++k){
1536
+ int w_index1 = batch_shift * height * width + i + (2 * k * width); // [batch,head,h+k, w]
1537
+ int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
1538
+ int zero_index = batch_shift * width + w; // [batch,head, w]
1539
+ if (w_index1 >= weight_total || w >= width || (2 * k + h) >= height) {
1540
+ weight[k] = __float2half2_rn(0);
1541
+ } else {
1542
+ float zero_f=__half2float(zeros[zero_index]);
1543
+ float scale_f= __half2float(scales[zero_index]);
1544
+ if (w_index2 >= weight_total){
1545
+ w_tmp1 = __float2half((as_unsigned(mat[w_index1]) -zero_f)*scale_f);
1546
+ w_tmp2 = __float2half(0);
1547
+ weight[k] = __halves2half2(w_tmp1,w_tmp2);
1548
+ //printf("zero_index is %d w is %d height is %d width is %d w_index1 is %d w_tmp1 is %f w_tmp2 is %f zero is %f scale is %f low is %f high is %f \n ",zero_index,w,height, width,w_index1,__half2float(w_tmp1),__half2float(w_tmp2),zero_f,scale_f,__low2float(weight[k]),__high2float(weight[k]));
1549
+ }else{
1550
+ w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
1551
+ w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
1552
+
1553
+ //weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero,zero)),__halves2half2(scale,scale));
1554
+ weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
1555
+ //printf("zero_index1 is %d zero_index2 is %d k is %d head is %d w is %d h is %d height is %d width is %d w_index1 is %d w_index2 is %d zero is %f scale is %f low is %f high is %f \n ",zero_index1,zero_index2,k,head,w,h,height, width,w_index1,w_index2,__half2float(zero1),__half2float(scale1),__low2float(weight[k]),__high2float(weight[k]));
1556
+ }
1557
+ }
1558
+ }
1559
+
1560
+
1561
+ for (int vr = 0; vr < vec_row; ++vr){
1562
+ float res=0;
1563
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
1564
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1565
+ if (vec_index < input_total) {
1566
+ //blockvec[tid] = __half2float(vec[vec_index]);// [batch, head, vr, tid(seq_len dim+)]
1567
+ blockvec[tid] = vec[vec_index];
1568
+ //printf("width is %d height is %d h is %d w is %d vec_index is %d out_index is %d vec_row is %d vec_height is %d,vr is %d tid is %d blockvec is %f\n",width,height, h,w,vec_index,out_index,vec_row,vec_height,vr,tid,blockvec[tid]);
1569
+ } else {
1570
+ blockvec[tid] = __float2half(0);
1571
+ }
1572
+ __syncthreads();
1573
+ if (out_index < out_total) {
1574
+ for (k = 0; k < BLOCKWIDTH_half; ++k){
1575
+ half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
1576
+ res += __low2float(res2) + __high2float(res2);
1577
+ }
1578
+ atomicAdd(&mul[out_index], __float2half(res));
1579
+ }
1580
+ __syncthreads();
1581
+ }
1582
+ }
1583
+ }
1584
+ }
1585
+
1586
+
1587
+ void vecquant8matmul_batched_column_compression_faster_old_cuda(
1588
+ torch::Tensor vec, // [batch,heads, seq_q, seq_v]
1589
+ torch::Tensor mat, // [batch,heads, seq_v, head_dim]
1590
+ torch::Tensor mul, // [batch,heads, seq_q,head_dim]
1591
+ torch::Tensor scales, // [batch,heads, head_dim]
1592
+ torch::Tensor zeros
1593
+ ) {
1594
+ int batch = vec.size(0);
1595
+ int heads = vec.size(1);
1596
+ int vec_row = vec.size(2); //ql
1597
+ int height = mat.size(2); //vl
1598
+ int width = mat.size(3); //head_dim
1599
+
1600
+ dim3 blocks(
1601
+ (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
1602
+ (width + BLOCKWIDTH - 1) / BLOCKWIDTH
1603
+ );
1604
+ dim3 threads(BLOCKWIDTH);
1605
+
1606
+ VecQuant8BatchMatMulColumnCompressionKernel_faster_old<<<blocks, threads>>>(
1607
+ (half*) vec.data_ptr(),
1608
+ (uint8_t*) mat.data_ptr(),
1609
+ (half*) mul.data_ptr(),
1610
+ (half*) scales.data_ptr(),
1611
+ (half*) zeros.data_ptr(),
1612
+ batch, heads, vec_row, height, width
1613
+ );
1614
+
1615
+ }
1616
+
1617
+
1618
+ __global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
1619
+ const half* __restrict__ vec, // [batch,heads, seq_q, seq_v]
1620
+ const uint8_t* __restrict__ mat, // [batch,heads, seq_v, head_dim]
1621
+ half* __restrict__ mul, // [batch,heads, seq_q,head_dim]
1622
+ const half* __restrict__ scales, // [batch,heads, seq_v]
1623
+ const half* __restrict__ zeros,
1624
+ int batch,
1625
+ int heads,
1626
+ int vec_row, //seq_q
1627
+ int height, //seq_v
1628
+ int width //head_dim
1629
+ ) {
1630
+ int weight_total = batch * heads * height * width;
1631
+ int input_total = batch * heads * vec_row * height;
1632
+ int out_total = batch * heads * vec_row * width;
1633
+ int tid = threadIdx.x;
1634
+ int h = BLOCKWIDTH * blockIdx.x; // vl
1635
+ int w = BLOCKWIDTH * blockIdx.y + tid; //head_dim + block
1636
+ if (w >= width && tid >= height) {
1637
+ return;
1638
+ }
1639
+ __shared__ half blockvec[BLOCKWIDTH];
1640
+ int k;
1641
+ half w_tmp1 = __float2half(0);
1642
+ half w_tmp2 = __float2half(0);
1643
+ int i = width * h + w;
1644
+ const int BLOCKWIDTH_half = BLOCKWIDTH/2;
1645
+ half2 weight[BLOCKWIDTH_half];
1646
+
1647
+ for (int b = 0; b < batch; ++b){
1648
+ for (int head = 0; head < heads; ++head){
1649
+ int batch_shift = b * heads + head;
1650
+ //int zero_index = batch_shift;
1651
+ for (k = 0; k < BLOCKWIDTH_half; ++k){
1652
+ int w_index1 = batch_shift * height * width + i + (2 * k) * width; // [batch,head, h+k, w]
1653
+ int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
1654
+ int zero_index1 = batch_shift * height + h + 2*k; // [batch,head, w]
1655
+ int zero_index2 = batch_shift * height + h + 2*k+1; // [batch,head, w]
1656
+
1657
+ if (w_index1 >= weight_total || (2 * k + h)>=height) {
1658
+ weight[k]=__float2half2_rn(0);
1659
+ } else{
1660
+ //int zero_index = batch_shift + h; // [batch,head, w]
1661
+ //float scale_f1 = __half2float(scales[zero_index1]);
1662
+ //float zero_f1 = __half2float(zeros[zero_index1]);
1663
+ if (w_index2>=weight_total){
1664
+ w_tmp1 = __float2half((as_unsigned(mat[w_index1]) - __half2float(zeros[zero_index1]))* __half2float(scales[zero_index1]));
1665
+ w_tmp2 = __float2half(0);
1666
+ weight[k] = __halves2half2(w_tmp1,w_tmp2);
1667
+ //printf("zero_index is %d k is %d w is %d head is %d height is %d width is %d w_index1 is %d w_tmp1 is %f w_tmp2 is %f zero is %f scale is %f low is %f high is %f \n ",zero_index,k,w,head,height, width,w_index1,__half2float(w_tmp1),__half2float(w_tmp2),zero_f,scale_f,__low2float(weight[k]),__high2float(weight[k]));
1668
+ }else{
1669
+ w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
1670
+ w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
1671
+ half zero1=zeros[zero_index1];
1672
+ half zero2=zeros[zero_index2];
1673
+ half scale1=scales[zero_index1];
1674
+ half scale2=scales[zero_index2];
1675
+ weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero1,zero2)),__halves2half2(scale1,scale2));
1676
+ //weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
1677
+ //printf("zero_index1 is %d zero_index2 is %d k is %d head is %d w is %d h is %d height is %d width is %d w_index1 is %d w_index2 is %d zero is %f scale is %f low is %f high is %f \n ",zero_index1,zero_index2,k,head,w,h,height, width,w_index1,w_index2,__half2float(zero1),__half2float(scale1),__low2float(weight[k]),__high2float(weight[k]));
1678
+ }
1679
+ }
1680
+ }
1681
+
1682
+
1683
+ for (int vr = 0; vr < vec_row; ++vr){
1684
+ float res=0;
1685
+ int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
1686
+ int out_index = (batch_shift * vec_row + vr) * width + w;
1687
+
1688
+ if (vec_index < input_total) {
1689
+ //blockvec[tid] = __half2float(vec[vec_index]);
1690
+ blockvec[tid] = vec[vec_index];
1691
+ //printf("vec_index is %d out_index is %d vec_row is %d ,vr is %d tid is %d blockvec is %f\n",vec_index,out_index,vec_row,vr,tid,blockvec[tid]);
1692
+ } else {
1693
+ blockvec[tid] = __float2half(0);
1694
+ //blockvec[tid] = 0;
1695
+ }
1696
+ __syncthreads();
1697
+ if (out_index < out_total) {
1698
+ for (k = 0; k < BLOCKWIDTH_half; ++k){
1699
+ half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
1700
+ res += __low2float(res2) + __high2float(res2);
1701
+ }
1702
+ atomicAdd(&mul[out_index], __float2half(res));
1703
+ }
1704
+ __syncthreads();
1705
+ }
1706
+ }
1707
+ }
1708
+ }
kernels/cpp_kernels.py ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from torch.utils import cpp_extension
2
+ import pathlib
3
+ import os
4
+ import subprocess
5
+
6
+ def _get_cuda_bare_metal_version(cuda_dir):
7
+ raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
8
+ universal_newlines=True)
9
+ output = raw_output.split()
10
+ release_idx = output.index("release") + 1
11
+ release = output[release_idx].split(".")
12
+ bare_metal_major = release[0]
13
+ bare_metal_minor = release[1][0]
14
+
15
+ return raw_output, bare_metal_major, bare_metal_minor
16
+
17
+ def _create_build_dir(buildpath):
18
+ try:
19
+ os.mkdir(buildpath)
20
+ except OSError:
21
+ if not os.path.isdir(buildpath):
22
+ print(f"Creation of the build directory {buildpath} failed")
23
+
24
+ # Check if cuda 11 is installed for compute capability 8.0
25
+ cc_flag = []
26
+ _, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
27
+ if int(bare_metal_major) >= 11:
28
+ cc_flag.append('-gencode')
29
+ cc_flag.append('arch=compute_80,code=sm_80')
30
+ if int(bare_metal_minor) >= 7:
31
+ cc_flag.append('-gencode')
32
+ cc_flag.append('arch=compute_90,code=sm_90')
33
+
34
+ # Build path
35
+ srcpath = pathlib.Path(__file__).parent.absolute()
36
+ buildpath = srcpath / 'build'
37
+ _create_build_dir(buildpath)
38
+
39
+ def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
40
+ return cpp_extension.load(
41
+ name=name,
42
+ sources=sources,
43
+ build_directory=buildpath,
44
+ extra_cflags=['-O3', ],
45
+ extra_cuda_cflags=['-O3',
46
+ '-gencode', 'arch=compute_70,code=sm_70',
47
+ '--use_fast_math'] + extra_cuda_flags + cc_flag,
48
+ verbose=1
49
+ )
50
+
51
+ extra_flags = []
52
+
53
+ cache_autogptq_cuda_256_sources = ["./kernels/cache_autogptq_cuda_256.cpp",
54
+ "./kernels/cache_autogptq_cuda_kernel_256.cu"]
55
+ cache_autogptq_cuda_256 = _cpp_extention_load_helper("cache_autogptq_cuda_256", cache_autogptq_cuda_256_sources, extra_flags)
modeling_qwen.py CHANGED
@@ -31,6 +31,7 @@ try:
31
  except ImportError:
32
  rearrange = None
33
  from torch import nn
 
34
 
35
  SUPPORT_CUDA = torch.cuda.is_available()
36
  SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
@@ -75,7 +76,6 @@ apply_rotary_emb_func = None
75
  rms_norm = None
76
  flash_attn_unpadded_func = None
77
 
78
-
79
  def _import_flash_attn():
80
  global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
81
  try:
@@ -112,6 +112,31 @@ def _import_flash_attn():
112
  "https://github.com/Dao-AILab/flash-attention"
113
  )
114
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
115
 
116
  class FlashSelfAttention(torch.nn.Module):
117
  def __init__(
@@ -254,19 +279,51 @@ class QWenAttention(nn.Module):
254
  self.register_buffer("logn_tensor", logn_tensor, persistent=False)
255
 
256
  self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
 
 
 
 
 
 
 
 
 
257
 
258
  def _attn(self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None):
259
- attn_weights = torch.matmul(query, key.transpose(-1, -2))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
260
 
261
  if self.scale_attn_weights:
 
 
 
 
262
  attn_weights = attn_weights / torch.full(
263
  [],
264
- value.size(-1) ** 0.5,
265
  dtype=attn_weights.dtype,
266
  device=attn_weights.device,
267
  )
268
-
269
- query_length, key_length = query.size(-2), key.size(-2)
 
 
270
  causal_mask = registered_causal_mask[
271
  :, :, key_length - query_length : key_length, :key_length
272
  ]
@@ -283,13 +340,32 @@ class QWenAttention(nn.Module):
283
 
284
  attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1)
285
 
286
- attn_weights = attn_weights.type(value.dtype)
287
  attn_weights = self.attn_dropout(attn_weights)
288
 
289
  if head_mask is not None:
290
  attn_weights = attn_weights * head_mask
291
 
292
- attn_output = torch.matmul(attn_weights, value)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
293
  attn_output = attn_output.transpose(1, 2)
294
 
295
  return attn_output, attn_weights
@@ -373,7 +449,6 @@ class QWenAttention(nn.Module):
373
  output_attentions: Optional[bool] = False,
374
  use_cache: Optional[bool] = False,
375
  ):
376
-
377
  mixed_x_layer = self.c_attn(hidden_states)
378
 
379
  query, key, value = mixed_x_layer.split(self.split_size, dim=2)
@@ -405,10 +480,34 @@ class QWenAttention(nn.Module):
405
  query = torch.cat(query_list, dim=0)
406
  key = torch.cat(key_list, dim=0)
407
 
 
 
 
 
 
 
 
 
 
 
 
408
  if layer_past is not None:
409
  past_key, past_value = layer_past[0], layer_past[1]
410
- key = torch.cat((past_key, key), dim=1)
411
- value = torch.cat((past_value, value), dim=1)
 
 
 
 
 
 
 
 
 
 
 
 
 
412
 
413
  if use_cache:
414
  present = (key, value)
@@ -416,8 +515,12 @@ class QWenAttention(nn.Module):
416
  present = None
417
 
418
  if self.use_logn_attn and not self.training:
419
- seq_start = key.size(1) - query.size(1)
420
- seq_end = key.size(1)
 
 
 
 
421
  logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
422
  query = query * logn_tensor.expand_as(query)
423
 
@@ -435,8 +538,9 @@ class QWenAttention(nn.Module):
435
 
436
  else:
437
  query = query.permute(0, 2, 1, 3)
438
- key = key.permute(0, 2, 1, 3)
439
- value = value.permute(0, 2, 1, 3)
 
440
  if (
441
  registered_causal_mask is None
442
  and self.use_flash_attn
@@ -597,6 +701,7 @@ class QWenModel(QWenPreTrainedModel):
597
  self.vocab_size = config.vocab_size
598
  self.num_hidden_layers = config.num_hidden_layers
599
  self.embed_dim = config.hidden_size
 
600
 
601
  self.gradient_checkpointing = False
602
  self.use_dynamic_ntk = config.use_dynamic_ntk
@@ -721,8 +826,10 @@ class QWenModel(QWenPreTrainedModel):
721
  past_length = 0
722
  past_key_values = tuple([None] * len(self.h))
723
  else:
724
- past_length = past_key_values[0][0].size(-2)
725
-
 
 
726
  if position_ids is None:
727
  position_ids = torch.arange(
728
  past_length,
@@ -750,7 +857,10 @@ class QWenModel(QWenPreTrainedModel):
750
  kv_seq_len = hidden_states.size()[1]
751
  if past_key_values[0] is not None:
752
  # past key values[0][0] shape: bs * seq_len * head_num * dim
753
- kv_seq_len += past_key_values[0][0].shape[1]
 
 
 
754
 
755
  if self.training or not self.use_dynamic_ntk:
756
  ntk_alpha_list = [1.0]
@@ -907,6 +1017,12 @@ class QWenLMHeadModel(QWenPreTrainedModel):
907
  if config.use_flash_attn:
908
  _import_flash_attn()
909
 
 
 
 
 
 
 
910
  self.transformer = QWenModel(config)
911
  self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
912
 
@@ -918,6 +1034,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
918
  self.lm_head.half()
919
  self.post_init()
920
 
 
921
  def get_output_embeddings(self):
922
  return self.lm_head
923
 
 
31
  except ImportError:
32
  rearrange = None
33
  from torch import nn
34
+ from kernels.cpp_kernels import cache_autogptq_cuda_256
35
 
36
  SUPPORT_CUDA = torch.cuda.is_available()
37
  SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
 
76
  rms_norm = None
77
  flash_attn_unpadded_func = None
78
 
 
79
  def _import_flash_attn():
80
  global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
81
  try:
 
112
  "https://github.com/Dao-AILab/flash-attention"
113
  )
114
 
115
+ def quantize_cache_v(fdata, bits, qmax, qmin):
116
+ # b, s, head, h-dim->b, head, s, h-dim
117
+ qtype = torch.uint8
118
+ device = fdata.device
119
+ shape = fdata.shape
120
+
121
+ fdata_cal = torch.flatten(fdata, 2)
122
+ fmax = torch.amax(fdata_cal, dim=-1, keepdim=True)
123
+ fmin = torch.amin(fdata_cal, dim=-1, keepdim=True)
124
+ # Compute params
125
+ if qmax.device != fmax.device:
126
+ qmax = qmax.to(device)
127
+ qmin = qmin.to(device)
128
+ scale = (fmax - fmin) / (qmax - qmin)
129
+ zero = qmin - fmin / scale
130
+ scale = scale.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
131
+ zero = zero.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
132
+ # Quantize
133
+ res_data = fdata / scale + zero
134
+ qdata = torch.clamp(res_data, qmin, qmax).to(qtype)
135
+ return qdata.contiguous(), scale, zero
136
+
137
+ def dequantize_cache_torch(qdata, scale, zero):
138
+ data = scale * (qdata - zero)
139
+ return data
140
 
141
  class FlashSelfAttention(torch.nn.Module):
142
  def __init__(
 
279
  self.register_buffer("logn_tensor", logn_tensor, persistent=False)
280
 
281
  self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
282
+ self.use_cache_quantization = config.use_cache_quantization if hasattr(config, 'use_cache_quantization') else False
283
+ self.use_cache_kernel = config.use_cache_kernel if hasattr(config,'use_cache_kernel') else False
284
+ cache_dtype = torch.float
285
+ if self.bf16:
286
+ cache_dtype=torch.bfloat16
287
+ elif config.fp16:
288
+ cache_dtype = torch.float16
289
+ self.cache_qmax = torch.tensor(torch.iinfo(torch.uint8).max, dtype=cache_dtype)
290
+ self.cache_qmin = torch.tensor(torch.iinfo(torch.uint8).min, dtype=cache_dtype)
291
 
292
  def _attn(self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None):
293
+ device = query.device
294
+ if self.use_cache_quantization:
295
+ qk, qk_scale, qk_zero = key
296
+ if self.use_cache_kernel:
297
+ shape = query.shape[:-1] + (qk.shape[-2],)
298
+ attn_weights = torch.zeros(shape, dtype=torch.float16, device=device)
299
+ cache_autogptq_cuda_256.vecquant8matmul_batched_faster_old(
300
+ query.contiguous() if query.dtype == torch.float16 else query.to(torch.float16).contiguous(),
301
+ qk.transpose(-1, -2).contiguous(),
302
+ attn_weights,
303
+ qk_scale.contiguous() if qk_scale.dtype == torch.float16 else qk_scale.to(torch.float16).contiguous(),
304
+ qk_zero.contiguous()if qk_zero.dtype == torch.float16 else qk_zero.to(torch.float16).contiguous())
305
+ # attn_weights = attn_weights.to(query.dtype).contiguous()
306
+ else:
307
+ key = dequantize_cache_torch(qk, qk_scale, qk_zero)
308
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
309
+ else:
310
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
311
 
312
  if self.scale_attn_weights:
313
+ if self.use_cache_quantization:
314
+ size_temp = value[0].size(-1)
315
+ else:
316
+ size_temp = value.size(-1)
317
  attn_weights = attn_weights / torch.full(
318
  [],
319
+ size_temp ** 0.5,
320
  dtype=attn_weights.dtype,
321
  device=attn_weights.device,
322
  )
323
+ if self.use_cache_quantization:
324
+ query_length, key_length = query.size(-2), key[0].size(-2)
325
+ else:
326
+ query_length, key_length = query.size(-2), key.size(-2)
327
  causal_mask = registered_causal_mask[
328
  :, :, key_length - query_length : key_length, :key_length
329
  ]
 
340
 
341
  attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1)
342
 
343
+ attn_weights = attn_weights.type(query.dtype)
344
  attn_weights = self.attn_dropout(attn_weights)
345
 
346
  if head_mask is not None:
347
  attn_weights = attn_weights * head_mask
348
 
349
+ if self.use_cache_quantization:
350
+ qv, qv_scale, qv_zero = value
351
+ if self.use_cache_kernel:
352
+ shape = attn_weights.shape[:-1] + (query.shape[-1],)
353
+ attn_output = torch.zeros(shape, dtype=torch.float16, device=device)
354
+ cache_autogptq_cuda_256.vecquant8matmul_batched_column_compression_faster_old(
355
+ attn_weights.contiguous() if attn_weights.dtype == torch.float16 else attn_weights.to(torch.float16).contiguous(),
356
+ qv.contiguous(), # dtype: int32
357
+ attn_output,
358
+ qv_scale.contiguous() if qv_scale.dtype == torch.float16 else qv_scale.to(torch.float16).contiguous(),
359
+ qv_zero.contiguous() if qv_zero.dtype == torch.float16 else qv_zero.to(torch.float16).contiguous())
360
+ if attn_output.dtype != query.dtype:
361
+ attn_output = attn_output.to(query.dtype)
362
+ attn_weights = attn_weights.to(query.dtype)
363
+ else:
364
+ value = dequantize_cache_torch(qv, qv_scale, qv_zero)
365
+ attn_output = torch.matmul(attn_weights, value)
366
+ else:
367
+ attn_output = torch.matmul(attn_weights, value)
368
+
369
  attn_output = attn_output.transpose(1, 2)
370
 
371
  return attn_output, attn_weights
 
449
  output_attentions: Optional[bool] = False,
450
  use_cache: Optional[bool] = False,
451
  ):
 
452
  mixed_x_layer = self.c_attn(hidden_states)
453
 
454
  query, key, value = mixed_x_layer.split(self.split_size, dim=2)
 
480
  query = torch.cat(query_list, dim=0)
481
  key = torch.cat(key_list, dim=0)
482
 
483
+ if self.use_cache_quantization:
484
+ key = quantize_cache_v(key.permute(0, 2, 1, 3),
485
+ bits=8,
486
+ qmin=self.cache_qmin,
487
+ qmax=self.cache_qmax)
488
+ value = quantize_cache_v(value.permute(0, 2, 1, 3),
489
+ bits=8,
490
+ qmin=self.cache_qmin,
491
+ qmax=self.cache_qmax)
492
+
493
+
494
  if layer_past is not None:
495
  past_key, past_value = layer_past[0], layer_past[1]
496
+ if self.use_cache_quantization:
497
+ # use_cache_quantization:
498
+ # present=((q_key,key_scale,key_zero_point),
499
+ # (q_value,value_scale,value_zero_point))
500
+ key = (torch.cat((past_key[0], key[0]), dim=2),
501
+ torch.cat((past_key[1], key[1]), dim=2),
502
+ torch.cat((past_key[2], key[2]), dim=2))
503
+ value = (torch.cat((past_value[0], value[0]), dim=2),
504
+ torch.cat((past_value[1], value[1]), dim=2),
505
+ torch.cat((past_value[2], value[2]), dim=2))
506
+ else:
507
+ # not use_cache_quantization:
508
+ # present=(key,value)
509
+ key = torch.cat((past_key, key), dim=1)
510
+ value = torch.cat((past_value, value), dim=1)
511
 
512
  if use_cache:
513
  present = (key, value)
 
515
  present = None
516
 
517
  if self.use_logn_attn and not self.training:
518
+ if self.use_cache_quantization:
519
+ seq_start = key[0].size(2) - query.size(1)
520
+ seq_end = key[0].size(2)
521
+ else:
522
+ seq_start = key.size(1) - query.size(1)
523
+ seq_end = key.size(1)
524
  logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
525
  query = query * logn_tensor.expand_as(query)
526
 
 
538
 
539
  else:
540
  query = query.permute(0, 2, 1, 3)
541
+ if not self.use_cache_quantization:
542
+ key = key.permute(0, 2, 1, 3)
543
+ value = value.permute(0, 2, 1, 3)
544
  if (
545
  registered_causal_mask is None
546
  and self.use_flash_attn
 
701
  self.vocab_size = config.vocab_size
702
  self.num_hidden_layers = config.num_hidden_layers
703
  self.embed_dim = config.hidden_size
704
+ self.use_cache_quantization = self.config.use_cache_quantization if hasattr(self.config, 'use_cache_quantization') else False
705
 
706
  self.gradient_checkpointing = False
707
  self.use_dynamic_ntk = config.use_dynamic_ntk
 
826
  past_length = 0
827
  past_key_values = tuple([None] * len(self.h))
828
  else:
829
+ if self.use_cache_quantization:
830
+ past_length = past_key_values[0][0][0].size(2)
831
+ else:
832
+ past_length = past_key_values[0][0].size(-2)
833
  if position_ids is None:
834
  position_ids = torch.arange(
835
  past_length,
 
857
  kv_seq_len = hidden_states.size()[1]
858
  if past_key_values[0] is not None:
859
  # past key values[0][0] shape: bs * seq_len * head_num * dim
860
+ if self.use_cache_quantization:
861
+ kv_seq_len += past_key_values[0][0][0].shape[2]
862
+ else:
863
+ kv_seq_len += past_key_values[0][0].shape[1]
864
 
865
  if self.training or not self.use_dynamic_ntk:
866
  ntk_alpha_list = [1.0]
 
1017
  if config.use_flash_attn:
1018
  _import_flash_attn()
1019
 
1020
+
1021
+ if hasattr(config, 'use_cache_quantization') and config.use_cache_quantization:
1022
+ config.use_flash_attn = False
1023
+ if hasattr(config, 'use_cache_kernel') and config.use_cache_kernel:
1024
+ from kernels.cpp_kernels import cache_autogptq_cuda_256
1025
+
1026
  self.transformer = QWenModel(config)
1027
  self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1028
 
 
1034
  self.lm_head.half()
1035
  self.post_init()
1036
 
1037
+
1038
  def get_output_embeddings(self):
1039
  return self.lm_head
1040