|
33 | 33 | // repack.cpp |
34 | 34 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
35 | 35 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
| 36 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
36 | 37 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
37 | 38 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
38 | 39 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
39 | 40 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
| 41 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
40 | 42 | #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K |
41 | 43 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
42 | 44 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
43 | 45 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
44 | 46 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
45 | 47 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
46 | 48 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
| 49 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
47 | 50 | #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K |
48 | 51 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
49 | 52 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
50 | 53 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
51 | 54 | #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64) |
52 | 55 | // repack.cpp |
| 56 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
53 | 57 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
54 | 58 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
55 | 59 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
|
58 | 62 | #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64) |
59 | 63 | // repack.cpp |
60 | 64 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
| 65 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
61 | 66 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
62 | 67 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
| 68 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
63 | 69 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
64 | 70 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
65 | 71 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
| 72 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
66 | 73 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
67 | 74 | #elif defined(__POWERPC__) || defined(__powerpc__) |
68 | 75 | // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679 |
|
74 | 81 | // repack.cpp |
75 | 82 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
76 | 83 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
| 84 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
77 | 85 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
78 | 86 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
79 | 87 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
80 | 88 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
| 89 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
81 | 90 | #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K |
82 | 91 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
83 | 92 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
84 | 93 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
85 | 94 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
86 | 95 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
87 | 96 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
| 97 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
88 | 98 | #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K |
89 | 99 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
90 | 100 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
|
99 | 109 | // repack.cpp |
100 | 110 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
101 | 111 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
| 112 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
102 | 113 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
103 | 114 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
104 | 115 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
105 | 116 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
| 117 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
106 | 118 | #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K |
107 | 119 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
108 | 120 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
109 | 121 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
110 | 122 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
111 | 123 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
112 | 124 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
| 125 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
113 | 126 | #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K |
114 | 127 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
115 | 128 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
|
132 | 145 | // repack.cpp |
133 | 146 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
134 | 147 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
| 148 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
135 | 149 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
136 | 150 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
137 | 151 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
| 152 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
138 | 153 | #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K |
139 | 154 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
140 | 155 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
141 | 156 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
142 | 157 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
143 | 158 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
| 159 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
144 | 160 | #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K |
145 | 161 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
146 | 162 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
|
161 | 177 | // repack.cpp |
162 | 178 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
163 | 179 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
| 180 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
164 | 181 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
165 | 182 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
166 | 183 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
167 | 184 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
| 185 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
168 | 186 | #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K |
169 | 187 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
170 | 188 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
171 | 189 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
172 | 190 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
173 | 191 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
174 | 192 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
| 193 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
175 | 194 | #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K |
176 | 195 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
177 | 196 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
|
194 | 213 | // repack.cpp |
195 | 214 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
196 | 215 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
| 216 | +#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
197 | 217 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
198 | 218 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
199 | 219 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
200 | 220 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
| 221 | +#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
201 | 222 | #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K |
202 | 223 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
203 | 224 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
204 | 225 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
205 | 226 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
206 | 227 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
207 | 228 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
| 229 | +#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
208 | 230 | #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K |
209 | 231 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
210 | 232 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
|
0 commit comments