Upload 5 files
Browse files- GPT2.cbp +0 -5
- GPT2.cscope_file_list +11 -15
- GPT2.depend +6 -5
- GPT2.layout +29 -29
- main-ctx.cpp +1213 -839
GPT2.cbp
CHANGED
|
@@ -34,11 +34,6 @@
|
|
| 34 |
</Compiler>
|
| 35 |
<Unit filename="GPT2.cbp" />
|
| 36 |
<Unit filename="GPT2.layout" />
|
| 37 |
-
<Unit filename="common-ggml.cpp" />
|
| 38 |
-
<Unit filename="common-ggml.h" />
|
| 39 |
-
<Unit filename="common.cpp" />
|
| 40 |
-
<Unit filename="common.h" />
|
| 41 |
-
<Unit filename="dr_wav.h" />
|
| 42 |
<Unit filename="ggml-aarch64.c">
|
| 43 |
<Option compilerVar="CC" />
|
| 44 |
</Unit>
|
|
|
|
| 34 |
</Compiler>
|
| 35 |
<Unit filename="GPT2.cbp" />
|
| 36 |
<Unit filename="GPT2.layout" />
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
<Unit filename="ggml-aarch64.c">
|
| 38 |
<Option compilerVar="CC" />
|
| 39 |
</Unit>
|
GPT2.cscope_file_list
CHANGED
|
@@ -1,22 +1,18 @@
|
|
| 1 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-alloc.h"
|
| 2 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-alloc.c"
|
| 3 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-impl.h"
|
| 4 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common-ggml.cpp"
|
| 5 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-quants.c"
|
| 6 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-aarch64.h"
|
| 7 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common.cpp"
|
| 8 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-quants.h"
|
| 9 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-backend-impl.h"
|
| 10 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml.h"
|
| 11 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml.c"
|
| 13 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-backend.h"
|
| 14 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common.h"
|
| 15 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\
|
| 16 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\
|
| 17 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-cpu-impl.h"
|
| 18 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-aarch64.c"
|
| 19 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\GPT2.layout"
|
| 20 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-backend.cpp"
|
| 21 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\dr_wav.h"
|
| 22 |
-
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-aarch64.h"
|
| 2 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common.cpp"
|
|
|
|
|
|
|
| 3 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml.h"
|
| 4 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-quants.c"
|
| 5 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-impl.h"
|
| 6 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common-ggml.cpp"
|
| 7 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-aarch64.c"
|
| 8 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\GPT2.layout"
|
| 9 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml.c"
|
|
|
|
| 10 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common.h"
|
| 11 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\main-ctx.cpp"
|
| 12 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-quants.h"
|
| 13 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-cpu-impl.h"
|
|
|
|
|
|
|
|
|
|
| 14 |
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\dr_wav.h"
|
| 15 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\common-ggml.h"
|
| 16 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\GPT2.cbp"
|
| 17 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\ggml-common.h"
|
| 18 |
+
"C:\Users\Caleb P. Nwokocha\CodeBlocksProjects\GPT2\quantize.cpp"
|
GPT2.depend
CHANGED
|
@@ -222,12 +222,8 @@
|
|
| 222 |
|
| 223 |
1730683892 source:c:\users\caleb p. nwokocha\codeblocksprojects\gpt2\main-alloc.cpp
|
| 224 |
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
1730534952 source:c:\users\caleb p. nwokocha\codeblocksprojects\gpt2\quantize.cpp
|
| 228 |
"ggml.h"
|
| 229 |
-
"common.h"
|
| 230 |
-
"common-ggml.h"
|
| 231 |
<cassert>
|
| 232 |
<cmath>
|
| 233 |
<cstdio>
|
|
@@ -236,5 +232,10 @@
|
|
| 236 |
<map>
|
| 237 |
<string>
|
| 238 |
<vector>
|
|
|
|
|
|
|
|
|
|
| 239 |
<regex>
|
| 240 |
|
|
|
|
|
|
|
|
|
| 222 |
|
| 223 |
1730683892 source:c:\users\caleb p. nwokocha\codeblocksprojects\gpt2\main-alloc.cpp
|
| 224 |
|
| 225 |
+
1731878749 source:c:\users\caleb p. nwokocha\codeblocksprojects\gpt2\main-ctx.cpp
|
|
|
|
|
|
|
| 226 |
"ggml.h"
|
|
|
|
|
|
|
| 227 |
<cassert>
|
| 228 |
<cmath>
|
| 229 |
<cstdio>
|
|
|
|
| 232 |
<map>
|
| 233 |
<string>
|
| 234 |
<vector>
|
| 235 |
+
<thread>
|
| 236 |
+
<ctime>
|
| 237 |
+
<random>
|
| 238 |
<regex>
|
| 239 |
|
| 240 |
+
1730831644 source:c:\users\caleb p. nwokocha\codeblocksprojects\gpt2\quantize.cpp
|
| 241 |
+
|
GPT2.layout
CHANGED
|
@@ -2,47 +2,42 @@
|
|
| 2 |
<CodeBlocks_layout_file>
|
| 3 |
<FileVersion major="1" minor="0" />
|
| 4 |
<ActiveTarget name="Debug" />
|
| 5 |
-
<File name="ggml-
|
| 6 |
<Cursor>
|
| 7 |
-
<Cursor1 position="
|
| 8 |
</Cursor>
|
| 9 |
</File>
|
| 10 |
-
<File name="
|
| 11 |
<Cursor>
|
| 12 |
-
<Cursor1 position="
|
| 13 |
</Cursor>
|
| 14 |
</File>
|
| 15 |
-
<File name="
|
| 16 |
<Cursor>
|
| 17 |
-
<Cursor1 position="
|
| 18 |
-
</Cursor>
|
| 19 |
-
</File>
|
| 20 |
-
<File name="ggml-aarch64.h" open="1" top="0" tabpos="8" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 21 |
-
<Cursor>
|
| 22 |
-
<Cursor1 position="1519" topLine="0" />
|
| 23 |
</Cursor>
|
| 24 |
</File>
|
| 25 |
-
<File name="common.
|
| 26 |
<Cursor>
|
| 27 |
-
<Cursor1 position="
|
| 28 |
</Cursor>
|
| 29 |
</File>
|
| 30 |
-
<File name="ggml-
|
| 31 |
<Cursor>
|
| 32 |
-
<Cursor1 position="
|
| 33 |
</Cursor>
|
| 34 |
</File>
|
| 35 |
-
<File name="
|
| 36 |
<Cursor>
|
| 37 |
-
<Cursor1 position="
|
| 38 |
</Cursor>
|
| 39 |
</File>
|
| 40 |
-
<File name="
|
| 41 |
<Cursor>
|
| 42 |
-
<Cursor1 position="
|
| 43 |
</Cursor>
|
| 44 |
</File>
|
| 45 |
-
<File name="ggml-
|
| 46 |
<Cursor>
|
| 47 |
<Cursor1 position="0" topLine="0" />
|
| 48 |
</Cursor>
|
|
@@ -52,29 +47,34 @@
|
|
| 52 |
<Cursor1 position="522" topLine="0" />
|
| 53 |
</Cursor>
|
| 54 |
</File>
|
| 55 |
-
<File name="common.h" open="1" top="0" tabpos="
|
| 56 |
<Cursor>
|
| 57 |
<Cursor1 position="0" topLine="0" />
|
| 58 |
</Cursor>
|
| 59 |
</File>
|
| 60 |
-
<File name="
|
| 61 |
<Cursor>
|
| 62 |
-
<Cursor1 position="
|
| 63 |
</Cursor>
|
| 64 |
</File>
|
| 65 |
-
<File name="
|
| 66 |
<Cursor>
|
| 67 |
-
<Cursor1 position="
|
| 68 |
</Cursor>
|
| 69 |
</File>
|
| 70 |
-
<File name="ggml
|
| 71 |
<Cursor>
|
| 72 |
-
<Cursor1 position="
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
</Cursor>
|
| 74 |
</File>
|
| 75 |
-
<File name="
|
| 76 |
<Cursor>
|
| 77 |
-
<Cursor1 position="
|
| 78 |
</Cursor>
|
| 79 |
</File>
|
| 80 |
</CodeBlocks_layout_file>
|
|
|
|
| 2 |
<CodeBlocks_layout_file>
|
| 3 |
<FileVersion major="1" minor="0" />
|
| 4 |
<ActiveTarget name="Debug" />
|
| 5 |
+
<File name="ggml-quants.c" open="1" top="0" tabpos="11" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 6 |
<Cursor>
|
| 7 |
+
<Cursor1 position="2705" topLine="0" />
|
| 8 |
</Cursor>
|
| 9 |
</File>
|
| 10 |
+
<File name="ggml-quants.h" open="1" top="0" tabpos="13" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 11 |
<Cursor>
|
| 12 |
+
<Cursor1 position="0" topLine="128" />
|
| 13 |
</Cursor>
|
| 14 |
</File>
|
| 15 |
+
<File name="quantize.cpp" open="1" top="0" tabpos="15" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 16 |
<Cursor>
|
| 17 |
+
<Cursor1 position="4241" topLine="139" />
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
</Cursor>
|
| 19 |
</File>
|
| 20 |
+
<File name="common.h" open="1" top="0" tabpos="2" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 21 |
<Cursor>
|
| 22 |
+
<Cursor1 position="0" topLine="0" />
|
| 23 |
</Cursor>
|
| 24 |
</File>
|
| 25 |
+
<File name="ggml-aarch64.h" open="1" top="0" tabpos="8" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 26 |
<Cursor>
|
| 27 |
+
<Cursor1 position="1519" topLine="0" />
|
| 28 |
</Cursor>
|
| 29 |
</File>
|
| 30 |
+
<File name="common-ggml.cpp" open="1" top="0" tabpos="4" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 31 |
<Cursor>
|
| 32 |
+
<Cursor1 position="223" topLine="135" />
|
| 33 |
</Cursor>
|
| 34 |
</File>
|
| 35 |
+
<File name="main-ctx.cpp" open="1" top="1" tabpos="12" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 36 |
<Cursor>
|
| 37 |
+
<Cursor1 position="26874" topLine="722" />
|
| 38 |
</Cursor>
|
| 39 |
</File>
|
| 40 |
+
<File name="ggml-cpu-impl.h" open="1" top="0" tabpos="9" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 41 |
<Cursor>
|
| 42 |
<Cursor1 position="0" topLine="0" />
|
| 43 |
</Cursor>
|
|
|
|
| 47 |
<Cursor1 position="522" topLine="0" />
|
| 48 |
</Cursor>
|
| 49 |
</File>
|
| 50 |
+
<File name="ggml-common.h" open="1" top="0" tabpos="14" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 51 |
<Cursor>
|
| 52 |
<Cursor1 position="0" topLine="0" />
|
| 53 |
</Cursor>
|
| 54 |
</File>
|
| 55 |
+
<File name="ggml-aarch64.c" open="1" top="0" tabpos="7" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 56 |
<Cursor>
|
| 57 |
+
<Cursor1 position="442" topLine="0" />
|
| 58 |
</Cursor>
|
| 59 |
</File>
|
| 60 |
+
<File name="common.cpp" open="1" top="0" tabpos="5" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 61 |
<Cursor>
|
| 62 |
+
<Cursor1 position="152" topLine="0" />
|
| 63 |
</Cursor>
|
| 64 |
</File>
|
| 65 |
+
<File name="ggml.h" open="1" top="0" tabpos="1" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 66 |
<Cursor>
|
| 67 |
+
<Cursor1 position="26839" topLine="760" />
|
| 68 |
+
</Cursor>
|
| 69 |
+
</File>
|
| 70 |
+
<File name="ggml-impl.h" open="1" top="0" tabpos="10" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 71 |
+
<Cursor>
|
| 72 |
+
<Cursor1 position="6388" topLine="0" />
|
| 73 |
</Cursor>
|
| 74 |
</File>
|
| 75 |
+
<File name="common-ggml.h" open="1" top="0" tabpos="3" split="0" active="1" splitpos="0" zoom_1="0" zoom_2="0">
|
| 76 |
<Cursor>
|
| 77 |
+
<Cursor1 position="141" topLine="0" />
|
| 78 |
</Cursor>
|
| 79 |
</File>
|
| 80 |
</CodeBlocks_layout_file>
|
main-ctx.cpp
CHANGED
|
@@ -1,841 +1,1215 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
//
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
//
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
//
|
| 29 |
-
|
| 30 |
-
//
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
//
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
//
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
//
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
//
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
//
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
//
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
//
|
| 97 |
-
|
| 98 |
-
//
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
//
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
//
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
//
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
//
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
//
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
//
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
//
|
| 183 |
-
//
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
//
|
| 205 |
-
|
| 206 |
-
//
|
| 207 |
-
|
| 208 |
-
//
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
//
|
| 214 |
-
//
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
//
|
| 218 |
-
|
| 219 |
-
//
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
//
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
//
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
//
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
//
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
//
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
//
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
-
|
| 309 |
-
|
| 310 |
-
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
|
| 316 |
-
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
|
| 320 |
-
//
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
//
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
|
| 363 |
-
|
| 364 |
-
|
| 365 |
-
|
| 366 |
-
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
|
| 372 |
-
|
| 373 |
-
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
|
| 385 |
-
|
| 386 |
-
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
|
| 392 |
-
|
| 393 |
-
|
| 394 |
-
|
| 395 |
-
|
| 396 |
-
|
| 397 |
-
|
| 398 |
-
|
| 399 |
-
|
| 400 |
-
//
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
|
| 404 |
-
|
| 405 |
-
|
| 406 |
-
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
|
| 415 |
-
|
| 416 |
-
|
| 417 |
-
|
| 418 |
-
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
|
| 422 |
-
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
|
| 429 |
-
|
| 430 |
-
|
| 431 |
-
|
| 432 |
-
|
| 433 |
-
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
//
|
| 437 |
-
|
| 438 |
-
|
| 439 |
-
|
| 440 |
-
//
|
| 441 |
-
|
| 442 |
-
|
| 443 |
-
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
|
| 451 |
-
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
//
|
| 456 |
-
|
| 457 |
-
//
|
| 458 |
-
|
| 459 |
-
|
| 460 |
-
|
| 461 |
-
|
| 462 |
-
|
| 463 |
-
//
|
| 464 |
-
|
| 465 |
-
|
| 466 |
-
|
| 467 |
-
|
| 468 |
-
|
| 469 |
-
//
|
| 470 |
-
|
| 471 |
-
|
| 472 |
-
|
| 473 |
-
|
| 474 |
-
|
| 475 |
-
|
| 476 |
-
//
|
| 477 |
-
|
| 478 |
-
|
| 479 |
-
|
| 480 |
-
|
| 481 |
-
//
|
| 482 |
-
|
| 483 |
-
|
| 484 |
-
|
| 485 |
-
|
| 486 |
-
|
| 487 |
-
|
| 488 |
-
|
| 489 |
-
|
| 490 |
-
|
| 491 |
-
|
| 492 |
-
|
| 493 |
-
|
| 494 |
-
|
| 495 |
-
|
| 496 |
-
|
| 497 |
-
|
| 498 |
-
|
| 499 |
-
|
| 500 |
-
|
| 501 |
-
|
| 502 |
-
|
| 503 |
-
|
| 504 |
-
|
| 505 |
-
|
| 506 |
-
|
| 507 |
-
|
| 508 |
-
|
| 509 |
-
|
| 510 |
-
|
| 511 |
-
|
| 512 |
-
|
| 513 |
-
|
| 514 |
-
|
| 515 |
-
|
| 516 |
-
|
| 517 |
-
|
| 518 |
-
|
| 519 |
-
|
| 520 |
-
|
| 521 |
-
|
| 522 |
-
|
| 523 |
-
|
| 524 |
-
|
| 525 |
-
|
| 526 |
-
|
| 527 |
-
|
| 528 |
-
|
| 529 |
-
|
| 530 |
-
|
| 531 |
-
|
| 532 |
-
|
| 533 |
-
|
| 534 |
-
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
|
| 538 |
-
//
|
| 539 |
-
|
| 540 |
-
|
| 541 |
-
|
| 542 |
-
|
| 543 |
-
|
| 544 |
-
|
| 545 |
-
|
| 546 |
-
|
| 547 |
-
|
| 548 |
-
|
| 549 |
-
|
| 550 |
-
|
| 551 |
-
|
| 552 |
-
|
| 553 |
-
|
| 554 |
-
|
| 555 |
-
|
| 556 |
-
|
| 557 |
-
|
| 558 |
-
|
| 559 |
-
|
| 560 |
-
|
| 561 |
-
|
| 562 |
-
|
| 563 |
-
|
| 564 |
-
|
| 565 |
-
|
| 566 |
-
|
| 567 |
-
|
| 568 |
-
|
| 569 |
-
|
| 570 |
-
|
| 571 |
-
|
| 572 |
-
|
| 573 |
-
|
| 574 |
-
|
| 575 |
-
//
|
| 576 |
-
|
| 577 |
-
|
| 578 |
-
|
| 579 |
-
|
| 580 |
-
|
| 581 |
-
|
| 582 |
-
|
| 583 |
-
|
| 584 |
-
|
| 585 |
-
|
| 586 |
-
|
| 587 |
-
|
| 588 |
-
|
| 589 |
-
|
| 590 |
-
//
|
| 591 |
-
|
| 592 |
-
|
| 593 |
-
|
| 594 |
-
|
| 595 |
-
|
| 596 |
-
|
| 597 |
-
|
| 598 |
-
|
| 599 |
-
|
| 600 |
-
|
| 601 |
-
//
|
| 602 |
-
|
| 603 |
-
|
| 604 |
-
|
| 605 |
-
|
| 606 |
-
|
| 607 |
-
|
| 608 |
-
|
| 609 |
-
|
| 610 |
-
|
| 611 |
-
|
| 612 |
-
|
| 613 |
-
|
| 614 |
-
|
| 615 |
-
|
| 616 |
-
|
| 617 |
-
|
| 618 |
-
|
| 619 |
-
|
| 620 |
-
|
| 621 |
-
|
| 622 |
-
|
| 623 |
-
|
| 624 |
-
//
|
| 625 |
-
//
|
| 626 |
-
//
|
| 627 |
-
//
|
| 628 |
-
//
|
| 629 |
-
|
| 630 |
-
|
| 631 |
-
|
| 632 |
-
|
| 633 |
-
|
| 634 |
-
|
| 635 |
-
|
| 636 |
-
//
|
| 637 |
-
// model.layers[il].c_mlp_proj_w,
|
| 638 |
-
// cur);
|
| 639 |
-
//
|
| 640 |
-
// cur = ggml_add(ctx0,
|
| 641 |
-
// ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
|
| 642 |
-
// cur);
|
| 643 |
-
// }
|
| 644 |
-
//
|
| 645 |
-
// // input for next layer
|
| 646 |
-
// inpL = ggml_add(ctx0, cur, inpFF);
|
| 647 |
-
// }
|
| 648 |
-
//
|
| 649 |
-
// // norm
|
| 650 |
-
// {
|
| 651 |
-
// // [ 768, N]
|
| 652 |
-
// inpL = ggml_norm(ctx0, inpL, hparams.eps);
|
| 653 |
-
//
|
| 654 |
-
// // inpL = ln_f_g*inpL + ln_f_b
|
| 655 |
-
// // [ 768, N]
|
| 656 |
-
// inpL = ggml_add(ctx0,
|
| 657 |
-
// ggml_mul(ctx0,
|
| 658 |
-
// ggml_repeat(ctx0, model.ln_f_g, inpL),
|
| 659 |
-
// inpL),
|
| 660 |
-
// ggml_repeat(ctx0, model.ln_f_b, inpL));
|
| 661 |
-
// }
|
| 662 |
-
//
|
| 663 |
-
// // inpL = WTE * inpL
|
| 664 |
-
// // [ 768, 50257] - model.lm_head
|
| 665 |
-
// // [ 768, N] - inpL
|
| 666 |
-
// inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
|
| 667 |
-
//
|
| 668 |
-
// // logits -> probs
|
| 669 |
-
// //inpL = ggml_soft_max_inplace(ctx0, inpL);
|
| 670 |
-
//
|
| 671 |
-
// // run the computation
|
| 672 |
-
// ggml_build_forward_expand(gf, inpL);
|
| 673 |
-
// ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
|
| 674 |
-
//
|
| 675 |
-
// //if (n_past%100 == 0) {
|
| 676 |
-
// // ggml_graph_print (&gf);
|
| 677 |
-
// // ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
| 678 |
-
// //}
|
| 679 |
-
//
|
| 680 |
-
// //embd_w.resize(n_vocab*N);
|
| 681 |
-
// //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
|
| 682 |
-
//
|
| 683 |
-
// // return result just for the last token
|
| 684 |
-
// embd_w.resize(n_vocab);
|
| 685 |
-
// memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
|
| 686 |
-
//
|
| 687 |
-
// if (mem_per_token == 0) {
|
| 688 |
-
// mem_per_token = ggml_used_mem(ctx0)/N;
|
| 689 |
-
// }
|
| 690 |
-
// //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
|
| 691 |
-
//
|
| 692 |
-
// ggml_free(ctx0);
|
| 693 |
-
//
|
| 694 |
-
// return true;
|
| 695 |
-
//}
|
| 696 |
-
//
|
| 697 |
-
//int main(int argc, char ** argv) {
|
| 698 |
-
// ggml_time_init();
|
| 699 |
-
//
|
| 700 |
-
// const int64_t t_main_start_us = ggml_time_us();
|
| 701 |
-
//
|
| 702 |
-
// gpt_params params;
|
| 703 |
-
// params.model = "ggml-model-gpt-2-774M.bin";
|
| 704 |
-
//
|
| 705 |
-
// if (gpt_params_parse(argc, argv, params) == false) {
|
| 706 |
-
// return 1;
|
| 707 |
-
// }
|
| 708 |
-
//
|
| 709 |
-
// if (params.seed < 0) {
|
| 710 |
-
// params.seed = time(NULL);
|
| 711 |
-
// }
|
| 712 |
-
//
|
| 713 |
-
// printf("%s: seed = %d\n", __func__, params.seed);
|
| 714 |
-
//
|
| 715 |
-
// std::mt19937 rng(params.seed);
|
| 716 |
-
// if (params.prompt.empty()) {
|
| 717 |
-
// params.prompt = gpt_random_prompt(rng);
|
| 718 |
-
// }
|
| 719 |
-
//
|
| 720 |
-
// int64_t t_load_us = 0;
|
| 721 |
-
//
|
| 722 |
-
// gpt_vocab vocab;
|
| 723 |
-
// gpt2_model model;
|
| 724 |
-
//
|
| 725 |
-
// // load the model
|
| 726 |
-
// {
|
| 727 |
-
// const int64_t t_start_us = ggml_time_us();
|
| 728 |
-
//
|
| 729 |
-
// if (!gpt2_model_load(params.model, model, vocab)) {
|
| 730 |
-
// fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
|
| 731 |
-
// return 1;
|
| 732 |
-
// }
|
| 733 |
-
//
|
| 734 |
-
// t_load_us = ggml_time_us() - t_start_us;
|
| 735 |
-
//
|
| 736 |
-
// test_gpt_tokenizer(vocab, params.token_test);
|
| 737 |
-
// }
|
| 738 |
//
|
| 739 |
-
//
|
| 740 |
-
//
|
|
|
|
|
|
|
|
|
|
| 741 |
//
|
| 742 |
-
|
| 743 |
-
|
| 744 |
-
|
| 745 |
-
|
| 746 |
-
|
| 747 |
-
|
| 748 |
-
|
| 749 |
-
|
| 750 |
-
|
| 751 |
-
|
| 752 |
-
|
| 753 |
-
|
| 754 |
-
|
| 755 |
-
|
| 756 |
-
|
| 757 |
-
|
| 758 |
-
|
| 759 |
-
|
| 760 |
-
|
| 761 |
-
|
| 762 |
-
|
| 763 |
-
|
| 764 |
-
//
|
| 765 |
-
|
| 766 |
-
//
|
| 767 |
-
|
| 768 |
-
|
| 769 |
-
|
| 770 |
-
|
| 771 |
-
|
| 772 |
-
|
| 773 |
-
|
| 774 |
-
|
| 775 |
-
|
| 776 |
-
|
| 777 |
-
|
| 778 |
-
|
| 779 |
-
|
| 780 |
-
|
| 781 |
-
|
| 782 |
-
|
| 783 |
-
|
| 784 |
-
|
| 785 |
-
|
| 786 |
-
|
| 787 |
-
|
| 788 |
-
|
| 789 |
-
|
| 790 |
-
|
| 791 |
-
|
| 792 |
-
//
|
| 793 |
-
|
| 794 |
-
|
| 795 |
-
|
| 796 |
-
|
| 797 |
-
|
| 798 |
-
|
| 799 |
-
|
| 800 |
-
|
| 801 |
-
//
|
| 802 |
-
|
| 803 |
-
//
|
| 804 |
-
|
| 805 |
-
|
| 806 |
-
//
|
| 807 |
-
//
|
| 808 |
-
|
| 809 |
-
|
| 810 |
-
|
| 811 |
-
|
| 812 |
-
|
| 813 |
-
|
| 814 |
-
|
| 815 |
-
//
|
| 816 |
-
//
|
| 817 |
-
//
|
| 818 |
-
//
|
| 819 |
-
//
|
| 820 |
-
//
|
| 821 |
-
//
|
| 822 |
-
//
|
| 823 |
-
|
| 824 |
-
|
| 825 |
-
|
| 826 |
-
|
| 827 |
-
|
| 828 |
-
|
| 829 |
-
|
| 830 |
-
|
| 831 |
-
|
| 832 |
-
|
| 833 |
-
//
|
| 834 |
-
|
| 835 |
-
|
| 836 |
-
|
| 837 |
-
|
| 838 |
-
|
| 839 |
-
//
|
| 840 |
-
|
| 841 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#include "ggml.h"
|
| 2 |
+
|
| 3 |
+
#include <cassert>
|
| 4 |
+
#include <cmath>
|
| 5 |
+
#include <cstdio>
|
| 6 |
+
#include <cstring>
|
| 7 |
+
#include <fstream>
|
| 8 |
+
#include <map>
|
| 9 |
+
#include <string>
|
| 10 |
+
#include <vector>
|
| 11 |
+
#include <thread>
|
| 12 |
+
#include <ctime>
|
| 13 |
+
#include <random>
|
| 14 |
+
#include <regex>
|
| 15 |
+
|
| 16 |
+
#if defined(_MSC_VER)
|
| 17 |
+
#pragma warning(disable: 4244 4267) // possible loss of data
|
| 18 |
+
#endif
|
| 19 |
+
|
| 20 |
+
// default hparams (GPT-2 117M)
|
| 21 |
+
struct gpt_hparams {
|
| 22 |
+
int32_t n_vocab = 50257; // Vocabulary size remains the same
|
| 23 |
+
//int32_t n_ctx = 1024; // Maximum context length (sequence length)
|
| 24 |
+
int32_t n_embd = 1024; // Embedding dimensionality
|
| 25 |
+
int32_t n_head = 16; // Number of attention heads
|
| 26 |
+
int32_t n_layer = 24; // Number of transformer layers
|
| 27 |
+
int32_t ftype = 1; // Set to 1 for FP16 precision (optional)
|
| 28 |
+
float eps = 1e-5f; // Small constant for numerical stability
|
| 29 |
+
|
| 30 |
+
int32_t seed = -1; // RNG seed
|
| 31 |
+
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
|
| 32 |
+
int32_t n_predict = 200; // new tokens to predict
|
| 33 |
+
int32_t n_parallel = 1; // number of parallel streams
|
| 34 |
+
int32_t n_batch = 32; // batch size for prompt processing
|
| 35 |
+
int32_t n_ctx = 2048; // context size (this is the KV cache max size)
|
| 36 |
+
int32_t n_gpu_layers = 0; // number of layers to offlload to the GPU
|
| 37 |
+
|
| 38 |
+
bool ignore_eos = false; // ignore EOS token when generating text
|
| 39 |
+
|
| 40 |
+
// sampling parameters
|
| 41 |
+
int32_t top_k = 40;
|
| 42 |
+
float top_p = 0.9f;
|
| 43 |
+
float temp = 0.9f;
|
| 44 |
+
int32_t repeat_last_n = 64;
|
| 45 |
+
float repeat_penalty = 1.00f;
|
| 46 |
+
|
| 47 |
+
std::string model = "ggml-model-gpt-2-774M.bin"; // model path
|
| 48 |
+
std::string prompt = "";
|
| 49 |
+
std::string token_test = "";
|
| 50 |
+
|
| 51 |
+
bool interactive = false;
|
| 52 |
+
int32_t interactive_port = -1;
|
| 53 |
+
};
|
| 54 |
+
|
| 55 |
+
struct gpt_vocab {
|
| 56 |
+
using id = int32_t;
|
| 57 |
+
using token = std::string;
|
| 58 |
+
|
| 59 |
+
std::map<token, id> token_to_id;
|
| 60 |
+
std::map<id, token> id_to_token;
|
| 61 |
+
std::vector<std::string> special_tokens;
|
| 62 |
+
|
| 63 |
+
void add_special_token(const std::string & token);
|
| 64 |
+
};
|
| 65 |
+
|
| 66 |
+
struct gpt_layer {
|
| 67 |
+
// normalization
|
| 68 |
+
struct ggml_tensor * ln_1_g;
|
| 69 |
+
struct ggml_tensor * ln_1_b;
|
| 70 |
+
|
| 71 |
+
struct ggml_tensor * ln_2_g;
|
| 72 |
+
struct ggml_tensor * ln_2_b;
|
| 73 |
+
|
| 74 |
+
// attention
|
| 75 |
+
struct ggml_tensor * c_attn_attn_w;
|
| 76 |
+
struct ggml_tensor * c_attn_attn_b;
|
| 77 |
+
|
| 78 |
+
struct ggml_tensor * c_attn_proj_w;
|
| 79 |
+
struct ggml_tensor * c_attn_proj_b;
|
| 80 |
+
|
| 81 |
+
// mlp
|
| 82 |
+
struct ggml_tensor * c_mlp_fc_w;
|
| 83 |
+
struct ggml_tensor * c_mlp_fc_b;
|
| 84 |
+
|
| 85 |
+
struct ggml_tensor * c_mlp_proj_w;
|
| 86 |
+
struct ggml_tensor * c_mlp_proj_b;
|
| 87 |
+
};
|
| 88 |
+
|
| 89 |
+
struct gpt_model {
|
| 90 |
+
gpt_hparams hparams;
|
| 91 |
+
|
| 92 |
+
// normalization
|
| 93 |
+
struct ggml_tensor * ln_f_g;
|
| 94 |
+
struct ggml_tensor * ln_f_b;
|
| 95 |
+
|
| 96 |
+
struct ggml_tensor * wte; // position embedding
|
| 97 |
+
struct ggml_tensor * wpe; // token embedding
|
| 98 |
+
struct ggml_tensor * lm_head; // language model head
|
| 99 |
+
|
| 100 |
+
std::vector<gpt_layer> layers;
|
| 101 |
+
|
| 102 |
+
// key + value memory
|
| 103 |
+
struct ggml_tensor * memory_k;
|
| 104 |
+
struct ggml_tensor * memory_v;
|
| 105 |
+
|
| 106 |
+
//
|
| 107 |
+
struct ggml_context * ctx_w;
|
| 108 |
+
std::map<std::string, struct ggml_tensor *> tensors;
|
| 109 |
+
};
|
| 110 |
+
|
| 111 |
+
// load the model's weights from a file
|
| 112 |
+
bool gpt_model_load(const std::string & fname, gpt_model & model, gpt_vocab & vocab) {
|
| 113 |
+
printf("%s: loading model from '%s'\n", __func__, fname.c_str());
|
| 114 |
+
|
| 115 |
+
auto fin = std::ifstream(fname, std::ios::binary);
|
| 116 |
+
if (!fin) {
|
| 117 |
+
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
|
| 118 |
+
return false;
|
| 119 |
+
}
|
| 120 |
+
|
| 121 |
+
// verify magic
|
| 122 |
+
{
|
| 123 |
+
uint32_t magic;
|
| 124 |
+
fin.read((char *) &magic, sizeof(magic));
|
| 125 |
+
if (magic != GGML_FILE_MAGIC) {
|
| 126 |
+
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
|
| 127 |
+
return false;
|
| 128 |
+
}
|
| 129 |
+
}
|
| 130 |
+
|
| 131 |
+
// load hparams
|
| 132 |
+
{
|
| 133 |
+
auto & hparams = model.hparams;
|
| 134 |
+
|
| 135 |
+
fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
| 136 |
+
fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
|
| 137 |
+
fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
|
| 138 |
+
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
|
| 139 |
+
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
| 140 |
+
fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
|
| 141 |
+
|
| 142 |
+
const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
|
| 143 |
+
|
| 144 |
+
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
|
| 145 |
+
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
|
| 146 |
+
printf("%s: n_embd = %d\n", __func__, hparams.n_embd);
|
| 147 |
+
printf("%s: n_head = %d\n", __func__, hparams.n_head);
|
| 148 |
+
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
|
| 149 |
+
printf("%s: ftype = %d\n", __func__, hparams.ftype);
|
| 150 |
+
printf("%s: qntvr = %d\n", __func__, qntvr);
|
| 151 |
+
|
| 152 |
+
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
|
| 153 |
+
}
|
| 154 |
+
|
| 155 |
+
// load vocab
|
| 156 |
+
{
|
| 157 |
+
int32_t n_vocab = 0;
|
| 158 |
+
fin.read((char *) &n_vocab, sizeof(n_vocab));
|
| 159 |
+
|
| 160 |
+
if (n_vocab != model.hparams.n_vocab) {
|
| 161 |
+
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
|
| 162 |
+
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
|
| 163 |
+
return false;
|
| 164 |
+
}
|
| 165 |
+
|
| 166 |
+
std::string word;
|
| 167 |
+
std::vector<char> buf(128);
|
| 168 |
+
|
| 169 |
+
for (int i = 0; i < n_vocab; i++) {
|
| 170 |
+
uint32_t len;
|
| 171 |
+
fin.read((char *) &len, sizeof(len));
|
| 172 |
+
|
| 173 |
+
buf.resize(len);
|
| 174 |
+
fin.read((char *) buf.data(), len);
|
| 175 |
+
word.assign(buf.data(), len);
|
| 176 |
+
|
| 177 |
+
vocab.token_to_id[word] = i;
|
| 178 |
+
vocab.id_to_token[i] = word;
|
| 179 |
+
}
|
| 180 |
+
}
|
| 181 |
+
|
| 182 |
+
// for the big tensors, we have the option to store the data in 16-bit floats or quantized
|
| 183 |
+
// in order to save memory and also to speed up the computation
|
| 184 |
+
ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
|
| 185 |
+
if (wtype == GGML_TYPE_COUNT) {
|
| 186 |
+
fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
|
| 187 |
+
__func__, fname.c_str(), model.hparams.ftype);
|
| 188 |
+
return false;
|
| 189 |
+
}
|
| 190 |
+
|
| 191 |
+
auto & ctx = model.ctx_w;
|
| 192 |
+
|
| 193 |
+
size_t ctx_size = 0;
|
| 194 |
+
|
| 195 |
+
{
|
| 196 |
+
const auto & hparams = model.hparams;
|
| 197 |
+
|
| 198 |
+
const int n_embd = hparams.n_embd;
|
| 199 |
+
const int n_layer = hparams.n_layer;
|
| 200 |
+
const int n_ctx = hparams.n_ctx;
|
| 201 |
+
const int n_vocab = hparams.n_vocab;
|
| 202 |
+
|
| 203 |
+
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
|
| 204 |
+
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
|
| 205 |
+
|
| 206 |
+
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // wte
|
| 207 |
+
ctx_size += ggml_row_size(GGML_TYPE_F32, n_ctx*n_embd); // wpe
|
| 208 |
+
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // lm_head
|
| 209 |
+
|
| 210 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
|
| 211 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
|
| 212 |
+
|
| 213 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
|
| 214 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
|
| 215 |
+
|
| 216 |
+
ctx_size += n_layer*(ggml_row_size(wtype, 3*n_embd*n_embd)); // c_attn_attn_w
|
| 217 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd)); // c_attn_attn_b
|
| 218 |
+
|
| 219 |
+
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_proj_w
|
| 220 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // c_attn_proj_b
|
| 221 |
+
|
| 222 |
+
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_fc_w
|
| 223 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_fc_b
|
| 224 |
+
|
| 225 |
+
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_proj_w
|
| 226 |
+
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_proj_b
|
| 227 |
+
|
| 228 |
+
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
|
| 229 |
+
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
|
| 230 |
+
|
| 231 |
+
ctx_size += (6 + 12*n_layer)*512; // object overhead
|
| 232 |
+
|
| 233 |
+
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
|
| 234 |
+
printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
|
| 235 |
+
}
|
| 236 |
+
|
| 237 |
+
// create the ggml context
|
| 238 |
+
{
|
| 239 |
+
struct ggml_init_params params = {
|
| 240 |
+
/*.mem_size =*/ ctx_size,
|
| 241 |
+
/*.mem_buffer =*/ NULL,
|
| 242 |
+
/*.no_alloc =*/ false,
|
| 243 |
+
};
|
| 244 |
+
|
| 245 |
+
model.ctx_w = ggml_init(params);
|
| 246 |
+
if (!model.ctx_w) {
|
| 247 |
+
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
|
| 248 |
+
return false;
|
| 249 |
+
}
|
| 250 |
+
}
|
| 251 |
+
|
| 252 |
+
// prepare memory for the weights
|
| 253 |
+
{
|
| 254 |
+
const auto & hparams = model.hparams;
|
| 255 |
+
|
| 256 |
+
const int n_embd = hparams.n_embd;
|
| 257 |
+
const int n_layer = hparams.n_layer;
|
| 258 |
+
const int n_ctx = hparams.n_ctx;
|
| 259 |
+
const int n_vocab = hparams.n_vocab;
|
| 260 |
+
|
| 261 |
+
model.layers.resize(n_layer);
|
| 262 |
+
|
| 263 |
+
model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 264 |
+
model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 265 |
+
|
| 266 |
+
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
| 267 |
+
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
|
| 268 |
+
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
| 269 |
+
|
| 270 |
+
// map by name
|
| 271 |
+
model.tensors["model/ln_f/g"] = model.ln_f_g;
|
| 272 |
+
model.tensors["model/ln_f/b"] = model.ln_f_b;
|
| 273 |
+
|
| 274 |
+
model.tensors["model/wte"] = model.wte;
|
| 275 |
+
model.tensors["model/wpe"] = model.wpe;
|
| 276 |
+
model.tensors["model/lm_head"] = model.lm_head;
|
| 277 |
+
|
| 278 |
+
for (int i = 0; i < n_layer; ++i) {
|
| 279 |
+
auto & layer = model.layers[i];
|
| 280 |
+
|
| 281 |
+
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 282 |
+
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 283 |
+
|
| 284 |
+
layer.ln_2_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 285 |
+
layer.ln_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 286 |
+
|
| 287 |
+
layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 3*n_embd);
|
| 288 |
+
layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
|
| 289 |
+
|
| 290 |
+
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
| 291 |
+
layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 292 |
+
|
| 293 |
+
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
|
| 294 |
+
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
|
| 295 |
+
|
| 296 |
+
layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
|
| 297 |
+
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
| 298 |
+
|
| 299 |
+
// map by name
|
| 300 |
+
model.tensors["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
|
| 301 |
+
model.tensors["model/h" + std::to_string(i) + "/ln_1/b"] = layer.ln_1_b;
|
| 302 |
+
|
| 303 |
+
model.tensors["model/h" + std::to_string(i) + "/ln_2/g"] = layer.ln_2_g;
|
| 304 |
+
model.tensors["model/h" + std::to_string(i) + "/ln_2/b"] = layer.ln_2_b;
|
| 305 |
+
|
| 306 |
+
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
|
| 307 |
+
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
|
| 308 |
+
|
| 309 |
+
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
|
| 310 |
+
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
|
| 311 |
+
|
| 312 |
+
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w;
|
| 313 |
+
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b;
|
| 314 |
+
|
| 315 |
+
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w;
|
| 316 |
+
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = layer.c_mlp_proj_b;
|
| 317 |
+
}
|
| 318 |
+
}
|
| 319 |
+
|
| 320 |
+
// key + value memory
|
| 321 |
+
{
|
| 322 |
+
const auto & hparams = model.hparams;
|
| 323 |
+
|
| 324 |
+
const int n_embd = hparams.n_embd;
|
| 325 |
+
const int n_layer = hparams.n_layer;
|
| 326 |
+
const int n_ctx = hparams.n_ctx;
|
| 327 |
+
|
| 328 |
+
const int n_mem = n_layer*n_ctx;
|
| 329 |
+
const int n_elements = n_embd*n_mem;
|
| 330 |
+
|
| 331 |
+
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
| 332 |
+
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
| 333 |
+
|
| 334 |
+
const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
|
| 335 |
+
|
| 336 |
+
printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
|
| 337 |
+
}
|
| 338 |
+
|
| 339 |
+
// load weights
|
| 340 |
+
{
|
| 341 |
+
size_t total_size = 0;
|
| 342 |
+
|
| 343 |
+
bool has_lm_head = false;
|
| 344 |
+
|
| 345 |
+
while (true) {
|
| 346 |
+
int32_t n_dims;
|
| 347 |
+
int32_t length;
|
| 348 |
+
int32_t ttype;
|
| 349 |
+
|
| 350 |
+
fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
|
| 351 |
+
fin.read(reinterpret_cast<char *>(&length), sizeof(length));
|
| 352 |
+
fin.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
|
| 353 |
+
|
| 354 |
+
if (fin.eof()) {
|
| 355 |
+
break;
|
| 356 |
+
}
|
| 357 |
+
|
| 358 |
+
int32_t nelements = 1;
|
| 359 |
+
int32_t ne[2] = { 1, 1 };
|
| 360 |
+
for (int i = 0; i < n_dims; ++i) {
|
| 361 |
+
fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
|
| 362 |
+
nelements *= ne[i];
|
| 363 |
+
}
|
| 364 |
+
|
| 365 |
+
std::string name(length, 0);
|
| 366 |
+
fin.read(&name[0], length);
|
| 367 |
+
|
| 368 |
+
if (model.tensors.find(name) == model.tensors.end()) {
|
| 369 |
+
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
|
| 370 |
+
return false;
|
| 371 |
+
}
|
| 372 |
+
|
| 373 |
+
auto tensor = model.tensors[name];
|
| 374 |
+
if (ggml_nelements(tensor) != nelements) {
|
| 375 |
+
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
|
| 376 |
+
return false;
|
| 377 |
+
}
|
| 378 |
+
|
| 379 |
+
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
|
| 380 |
+
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
|
| 381 |
+
__func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
|
| 382 |
+
return false;
|
| 383 |
+
}
|
| 384 |
+
|
| 385 |
+
// for debugging
|
| 386 |
+
if (0) {
|
| 387 |
+
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
|
| 388 |
+
}
|
| 389 |
+
|
| 390 |
+
const size_t bpe = ggml_type_size(ggml_type(ttype));
|
| 391 |
+
|
| 392 |
+
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
|
| 393 |
+
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
|
| 394 |
+
__func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
|
| 395 |
+
return false;
|
| 396 |
+
}
|
| 397 |
+
|
| 398 |
+
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
| 399 |
+
|
| 400 |
+
// GPT-2 models share the WTE tensor as the LM head
|
| 401 |
+
if (name == "model/wte" && has_lm_head == false) {
|
| 402 |
+
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
|
| 403 |
+
}
|
| 404 |
+
|
| 405 |
+
if (name == "model/lm_head") {
|
| 406 |
+
has_lm_head = true;
|
| 407 |
+
}
|
| 408 |
+
|
| 409 |
+
total_size += ggml_nbytes(tensor);
|
| 410 |
+
}
|
| 411 |
+
|
| 412 |
+
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
|
| 413 |
+
}
|
| 414 |
+
|
| 415 |
+
fin.close();
|
| 416 |
+
|
| 417 |
+
return true;
|
| 418 |
+
}
|
| 419 |
+
|
| 420 |
+
void gpt_split_words(std::string str, std::vector<std::string>& words) {
|
| 421 |
+
const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
|
| 422 |
+
const std::regex re(pattern);
|
| 423 |
+
std::smatch m;
|
| 424 |
+
|
| 425 |
+
while (std::regex_search(str, m, re)) {
|
| 426 |
+
for (auto x : m) {
|
| 427 |
+
words.push_back(x);
|
| 428 |
+
}
|
| 429 |
+
str = m.suffix();
|
| 430 |
+
}
|
| 431 |
+
}
|
| 432 |
+
|
| 433 |
+
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
|
| 434 |
+
std::vector<std::string> words;
|
| 435 |
+
|
| 436 |
+
// first split the text into words
|
| 437 |
+
{
|
| 438 |
+
std::string str = text;
|
| 439 |
+
|
| 440 |
+
// Generate the subpattern from the special_tokens vector if it's not empty
|
| 441 |
+
if (!vocab.special_tokens.empty()) {
|
| 442 |
+
const std::regex escape(R"([\[\\\^\$\.\|\?\*\+\(\)\{\}])");
|
| 443 |
+
std::string special_tokens_subpattern;
|
| 444 |
+
for (const auto & token : vocab.special_tokens) {
|
| 445 |
+
if (!special_tokens_subpattern.empty()) {
|
| 446 |
+
special_tokens_subpattern += "|";
|
| 447 |
+
}
|
| 448 |
+
special_tokens_subpattern += std::regex_replace(token, escape, R"(\$&)");
|
| 449 |
+
}
|
| 450 |
+
|
| 451 |
+
std::regex re(special_tokens_subpattern);
|
| 452 |
+
std::smatch m;
|
| 453 |
+
// Split the text by special tokens.
|
| 454 |
+
while (std::regex_search(str, m, re)) {
|
| 455 |
+
// Split the substrings in-between special tokens into words.
|
| 456 |
+
gpt_split_words(m.prefix(), words);
|
| 457 |
+
// Add matched special tokens as words.
|
| 458 |
+
for (auto x : m) {
|
| 459 |
+
words.push_back(x);
|
| 460 |
+
}
|
| 461 |
+
str = m.suffix();
|
| 462 |
+
}
|
| 463 |
+
// Remaining text without special tokens will be handled below.
|
| 464 |
+
}
|
| 465 |
+
|
| 466 |
+
gpt_split_words(str, words);
|
| 467 |
+
}
|
| 468 |
+
|
| 469 |
+
// find the longest token that forms each word in words:
|
| 470 |
+
std::vector<gpt_vocab::id> tokens;
|
| 471 |
+
for (const auto & word : words) {
|
| 472 |
+
for (int i = 0; i < (int) word.size(); ){
|
| 473 |
+
for (int j = word.size() - 1; j >= i; j--){
|
| 474 |
+
auto cand = word.substr(i, j-i+1);
|
| 475 |
+
auto it = vocab.token_to_id.find(cand);
|
| 476 |
+
if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab
|
| 477 |
+
tokens.push_back(it->second);
|
| 478 |
+
i = j + 1;
|
| 479 |
+
break;
|
| 480 |
+
}
|
| 481 |
+
else if (j == i){ // word.substr(i, 1) has no matching
|
| 482 |
+
fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
|
| 483 |
+
i++;
|
| 484 |
+
}
|
| 485 |
+
}
|
| 486 |
+
}
|
| 487 |
+
}
|
| 488 |
+
|
| 489 |
+
return tokens;
|
| 490 |
+
}
|
| 491 |
+
|
| 492 |
+
static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
|
| 493 |
+
std::vector<gpt_vocab::id> output;
|
| 494 |
+
std::stringstream ss(input);
|
| 495 |
+
std::string token;
|
| 496 |
+
|
| 497 |
+
while (std::getline(ss, token, delimiter)) {
|
| 498 |
+
output.push_back(std::stoi(token));
|
| 499 |
+
}
|
| 500 |
+
|
| 501 |
+
return output;
|
| 502 |
+
}
|
| 503 |
+
|
| 504 |
+
static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
|
| 505 |
+
if (fpath_test.empty()){
|
| 506 |
+
fprintf(stderr, "%s : No test file found.\n", __func__);
|
| 507 |
+
return std::map<std::string, std::vector<gpt_vocab::id>>();
|
| 508 |
+
}
|
| 509 |
+
|
| 510 |
+
std::map<std::string, std::vector<gpt_vocab::id>> tests;
|
| 511 |
+
|
| 512 |
+
auto fin = std::ifstream(fpath_test, std::ios_base::in);
|
| 513 |
+
const char * delimeter = " => ";
|
| 514 |
+
const char del_tok = ',';
|
| 515 |
+
std::string line;
|
| 516 |
+
while (std::getline(fin, line)) {
|
| 517 |
+
size_t delimiterPos = line.find(delimeter);
|
| 518 |
+
if (delimiterPos != std::string::npos) {
|
| 519 |
+
std::string text = line.substr(0, delimiterPos);
|
| 520 |
+
std::string s_tokens = line.substr(delimiterPos + std::strlen(delimeter));
|
| 521 |
+
tests[text] = parse_tokens_from_string(s_tokens, del_tok);
|
| 522 |
+
}
|
| 523 |
+
}
|
| 524 |
+
return tests;
|
| 525 |
+
}
|
| 526 |
+
|
| 527 |
+
void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test){
|
| 528 |
+
std::map<std::string, std::vector<gpt_vocab::id>> tests = extract_tests_from_file(fpath_test);
|
| 529 |
+
|
| 530 |
+
size_t n_fails = 0;
|
| 531 |
+
|
| 532 |
+
for (const auto & test : tests) {
|
| 533 |
+
std::vector<gpt_vocab::id> tokens = gpt_tokenize(vocab, test.first);
|
| 534 |
+
|
| 535 |
+
if (tokens != test.second){
|
| 536 |
+
n_fails++;
|
| 537 |
+
|
| 538 |
+
// print out failure cases
|
| 539 |
+
fprintf(stderr, "%s : failed test: '%s'\n", __func__, test.first.c_str());
|
| 540 |
+
fprintf(stderr, "%s : tokens in hf: ", __func__);
|
| 541 |
+
for (const auto & t : test.second) {
|
| 542 |
+
fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
|
| 543 |
+
}
|
| 544 |
+
fprintf(stderr, "\n");
|
| 545 |
+
fprintf(stderr, "%s : tokens in ggml: ", __func__);
|
| 546 |
+
for (const auto & t : tokens) {
|
| 547 |
+
fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
|
| 548 |
+
}
|
| 549 |
+
fprintf(stderr, "\n");
|
| 550 |
+
}
|
| 551 |
+
}
|
| 552 |
+
|
| 553 |
+
fprintf(stderr, "%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
|
| 554 |
+
}
|
| 555 |
+
|
| 556 |
+
gpt_vocab::id gpt_sample_top_k_top_p(
|
| 557 |
+
const gpt_vocab & vocab,
|
| 558 |
+
const float * logits,
|
| 559 |
+
int top_k,
|
| 560 |
+
double top_p,
|
| 561 |
+
double temp,
|
| 562 |
+
std::mt19937 & rng) {
|
| 563 |
+
int n_logits = vocab.id_to_token.size();
|
| 564 |
+
|
| 565 |
+
std::vector<std::pair<double, gpt_vocab::id>> logits_id;
|
| 566 |
+
logits_id.reserve(n_logits);
|
| 567 |
+
|
| 568 |
+
{
|
| 569 |
+
const double scale = 1.0/temp;
|
| 570 |
+
for (int i = 0; i < n_logits; ++i) {
|
| 571 |
+
logits_id.push_back(std::make_pair(logits[i]*scale, i));
|
| 572 |
+
}
|
| 573 |
+
}
|
| 574 |
+
|
| 575 |
+
// find the top K tokens
|
| 576 |
+
std::partial_sort(
|
| 577 |
+
logits_id.begin(),
|
| 578 |
+
logits_id.begin() + top_k, logits_id.end(),
|
| 579 |
+
[](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
|
| 580 |
+
return a.first > b.first;
|
| 581 |
+
});
|
| 582 |
+
|
| 583 |
+
logits_id.resize(top_k);
|
| 584 |
+
|
| 585 |
+
double maxl = -INFINITY;
|
| 586 |
+
for (const auto & kv : logits_id) {
|
| 587 |
+
maxl = std::max(maxl, kv.first);
|
| 588 |
+
}
|
| 589 |
+
|
| 590 |
+
// compute probs for the top K tokens
|
| 591 |
+
std::vector<double> probs;
|
| 592 |
+
probs.reserve(logits_id.size());
|
| 593 |
+
|
| 594 |
+
double sum = 0.0;
|
| 595 |
+
for (const auto & kv : logits_id) {
|
| 596 |
+
double p = exp(kv.first - maxl);
|
| 597 |
+
probs.push_back(p);
|
| 598 |
+
sum += p;
|
| 599 |
+
}
|
| 600 |
+
|
| 601 |
+
// normalize the probs
|
| 602 |
+
for (auto & p : probs) {
|
| 603 |
+
p /= sum;
|
| 604 |
+
}
|
| 605 |
+
|
| 606 |
+
if (top_p < 1.0f) {
|
| 607 |
+
double cumsum = 0.0f;
|
| 608 |
+
for (int i = 0; i < top_k; i++) {
|
| 609 |
+
cumsum += probs[i];
|
| 610 |
+
if (cumsum >= top_p) {
|
| 611 |
+
top_k = i + 1;
|
| 612 |
+
probs.resize(top_k);
|
| 613 |
+
logits_id.resize(top_k);
|
| 614 |
+
break;
|
| 615 |
+
}
|
| 616 |
+
}
|
| 617 |
+
|
| 618 |
+
cumsum = 1.0/cumsum;
|
| 619 |
+
for (int i = 0; i < (int) probs.size(); i++) {
|
| 620 |
+
probs[i] *= cumsum;
|
| 621 |
+
}
|
| 622 |
+
}
|
| 623 |
+
|
| 624 |
+
//printf("\n");
|
| 625 |
+
//for (int i = 0; i < (int) probs.size(); i++) {
|
| 626 |
+
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
|
| 627 |
+
//}
|
| 628 |
+
//exit(0);
|
| 629 |
+
|
| 630 |
+
std::discrete_distribution<> dist(probs.begin(), probs.end());
|
| 631 |
+
int idx = dist(rng);
|
| 632 |
+
|
| 633 |
+
return logits_id[idx].second;
|
| 634 |
+
}
|
| 635 |
+
|
| 636 |
+
// evaluate the transformer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 637 |
//
|
| 638 |
+
// - model: the model
|
| 639 |
+
// - n_threads: number of threads to use
|
| 640 |
+
// - n_past: the context size so far
|
| 641 |
+
// - embd_inp: the embeddings of the tokens in the context
|
| 642 |
+
// - embd_w: the predicted logits for the next token
|
| 643 |
//
|
| 644 |
+
bool gpt_eval(
|
| 645 |
+
const gpt_model & model,
|
| 646 |
+
const int n_threads,
|
| 647 |
+
const int n_past,
|
| 648 |
+
const std::vector<gpt_vocab::id> & embd_inp,
|
| 649 |
+
std::vector<float> & embd_w,
|
| 650 |
+
size_t & mem_per_token) {
|
| 651 |
+
const int N = embd_inp.size();
|
| 652 |
+
|
| 653 |
+
const auto & hparams = model.hparams;
|
| 654 |
+
|
| 655 |
+
const int n_embd = hparams.n_embd;
|
| 656 |
+
const int n_layer = hparams.n_layer;
|
| 657 |
+
const int n_ctx = hparams.n_ctx;
|
| 658 |
+
const int n_head = hparams.n_head;
|
| 659 |
+
const int n_vocab = hparams.n_vocab;
|
| 660 |
+
|
| 661 |
+
static size_t buf_size = 256u*1024*1024;
|
| 662 |
+
static void * buf = malloc(buf_size);
|
| 663 |
+
|
| 664 |
+
if (mem_per_token > 0 && mem_per_token*N > buf_size) {
|
| 665 |
+
const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
|
| 666 |
+
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
|
| 667 |
+
|
| 668 |
+
// reallocate
|
| 669 |
+
buf_size = buf_size_new;
|
| 670 |
+
buf = realloc(buf, buf_size);
|
| 671 |
+
if (buf == nullptr) {
|
| 672 |
+
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
|
| 673 |
+
return false;
|
| 674 |
+
}
|
| 675 |
+
}
|
| 676 |
+
|
| 677 |
+
struct ggml_init_params params = {
|
| 678 |
+
/*.mem_size =*/ buf_size,
|
| 679 |
+
/*.mem_buffer =*/ buf,
|
| 680 |
+
/*.no_alloc =*/ false,
|
| 681 |
+
};
|
| 682 |
+
|
| 683 |
+
struct ggml_context * ctx0 = ggml_init(params);
|
| 684 |
+
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
|
| 685 |
+
|
| 686 |
+
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
| 687 |
+
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
|
| 688 |
+
|
| 689 |
+
struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
| 690 |
+
for (int i = 0; i < N; ++i) {
|
| 691 |
+
((int32_t *) position->data)[i] = n_past + i;
|
| 692 |
+
}
|
| 693 |
+
|
| 694 |
+
// wte + wpe
|
| 695 |
+
struct ggml_tensor * inpL =
|
| 696 |
+
ggml_add(ctx0,
|
| 697 |
+
ggml_get_rows(ctx0, model.wte, embd),
|
| 698 |
+
ggml_get_rows(ctx0, model.wpe, position));
|
| 699 |
+
|
| 700 |
+
for (int il = 0; il < n_layer; ++il) {
|
| 701 |
+
struct ggml_tensor * cur;
|
| 702 |
+
|
| 703 |
+
// norm
|
| 704 |
+
{
|
| 705 |
+
// [ 768, N]
|
| 706 |
+
cur = ggml_norm(ctx0, inpL, hparams.eps);
|
| 707 |
+
|
| 708 |
+
// cur = ln_1_g*cur + ln_1_b
|
| 709 |
+
// [ 768, N]
|
| 710 |
+
cur = ggml_add(ctx0,
|
| 711 |
+
ggml_mul(ctx0,
|
| 712 |
+
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
|
| 713 |
+
cur),
|
| 714 |
+
ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
|
| 715 |
+
}
|
| 716 |
+
|
| 717 |
+
// attn
|
| 718 |
+
// [2304, 768] - model.layers[il].c_attn_attn_w
|
| 719 |
+
// [2304, 1] - model.layers[il].c_attn_attn_b
|
| 720 |
+
// [ 768, N] - cur (in)
|
| 721 |
+
// [2304, N] - cur (out)
|
| 722 |
+
//
|
| 723 |
+
// cur = attn_w*cur + attn_b
|
| 724 |
+
// [2304, N]
|
| 725 |
+
{
|
| 726 |
+
cur = ggml_mul_mat(ctx0,
|
| 727 |
+
model.layers[il].c_attn_attn_w,
|
| 728 |
+
cur);
|
| 729 |
+
|
| 730 |
+
cur = ggml_add(ctx0,
|
| 731 |
+
ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
|
| 732 |
+
cur);
|
| 733 |
+
}
|
| 734 |
+
|
| 735 |
+
// self-attention
|
| 736 |
+
{
|
| 737 |
+
struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
|
| 738 |
+
struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
|
| 739 |
+
struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
|
| 740 |
+
|
| 741 |
+
// store key and value to memory
|
| 742 |
+
if (N >= 1) {
|
| 743 |
+
struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
|
| 744 |
+
struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
|
| 745 |
+
|
| 746 |
+
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
|
| 747 |
+
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
|
| 748 |
+
}
|
| 749 |
+
|
| 750 |
+
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
|
| 751 |
+
// [64, N, 12]
|
| 752 |
+
struct ggml_tensor * Q =
|
| 753 |
+
ggml_permute(ctx0,
|
| 754 |
+
ggml_cpy(ctx0,
|
| 755 |
+
Qcur,
|
| 756 |
+
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
|
| 757 |
+
0, 2, 1, 3);
|
| 758 |
+
|
| 759 |
+
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
| 760 |
+
// [64, n_past + N, 12]
|
| 761 |
+
struct ggml_tensor * K =
|
| 762 |
+
ggml_permute(ctx0,
|
| 763 |
+
ggml_reshape_3d(ctx0,
|
| 764 |
+
ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
|
| 765 |
+
n_embd/n_head, n_head, n_past + N),
|
| 766 |
+
0, 2, 1, 3);
|
| 767 |
+
|
| 768 |
+
// GG: flash attention
|
| 769 |
+
//struct ggml_tensor * V =
|
| 770 |
+
// ggml_cpy(ctx0,
|
| 771 |
+
// ggml_permute(ctx0,
|
| 772 |
+
// ggml_reshape_3d(ctx0,
|
| 773 |
+
// ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
| 774 |
+
// n_embd/n_head, n_head, n_past + N),
|
| 775 |
+
// 1, 2, 0, 3),
|
| 776 |
+
// ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
|
| 777 |
+
|
| 778 |
+
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
|
| 779 |
+
|
| 780 |
+
// K * Q
|
| 781 |
+
// [n_past + N, N, 12]
|
| 782 |
+
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
| 783 |
+
|
| 784 |
+
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
| 785 |
+
// [n_past + N, N, 12]
|
| 786 |
+
struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, 1.0f/sqrt(float(n_embd)/n_head));
|
| 787 |
+
|
| 788 |
+
// KQ_masked = mask_past(KQ_scaled)
|
| 789 |
+
// [n_past + N, N, 12]
|
| 790 |
+
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
|
| 791 |
+
|
| 792 |
+
// KQ = soft_max(KQ_masked)
|
| 793 |
+
// [n_past + N, N, 12]
|
| 794 |
+
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
|
| 795 |
+
|
| 796 |
+
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
|
| 797 |
+
// [n_past + N, 64, 12]
|
| 798 |
+
struct ggml_tensor * V_trans =
|
| 799 |
+
ggml_cpy(ctx0,
|
| 800 |
+
ggml_permute(ctx0,
|
| 801 |
+
ggml_reshape_3d(ctx0,
|
| 802 |
+
ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
| 803 |
+
n_embd/n_head, n_head, n_past + N),
|
| 804 |
+
1, 2, 0, 3),
|
| 805 |
+
ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
|
| 806 |
+
|
| 807 |
+
// KQV = transpose(V) * KQ_soft_max
|
| 808 |
+
// [64, N, 12]
|
| 809 |
+
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
|
| 810 |
+
|
| 811 |
+
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
| 812 |
+
// [64, 12, N]
|
| 813 |
+
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
| 814 |
+
|
| 815 |
+
// cur = KQV_merged.contiguous().view(n_embd, N)
|
| 816 |
+
// [768, N]
|
| 817 |
+
cur = ggml_cpy(ctx0,
|
| 818 |
+
KQV_merged,
|
| 819 |
+
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
|
| 820 |
+
}
|
| 821 |
+
|
| 822 |
+
// projection
|
| 823 |
+
// [ 768, 768] - model.layers[il].c_attn_proj_w
|
| 824 |
+
// [ 768, 1] - model.layers[il].c_attn_proj_b
|
| 825 |
+
// [ 768, N] - cur (in)
|
| 826 |
+
// [ 768, N] - cur (out)
|
| 827 |
+
//
|
| 828 |
+
// cur = proj_w*cur + proj_b
|
| 829 |
+
// [768, N]
|
| 830 |
+
{
|
| 831 |
+
cur = ggml_mul_mat(ctx0,
|
| 832 |
+
model.layers[il].c_attn_proj_w,
|
| 833 |
+
cur);
|
| 834 |
+
|
| 835 |
+
cur = ggml_add(ctx0,
|
| 836 |
+
ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
|
| 837 |
+
cur);
|
| 838 |
+
}
|
| 839 |
+
|
| 840 |
+
// add the input
|
| 841 |
+
cur = ggml_add(ctx0, cur, inpL);
|
| 842 |
+
|
| 843 |
+
struct ggml_tensor * inpFF = cur;
|
| 844 |
+
|
| 845 |
+
// feed-forward network
|
| 846 |
+
{
|
| 847 |
+
// norm
|
| 848 |
+
{
|
| 849 |
+
cur = ggml_norm(ctx0, inpFF, hparams.eps);
|
| 850 |
+
|
| 851 |
+
// cur = ln_2_g*cur + ln_2_b
|
| 852 |
+
// [ 768, N]
|
| 853 |
+
cur = ggml_add(ctx0,
|
| 854 |
+
ggml_mul(ctx0,
|
| 855 |
+
ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
|
| 856 |
+
cur),
|
| 857 |
+
ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
|
| 858 |
+
}
|
| 859 |
+
|
| 860 |
+
// fully connected
|
| 861 |
+
// [3072, 768] - model.layers[il].c_mlp_fc_w
|
| 862 |
+
// [3072, 1] - model.layers[il].c_mlp_fc_b
|
| 863 |
+
// [ 768, N] - cur (in)
|
| 864 |
+
// [3072, N] - cur (out)
|
| 865 |
+
//
|
| 866 |
+
// cur = fc_w*cur + fc_b
|
| 867 |
+
// [3072, N]
|
| 868 |
+
cur = ggml_mul_mat(ctx0,
|
| 869 |
+
model.layers[il].c_mlp_fc_w,
|
| 870 |
+
cur);
|
| 871 |
+
|
| 872 |
+
cur = ggml_add(ctx0,
|
| 873 |
+
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
|
| 874 |
+
cur);
|
| 875 |
+
|
| 876 |
+
// GELU activation
|
| 877 |
+
// [3072, N]
|
| 878 |
+
cur = ggml_gelu(ctx0, cur);
|
| 879 |
+
|
| 880 |
+
// projection
|
| 881 |
+
// [ 768, 3072] - model.layers[il].c_mlp_proj_w
|
| 882 |
+
// [ 768, 1] - model.layers[il].c_mlp_proj_b
|
| 883 |
+
// [3072, N] - cur (in)
|
| 884 |
+
// [ 768, N] - cur (out)
|
| 885 |
+
//
|
| 886 |
+
// cur = proj_w*cur + proj_b
|
| 887 |
+
// [768, N]
|
| 888 |
+
cur = ggml_mul_mat(ctx0,
|
| 889 |
+
model.layers[il].c_mlp_proj_w,
|
| 890 |
+
cur);
|
| 891 |
+
|
| 892 |
+
cur = ggml_add(ctx0,
|
| 893 |
+
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
|
| 894 |
+
cur);
|
| 895 |
+
}
|
| 896 |
+
|
| 897 |
+
// input for next layer
|
| 898 |
+
inpL = ggml_add(ctx0, cur, inpFF);
|
| 899 |
+
}
|
| 900 |
+
|
| 901 |
+
// norm
|
| 902 |
+
{
|
| 903 |
+
// [ 768, N]
|
| 904 |
+
inpL = ggml_norm(ctx0, inpL, hparams.eps);
|
| 905 |
+
|
| 906 |
+
// inpL = ln_f_g*inpL + ln_f_b
|
| 907 |
+
// [ 768, N]
|
| 908 |
+
inpL = ggml_add(ctx0,
|
| 909 |
+
ggml_mul(ctx0,
|
| 910 |
+
ggml_repeat(ctx0, model.ln_f_g, inpL),
|
| 911 |
+
inpL),
|
| 912 |
+
ggml_repeat(ctx0, model.ln_f_b, inpL));
|
| 913 |
+
}
|
| 914 |
+
|
| 915 |
+
// inpL = WTE * inpL
|
| 916 |
+
// [ 768, 50257] - model.lm_head
|
| 917 |
+
// [ 768, N] - inpL
|
| 918 |
+
inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
|
| 919 |
+
|
| 920 |
+
// logits -> probs
|
| 921 |
+
//inpL = ggml_soft_max_inplace(ctx0, inpL);
|
| 922 |
+
|
| 923 |
+
// run the computation
|
| 924 |
+
ggml_build_forward_expand(gf, inpL);
|
| 925 |
+
ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
|
| 926 |
+
|
| 927 |
+
//if (n_past%100 == 0) {
|
| 928 |
+
// ggml_graph_print (&gf);
|
| 929 |
+
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
| 930 |
+
//}
|
| 931 |
+
|
| 932 |
+
//embd_w.resize(n_vocab*N);
|
| 933 |
+
//memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
|
| 934 |
+
|
| 935 |
+
// return result just for the last token
|
| 936 |
+
embd_w.resize(n_vocab);
|
| 937 |
+
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
|
| 938 |
+
|
| 939 |
+
if (mem_per_token == 0) {
|
| 940 |
+
mem_per_token = ggml_used_mem(ctx0)/N;
|
| 941 |
+
}
|
| 942 |
+
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
|
| 943 |
+
|
| 944 |
+
ggml_free(ctx0);
|
| 945 |
+
|
| 946 |
+
return true;
|
| 947 |
+
}
|
| 948 |
+
|
| 949 |
+
void gpt_print_usage(int argc, char ** argv, const gpt_hparams & params) {
|
| 950 |
+
fprintf(stderr, "usage: %s [options]\n", argv[0]);
|
| 951 |
+
fprintf(stderr, "\n");
|
| 952 |
+
fprintf(stderr, "options:\n");
|
| 953 |
+
fprintf(stderr, " -h, --help show this help message and exit\n");
|
| 954 |
+
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
|
| 955 |
+
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
|
| 956 |
+
fprintf(stderr, " -p PROMPT, --prompt PROMPT\n");
|
| 957 |
+
fprintf(stderr, " prompt to start generation with (default: random)\n");
|
| 958 |
+
fprintf(stderr, " -f FNAME, --file FNAME\n");
|
| 959 |
+
fprintf(stderr, " load prompt from a file\n");
|
| 960 |
+
fprintf(stderr, " -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
|
| 961 |
+
fprintf(stderr, " test tokenization\n");
|
| 962 |
+
fprintf(stderr, " -n N, --n_predict N number of tokens to predict (default: %d)\n", params.n_predict);
|
| 963 |
+
fprintf(stderr, " --top_k N top-k sampling (default: %d)\n", params.top_k);
|
| 964 |
+
fprintf(stderr, " --top_p N top-p sampling (default: %.1f)\n", params.top_p);
|
| 965 |
+
fprintf(stderr, " --temp N temperature (default: %.1f)\n", params.temp);
|
| 966 |
+
fprintf(stderr, " --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
|
| 967 |
+
fprintf(stderr, " --repeat-penalty N penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
|
| 968 |
+
fprintf(stderr, " -b N, --batch_size N batch size for prompt processing (default: %d)\n", params.n_batch);
|
| 969 |
+
fprintf(stderr, " -c N, --context N context / KV cache size (default: %d)\n", params.n_ctx);
|
| 970 |
+
fprintf(stderr, " --ignore-eos ignore EOS token during generation\n");
|
| 971 |
+
fprintf(stderr, " -ngl N, --gpu-layers N number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
|
| 972 |
+
fprintf(stderr, " -m FNAME, --model FNAME\n");
|
| 973 |
+
fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
|
| 974 |
+
fprintf(stderr, "\n");
|
| 975 |
+
}
|
| 976 |
+
|
| 977 |
+
// Function to check if the next argument exists
|
| 978 |
+
static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_hparams& params) {
|
| 979 |
+
if (i + 1 < argc && argv[i + 1][0] != '-') {
|
| 980 |
+
return argv[++i];
|
| 981 |
+
} else {
|
| 982 |
+
fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
|
| 983 |
+
gpt_print_usage(argc, argv, params);
|
| 984 |
+
exit(0);
|
| 985 |
+
}
|
| 986 |
+
}
|
| 987 |
+
|
| 988 |
+
bool gpt_params_parse(int argc, char ** argv, gpt_hparams & params) {
|
| 989 |
+
for (int i = 1; i < argc; i++) {
|
| 990 |
+
std::string arg = argv[i];
|
| 991 |
+
|
| 992 |
+
if (arg == "-s" || arg == "--seed") {
|
| 993 |
+
params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 994 |
+
} else if (arg == "-t" || arg == "--threads") {
|
| 995 |
+
params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 996 |
+
} else if (arg == "-p" || arg == "--prompt") {
|
| 997 |
+
params.prompt = get_next_arg(i, argc, argv, arg, params);
|
| 998 |
+
} else if (arg == "-n" || arg == "--n_predict") {
|
| 999 |
+
params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1000 |
+
} else if (arg == "-np" || arg == "--n_parallel") {
|
| 1001 |
+
params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1002 |
+
} else if (arg == "--top_k") {
|
| 1003 |
+
params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1004 |
+
} else if (arg == "--top_p") {
|
| 1005 |
+
params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
|
| 1006 |
+
} else if (arg == "--temp") {
|
| 1007 |
+
params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
|
| 1008 |
+
} else if (arg == "--repeat-last-n") {
|
| 1009 |
+
params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1010 |
+
} else if (arg == "--repeat-penalty") {
|
| 1011 |
+
params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
|
| 1012 |
+
} else if (arg == "-b" || arg == "--batch_size") {
|
| 1013 |
+
params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1014 |
+
} else if (arg == "-c" || arg == "--context") {
|
| 1015 |
+
params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1016 |
+
} else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
|
| 1017 |
+
params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1018 |
+
} else if (arg == "--ignore-eos") {
|
| 1019 |
+
params.ignore_eos = true;
|
| 1020 |
+
} else if (arg == "-m" || arg == "--model") {
|
| 1021 |
+
params.model = get_next_arg(i, argc, argv, arg, params);
|
| 1022 |
+
} else if (arg == "-i" || arg == "--interactive") {
|
| 1023 |
+
params.interactive = true;
|
| 1024 |
+
} else if (arg == "-ip" || arg == "--interactive-port") {
|
| 1025 |
+
params.interactive = true;
|
| 1026 |
+
params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
|
| 1027 |
+
} else if (arg == "-h" || arg == "--help") {
|
| 1028 |
+
gpt_print_usage(argc, argv, params);
|
| 1029 |
+
exit(0);
|
| 1030 |
+
} else if (arg == "-f" || arg == "--file") {
|
| 1031 |
+
get_next_arg(i, argc, argv, arg, params);
|
| 1032 |
+
std::ifstream file(argv[i]);
|
| 1033 |
+
if (!file) {
|
| 1034 |
+
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
|
| 1035 |
+
break;
|
| 1036 |
+
}
|
| 1037 |
+
std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
|
| 1038 |
+
if (params.prompt.back() == '\n') {
|
| 1039 |
+
params.prompt.pop_back();
|
| 1040 |
+
}
|
| 1041 |
+
} else if (arg == "-tt" || arg == "--token_test") {
|
| 1042 |
+
params.token_test = get_next_arg(i, argc, argv, arg, params);
|
| 1043 |
+
}
|
| 1044 |
+
else {
|
| 1045 |
+
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
|
| 1046 |
+
gpt_print_usage(argc, argv, params);
|
| 1047 |
+
exit(0);
|
| 1048 |
+
}
|
| 1049 |
+
}
|
| 1050 |
+
|
| 1051 |
+
return true;
|
| 1052 |
+
}
|
| 1053 |
+
|
| 1054 |
+
std::string gpt_random_prompt(std::mt19937 & rng) {
|
| 1055 |
+
const int r = rng() % 10;
|
| 1056 |
+
switch (r) {
|
| 1057 |
+
case 0: return "So";
|
| 1058 |
+
case 1: return "Once upon a time";
|
| 1059 |
+
case 2: return "When";
|
| 1060 |
+
case 3: return "The";
|
| 1061 |
+
case 4: return "After";
|
| 1062 |
+
case 5: return "If";
|
| 1063 |
+
case 6: return "import";
|
| 1064 |
+
case 7: return "He";
|
| 1065 |
+
case 8: return "She";
|
| 1066 |
+
case 9: return "They";
|
| 1067 |
+
}
|
| 1068 |
+
|
| 1069 |
+
return "The";
|
| 1070 |
+
}
|
| 1071 |
+
|
| 1072 |
+
int main(int argc, char ** argv) {
|
| 1073 |
+
ggml_time_init();
|
| 1074 |
+
|
| 1075 |
+
const int64_t t_main_start_us = ggml_time_us();
|
| 1076 |
+
|
| 1077 |
+
gpt_hparams params;
|
| 1078 |
+
|
| 1079 |
+
if (gpt_params_parse(argc, argv, params) == false) {
|
| 1080 |
+
return 1;
|
| 1081 |
+
}
|
| 1082 |
+
|
| 1083 |
+
if (params.seed < 0) {
|
| 1084 |
+
params.seed = time(NULL);
|
| 1085 |
+
}
|
| 1086 |
+
|
| 1087 |
+
printf("%s: seed = %d\n", __func__, params.seed);
|
| 1088 |
+
|
| 1089 |
+
std::mt19937 rng(params.seed);
|
| 1090 |
+
if (params.prompt.empty()) {
|
| 1091 |
+
params.prompt = gpt_random_prompt(rng);
|
| 1092 |
+
}
|
| 1093 |
+
|
| 1094 |
+
int64_t t_load_us = 0;
|
| 1095 |
+
|
| 1096 |
+
gpt_vocab vocab;
|
| 1097 |
+
gpt_model model;
|
| 1098 |
+
|
| 1099 |
+
// load the model
|
| 1100 |
+
{
|
| 1101 |
+
const int64_t t_start_us = ggml_time_us();
|
| 1102 |
+
|
| 1103 |
+
if (!gpt_model_load(params.model, model, vocab)) {
|
| 1104 |
+
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
|
| 1105 |
+
return 1;
|
| 1106 |
+
}
|
| 1107 |
+
|
| 1108 |
+
t_load_us = ggml_time_us() - t_start_us;
|
| 1109 |
+
|
| 1110 |
+
test_gpt_tokenizer(vocab, params.token_test);
|
| 1111 |
+
}
|
| 1112 |
+
|
| 1113 |
+
while(true) {
|
| 1114 |
+
int n_past = 0;
|
| 1115 |
+
|
| 1116 |
+
int64_t t_sample_us = 0;
|
| 1117 |
+
int64_t t_predict_us = 0;
|
| 1118 |
+
|
| 1119 |
+
std::vector<float> logits;
|
| 1120 |
+
|
| 1121 |
+
// tokenize the prompt
|
| 1122 |
+
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
|
| 1123 |
+
|
| 1124 |
+
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
|
| 1125 |
+
|
| 1126 |
+
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
|
| 1127 |
+
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
|
| 1128 |
+
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
|
| 1129 |
+
printf("%d ", embd_inp[i]);
|
| 1130 |
+
}
|
| 1131 |
+
printf("\n\n");
|
| 1132 |
+
|
| 1133 |
+
// submit the input prompt token-by-token
|
| 1134 |
+
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
|
| 1135 |
+
std::vector<gpt_vocab::id> embd;
|
| 1136 |
+
|
| 1137 |
+
// determine the required inference memory per token:
|
| 1138 |
+
size_t mem_per_token = 0;
|
| 1139 |
+
gpt_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
|
| 1140 |
+
|
| 1141 |
+
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
|
| 1142 |
+
// predict
|
| 1143 |
+
if (embd.size() > 0) {
|
| 1144 |
+
const int64_t t_start_us = ggml_time_us();
|
| 1145 |
+
|
| 1146 |
+
if (!gpt_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
|
| 1147 |
+
printf("Failed to predict\n");
|
| 1148 |
+
return 1;
|
| 1149 |
+
}
|
| 1150 |
+
|
| 1151 |
+
t_predict_us += ggml_time_us() - t_start_us;
|
| 1152 |
+
}
|
| 1153 |
+
|
| 1154 |
+
n_past += embd.size();
|
| 1155 |
+
embd.clear();
|
| 1156 |
+
|
| 1157 |
+
if (i >= embd_inp.size()) {
|
| 1158 |
+
// sample next token
|
| 1159 |
+
const int top_k = params.top_k;
|
| 1160 |
+
const float top_p = params.top_p;
|
| 1161 |
+
const float temp = params.temp;
|
| 1162 |
+
|
| 1163 |
+
const int n_vocab = model.hparams.n_vocab;
|
| 1164 |
+
|
| 1165 |
+
gpt_vocab::id id = 0;
|
| 1166 |
+
|
| 1167 |
+
{
|
| 1168 |
+
const int64_t t_start_sample_us = ggml_time_us();
|
| 1169 |
+
|
| 1170 |
+
id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
|
| 1171 |
+
|
| 1172 |
+
t_sample_us += ggml_time_us() - t_start_sample_us;
|
| 1173 |
+
}
|
| 1174 |
+
|
| 1175 |
+
// add it to the context
|
| 1176 |
+
embd.push_back(id);
|
| 1177 |
+
} else {
|
| 1178 |
+
// if here, it means we are still processing the input prompt
|
| 1179 |
+
for (size_t k = i; k < embd_inp.size(); k++) {
|
| 1180 |
+
embd.push_back(embd_inp[k]);
|
| 1181 |
+
if (int32_t(embd.size()) >= params.n_batch) {
|
| 1182 |
+
break;
|
| 1183 |
+
}
|
| 1184 |
+
}
|
| 1185 |
+
i += embd.size() - 1;
|
| 1186 |
+
}
|
| 1187 |
+
|
| 1188 |
+
// display text
|
| 1189 |
+
for (auto id : embd) {
|
| 1190 |
+
printf("%s", vocab.id_to_token[id].c_str());
|
| 1191 |
+
}
|
| 1192 |
+
fflush(stdout);
|
| 1193 |
+
|
| 1194 |
+
// end of text token
|
| 1195 |
+
if (embd.back() == 50256) {
|
| 1196 |
+
// report timing
|
| 1197 |
+
{
|
| 1198 |
+
const int64_t t_main_end_us = ggml_time_us();
|
| 1199 |
+
|
| 1200 |
+
printf("\n\n");
|
| 1201 |
+
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
|
| 1202 |
+
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
|
| 1203 |
+
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
|
| 1204 |
+
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
|
| 1205 |
+
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
|
| 1206 |
+
}
|
| 1207 |
+
break;
|
| 1208 |
+
}
|
| 1209 |
+
}
|
| 1210 |
+
}
|
| 1211 |
+
|
| 1212 |
+
ggml_free(model.ctx_w);
|
| 1213 |
+
|
| 1214 |
+
return 0;
|
| 1215 |
+
}
|