llama_cpp 0.16.2 → 0.17.0

Sign up to get free protection for your applications and to get access to all the features.
Files changed (177) hide show
  1. checksums.yaml +4 -4
  2. data/CHANGELOG.md +18 -0
  3. data/README.md +7 -12
  4. data/ext/llama_cpp/extconf.rb +2 -43
  5. data/ext/llama_cpp/llama_cpp.cpp +8 -0
  6. data/lib/llama_cpp/version.rb +3 -3
  7. data/sig/llama_cpp.rbs +3 -0
  8. metadata +2 -171
  9. data/vendor/include/.gitkeep +0 -0
  10. data/vendor/lib/.gitkeep +0 -0
  11. data/vendor/tmp/llama.cpp/LICENSE +0 -21
  12. data/vendor/tmp/llama.cpp/Makefile +0 -1124
  13. data/vendor/tmp/llama.cpp/ggml-alloc.c +0 -1041
  14. data/vendor/tmp/llama.cpp/ggml-alloc.h +0 -76
  15. data/vendor/tmp/llama.cpp/ggml-backend-impl.h +0 -153
  16. data/vendor/tmp/llama.cpp/ggml-backend.c +0 -2225
  17. data/vendor/tmp/llama.cpp/ggml-backend.h +0 -236
  18. data/vendor/tmp/llama.cpp/ggml-blas.cpp +0 -363
  19. data/vendor/tmp/llama.cpp/ggml-blas.h +0 -23
  20. data/vendor/tmp/llama.cpp/ggml-common.h +0 -1805
  21. data/vendor/tmp/llama.cpp/ggml-cuda/acc.cu +0 -47
  22. data/vendor/tmp/llama.cpp/ggml-cuda/arange.cu +0 -34
  23. data/vendor/tmp/llama.cpp/ggml-cuda/argsort.cu +0 -104
  24. data/vendor/tmp/llama.cpp/ggml-cuda/binbcast.cu +0 -280
  25. data/vendor/tmp/llama.cpp/ggml-cuda/clamp.cu +0 -34
  26. data/vendor/tmp/llama.cpp/ggml-cuda/concat.cu +0 -196
  27. data/vendor/tmp/llama.cpp/ggml-cuda/convert.cu +0 -686
  28. data/vendor/tmp/llama.cpp/ggml-cuda/cpy.cu +0 -490
  29. data/vendor/tmp/llama.cpp/ggml-cuda/diagmask.cu +0 -40
  30. data/vendor/tmp/llama.cpp/ggml-cuda/dmmv.cu +0 -674
  31. data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f16.cu +0 -319
  32. data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f32.cu +0 -312
  33. data/vendor/tmp/llama.cpp/ggml-cuda/fattn.cu +0 -345
  34. data/vendor/tmp/llama.cpp/ggml-cuda/getrows.cu +0 -178
  35. data/vendor/tmp/llama.cpp/ggml-cuda/im2col.cu +0 -104
  36. data/vendor/tmp/llama.cpp/ggml-cuda/mmq.cu +0 -88
  37. data/vendor/tmp/llama.cpp/ggml-cuda/mmvq.cu +0 -419
  38. data/vendor/tmp/llama.cpp/ggml-cuda/norm.cu +0 -221
  39. data/vendor/tmp/llama.cpp/ggml-cuda/pad.cu +0 -49
  40. data/vendor/tmp/llama.cpp/ggml-cuda/pool2d.cu +0 -94
  41. data/vendor/tmp/llama.cpp/ggml-cuda/quantize.cu +0 -112
  42. data/vendor/tmp/llama.cpp/ggml-cuda/rope.cu +0 -271
  43. data/vendor/tmp/llama.cpp/ggml-cuda/scale.cu +0 -31
  44. data/vendor/tmp/llama.cpp/ggml-cuda/softmax.cu +0 -206
  45. data/vendor/tmp/llama.cpp/ggml-cuda/sumrows.cu +0 -40
  46. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
  47. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
  48. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
  49. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
  50. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
  51. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
  52. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
  53. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
  54. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
  55. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
  56. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
  57. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
  58. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
  59. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
  60. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
  61. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
  62. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
  63. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
  64. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
  65. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
  66. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
  67. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
  68. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
  69. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
  70. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
  71. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
  72. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
  73. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
  74. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
  75. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
  76. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
  77. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
  78. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
  79. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
  80. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
  81. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
  82. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
  83. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
  84. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
  85. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
  86. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
  87. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
  88. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
  89. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
  90. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
  91. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
  92. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
  93. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
  94. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
  95. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
  96. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
  97. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
  98. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
  99. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
  100. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
  101. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
  102. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
  103. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
  104. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
  105. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
  106. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
  107. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
  108. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
  109. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
  110. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
  111. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
  112. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
  113. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
  114. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
  115. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
  116. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
  117. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
  118. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
  119. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
  120. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
  121. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
  122. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
  123. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
  124. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
  125. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
  126. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
  127. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
  128. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
  129. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
  130. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
  131. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
  132. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu +0 -10
  133. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu +0 -9
  134. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu +0 -10
  135. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu +0 -10
  136. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu +0 -8
  137. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q2_k.cu +0 -5
  138. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q3_k.cu +0 -5
  139. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_0.cu +0 -5
  140. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_1.cu +0 -5
  141. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_k.cu +0 -5
  142. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_0.cu +0 -5
  143. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_1.cu +0 -5
  144. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_k.cu +0 -5
  145. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q6_k.cu +0 -5
  146. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q8_0.cu +0 -5
  147. data/vendor/tmp/llama.cpp/ggml-cuda/tsembd.cu +0 -47
  148. data/vendor/tmp/llama.cpp/ggml-cuda/unary.cu +0 -314
  149. data/vendor/tmp/llama.cpp/ggml-cuda/upscale.cu +0 -51
  150. data/vendor/tmp/llama.cpp/ggml-cuda.cu +0 -3069
  151. data/vendor/tmp/llama.cpp/ggml-cuda.h +0 -44
  152. data/vendor/tmp/llama.cpp/ggml-impl.h +0 -651
  153. data/vendor/tmp/llama.cpp/ggml-kompute.cpp +0 -2038
  154. data/vendor/tmp/llama.cpp/ggml-kompute.h +0 -46
  155. data/vendor/tmp/llama.cpp/ggml-metal.h +0 -66
  156. data/vendor/tmp/llama.cpp/ggml-metal.m +0 -3273
  157. data/vendor/tmp/llama.cpp/ggml-metal.metal +0 -6540
  158. data/vendor/tmp/llama.cpp/ggml-quants.c +0 -14994
  159. data/vendor/tmp/llama.cpp/ggml-quants.h +0 -133
  160. data/vendor/tmp/llama.cpp/ggml-rpc.cpp +0 -1178
  161. data/vendor/tmp/llama.cpp/ggml-rpc.h +0 -24
  162. data/vendor/tmp/llama.cpp/ggml-sycl.cpp +0 -6351
  163. data/vendor/tmp/llama.cpp/ggml-sycl.h +0 -40
  164. data/vendor/tmp/llama.cpp/ggml-vulkan-shaders.hpp +0 -144508
  165. data/vendor/tmp/llama.cpp/ggml-vulkan.cpp +0 -7183
  166. data/vendor/tmp/llama.cpp/ggml-vulkan.h +0 -29
  167. data/vendor/tmp/llama.cpp/ggml.c +0 -22506
  168. data/vendor/tmp/llama.cpp/ggml.h +0 -2458
  169. data/vendor/tmp/llama.cpp/llama.cpp +0 -18985
  170. data/vendor/tmp/llama.cpp/llama.h +0 -1147
  171. data/vendor/tmp/llama.cpp/scripts/get-flags.mk +0 -38
  172. data/vendor/tmp/llama.cpp/sgemm.cpp +0 -1032
  173. data/vendor/tmp/llama.cpp/sgemm.h +0 -14
  174. data/vendor/tmp/llama.cpp/unicode-data.cpp +0 -7033
  175. data/vendor/tmp/llama.cpp/unicode-data.h +0 -20
  176. data/vendor/tmp/llama.cpp/unicode.cpp +0 -810
  177. data/vendor/tmp/llama.cpp/unicode.h +0 -63
@@ -1,1147 +0,0 @@
1
- #ifndef LLAMA_H
2
- #define LLAMA_H
3
-
4
- #include "ggml.h"
5
- #include "ggml-backend.h"
6
-
7
- #include <stddef.h>
8
- #include <stdint.h>
9
- #include <stdio.h>
10
- #include <stdbool.h>
11
-
12
- #ifdef LLAMA_SHARED
13
- # if defined(_WIN32) && !defined(__MINGW32__)
14
- # ifdef LLAMA_BUILD
15
- # define LLAMA_API __declspec(dllexport)
16
- # else
17
- # define LLAMA_API __declspec(dllimport)
18
- # endif
19
- # else
20
- # define LLAMA_API __attribute__ ((visibility ("default")))
21
- # endif
22
- #else
23
- # define LLAMA_API
24
- #endif
25
-
26
- #ifdef __GNUC__
27
- # define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
28
- #elif defined(_MSC_VER)
29
- # define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
30
- #else
31
- # define DEPRECATED(func, hint) func
32
- #endif
33
-
34
- #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
35
-
36
- #define LLAMA_MAX_RNG_STATE (64*1024)
37
-
38
- #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
39
- #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
40
- #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
41
-
42
- #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN
43
- #define LLAMA_SESSION_VERSION 6
44
-
45
- #define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ
46
- #define LLAMA_STATE_SEQ_VERSION 1
47
-
48
- #ifdef __cplusplus
49
- extern "C" {
50
- #endif
51
-
52
- //
53
- // C interface
54
- //
55
- // TODO: show sample usage
56
- //
57
-
58
- struct llama_model;
59
- struct llama_context;
60
-
61
- typedef int32_t llama_pos;
62
- typedef int32_t llama_token;
63
- typedef int32_t llama_seq_id;
64
-
65
- enum llama_vocab_type {
66
- LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
67
- LLAMA_VOCAB_TYPE_SPM = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
68
- LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE
69
- LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece
70
- };
71
-
72
- // pre-tokenization types
73
- enum llama_vocab_pre_type {
74
- LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0,
75
- LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1,
76
- LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2,
77
- LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
78
- LLAMA_VOCAB_PRE_TYPE_FALCON = 4,
79
- LLAMA_VOCAB_PRE_TYPE_MPT = 5,
80
- LLAMA_VOCAB_PRE_TYPE_STARCODER = 6,
81
- LLAMA_VOCAB_PRE_TYPE_GPT2 = 7,
82
- LLAMA_VOCAB_PRE_TYPE_REFACT = 8,
83
- LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9,
84
- LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10,
85
- LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11,
86
- LLAMA_VOCAB_PRE_TYPE_OLMO = 12,
87
- LLAMA_VOCAB_PRE_TYPE_DBRX = 13,
88
- LLAMA_VOCAB_PRE_TYPE_SMAUG = 14,
89
- LLAMA_VOCAB_PRE_TYPE_PORO = 15,
90
- };
91
-
92
- // note: these values should be synchronized with ggml_rope
93
- // TODO: maybe move this enum to ggml.h (ggml_rope_type)
94
- enum llama_rope_type {
95
- LLAMA_ROPE_TYPE_NONE = -1,
96
- LLAMA_ROPE_TYPE_NORM = 0,
97
- LLAMA_ROPE_TYPE_NEOX = 2,
98
- LLAMA_ROPE_TYPE_GLM = 4,
99
- };
100
-
101
- enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
102
- LLAMA_TOKEN_TYPE_UNDEFINED = 0,
103
- LLAMA_TOKEN_TYPE_NORMAL = 1,
104
- LLAMA_TOKEN_TYPE_UNKNOWN = 2,
105
- LLAMA_TOKEN_TYPE_CONTROL = 3,
106
- LLAMA_TOKEN_TYPE_USER_DEFINED = 4,
107
- LLAMA_TOKEN_TYPE_UNUSED = 5,
108
- LLAMA_TOKEN_TYPE_BYTE = 6,
109
- };
110
-
111
- enum llama_token_attr {
112
- LLAMA_TOKEN_ATTR_UNDEFINED = 0,
113
- LLAMA_TOKEN_ATTR_UNKNOWN = 1 << 0,
114
- LLAMA_TOKEN_ATTR_UNUSED = 1 << 1,
115
- LLAMA_TOKEN_ATTR_NORMAL = 1 << 2,
116
- LLAMA_TOKEN_ATTR_CONTROL = 1 << 3, // SPECIAL?
117
- LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4,
118
- LLAMA_TOKEN_ATTR_BYTE = 1 << 5,
119
- LLAMA_TOKEN_ATTR_NORMALIZED = 1 << 6,
120
- LLAMA_TOKEN_ATTR_LSTRIP = 1 << 7,
121
- LLAMA_TOKEN_ATTR_RSTRIP = 1 << 8,
122
- LLAMA_TOKEN_ATTR_SINGLE_WORD = 1 << 9,
123
- };
124
-
125
- // model file types
126
- enum llama_ftype {
127
- LLAMA_FTYPE_ALL_F32 = 0,
128
- LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
129
- LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
130
- LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
131
- LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
132
- // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
133
- // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed
134
- LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
135
- LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
136
- LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
137
- LLAMA_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
138
- LLAMA_FTYPE_MOSTLY_Q3_K_S = 11, // except 1d tensors
139
- LLAMA_FTYPE_MOSTLY_Q3_K_M = 12, // except 1d tensors
140
- LLAMA_FTYPE_MOSTLY_Q3_K_L = 13, // except 1d tensors
141
- LLAMA_FTYPE_MOSTLY_Q4_K_S = 14, // except 1d tensors
142
- LLAMA_FTYPE_MOSTLY_Q4_K_M = 15, // except 1d tensors
143
- LLAMA_FTYPE_MOSTLY_Q5_K_S = 16, // except 1d tensors
144
- LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors
145
- LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors
146
- LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors
147
- LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors
148
- LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors
149
- LLAMA_FTYPE_MOSTLY_IQ3_XS = 22, // except 1d tensors
150
- LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors
151
- LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors
152
- LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors
153
- LLAMA_FTYPE_MOSTLY_IQ3_S = 26, // except 1d tensors
154
- LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors
155
- LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors
156
- LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors
157
- LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors
158
- LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors
159
- LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors
160
-
161
- LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
162
- };
163
-
164
- enum llama_rope_scaling_type {
165
- LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1,
166
- LLAMA_ROPE_SCALING_TYPE_NONE = 0,
167
- LLAMA_ROPE_SCALING_TYPE_LINEAR = 1,
168
- LLAMA_ROPE_SCALING_TYPE_YARN = 2,
169
- LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN,
170
- };
171
-
172
- enum llama_pooling_type {
173
- LLAMA_POOLING_TYPE_UNSPECIFIED = -1,
174
- LLAMA_POOLING_TYPE_NONE = 0,
175
- LLAMA_POOLING_TYPE_MEAN = 1,
176
- LLAMA_POOLING_TYPE_CLS = 2,
177
- LLAMA_POOLING_TYPE_LAST = 3,
178
- };
179
-
180
- enum llama_split_mode {
181
- LLAMA_SPLIT_MODE_NONE = 0, // single GPU
182
- LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
183
- LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs
184
- };
185
-
186
- typedef struct llama_token_data {
187
- llama_token id; // token id
188
- float logit; // log-odds of the token
189
- float p; // probability of the token
190
- } llama_token_data;
191
-
192
- typedef struct llama_token_data_array {
193
- llama_token_data * data;
194
- size_t size;
195
- bool sorted;
196
- } llama_token_data_array;
197
-
198
- typedef bool (*llama_progress_callback)(float progress, void * user_data);
199
-
200
- // Input data for llama_decode
201
- // A llama_batch object can contain input about one or many sequences
202
- // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
203
- //
204
- // - token : the token ids of the input (used when embd is NULL)
205
- // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
206
- // - pos : the positions of the respective token in the sequence
207
- // - seq_id : the sequence to which the respective token belongs
208
- // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
209
- //
210
- typedef struct llama_batch {
211
- int32_t n_tokens;
212
-
213
- llama_token * token;
214
- float * embd;
215
- llama_pos * pos;
216
- int32_t * n_seq_id;
217
- llama_seq_id ** seq_id;
218
- int8_t * logits; // TODO: rename this to "output"
219
-
220
- // NOTE: helpers for smooth API transition - can be deprecated in the future
221
- // for future-proof code, use the above fields instead and ignore everything below
222
- //
223
- // pos[i] = all_pos_0 + i*all_pos_1
224
- //
225
- llama_pos all_pos_0; // used if pos == NULL
226
- llama_pos all_pos_1; // used if pos == NULL
227
- llama_seq_id all_seq_id; // used if seq_id == NULL
228
- } llama_batch;
229
-
230
- enum llama_model_kv_override_type {
231
- LLAMA_KV_OVERRIDE_TYPE_INT,
232
- LLAMA_KV_OVERRIDE_TYPE_FLOAT,
233
- LLAMA_KV_OVERRIDE_TYPE_BOOL,
234
- LLAMA_KV_OVERRIDE_TYPE_STR,
235
- };
236
-
237
- struct llama_model_kv_override {
238
- enum llama_model_kv_override_type tag;
239
-
240
- char key[128];
241
-
242
- union {
243
- int64_t val_i64;
244
- double val_f64;
245
- bool val_bool;
246
- char val_str[128];
247
- };
248
- };
249
-
250
- struct llama_model_params {
251
- int32_t n_gpu_layers; // number of layers to store in VRAM
252
- enum llama_split_mode split_mode; // how to split the model across multiple GPUs
253
-
254
- // main_gpu interpretation depends on split_mode:
255
- // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
256
- // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
257
- // LLAMA_SPLIT_LAYER: ignored
258
- int32_t main_gpu;
259
-
260
- // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
261
- const float * tensor_split;
262
-
263
- // comma separated list of RPC servers to use for offloading
264
- const char * rpc_servers;
265
-
266
- // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
267
- // If the provided progress_callback returns true, model loading continues.
268
- // If it returns false, model loading is immediately aborted.
269
- llama_progress_callback progress_callback;
270
-
271
- // context pointer passed to the progress callback
272
- void * progress_callback_user_data;
273
-
274
- // override key-value pairs of the model meta data
275
- const struct llama_model_kv_override * kv_overrides;
276
-
277
- // Keep the booleans together to avoid misalignment during copy-by-value.
278
- bool vocab_only; // only load the vocabulary, no weights
279
- bool use_mmap; // use mmap if possible
280
- bool use_mlock; // force system to keep model in RAM
281
- bool check_tensors; // validate model tensor data
282
- };
283
-
284
- // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
285
- // https://github.com/ggerganov/llama.cpp/pull/7544
286
- struct llama_context_params {
287
- uint32_t seed; // RNG seed, -1 for random
288
- uint32_t n_ctx; // text context, 0 = from model
289
- uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode
290
- uint32_t n_ubatch; // physical maximum batch size
291
- uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models)
292
- uint32_t n_threads; // number of threads to use for generation
293
- uint32_t n_threads_batch; // number of threads to use for batch processing
294
-
295
- enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
296
- enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id
297
-
298
- // ref: https://github.com/ggerganov/llama.cpp/pull/2054
299
- float rope_freq_base; // RoPE base frequency, 0 = from model
300
- float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
301
- float yarn_ext_factor; // YaRN extrapolation mix factor, negative = from model
302
- float yarn_attn_factor; // YaRN magnitude scaling factor
303
- float yarn_beta_fast; // YaRN low correction dim
304
- float yarn_beta_slow; // YaRN high correction dim
305
- uint32_t yarn_orig_ctx; // YaRN original context size
306
- float defrag_thold; // defragment the KV cache if holes/size > thold, < 0 disabled (default)
307
-
308
- ggml_backend_sched_eval_callback cb_eval;
309
- void * cb_eval_user_data;
310
-
311
- enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
312
- enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
313
-
314
- // Keep the booleans together to avoid misalignment during copy-by-value.
315
- bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
316
- bool embeddings; // if true, extract embeddings (together with logits)
317
- bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
318
- bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
319
-
320
- // Abort callback
321
- // if it returns true, execution of llama_decode() will be aborted
322
- // currently works only with CPU execution
323
- ggml_abort_callback abort_callback;
324
- void * abort_callback_data;
325
- };
326
-
327
- // model quantization parameters
328
- typedef struct llama_model_quantize_params {
329
- int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
330
- enum llama_ftype ftype; // quantize to this llama_ftype
331
- enum ggml_type output_tensor_type; // output tensor type
332
- enum ggml_type token_embedding_type; // itoken embeddings tensor type
333
- bool allow_requantize; // allow quantizing non-f32/f16 tensors
334
- bool quantize_output_tensor; // quantize output.weight
335
- bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
336
- bool pure; // quantize all tensors to the default type
337
- bool keep_split; // quantize to the same number of shards
338
- void * imatrix; // pointer to importance matrix data
339
- void * kv_overrides; // pointer to vector containing overrides
340
- } llama_model_quantize_params;
341
-
342
- // grammar types
343
- struct llama_grammar;
344
-
345
- // grammar element type
346
- enum llama_gretype {
347
- // end of rule definition
348
- LLAMA_GRETYPE_END = 0,
349
-
350
- // start of alternate definition for rule
351
- LLAMA_GRETYPE_ALT = 1,
352
-
353
- // non-terminal element: reference to rule
354
- LLAMA_GRETYPE_RULE_REF = 2,
355
-
356
- // terminal element: character (code point)
357
- LLAMA_GRETYPE_CHAR = 3,
358
-
359
- // inverse char(s) ([^a], [^a-b] [^abc])
360
- LLAMA_GRETYPE_CHAR_NOT = 4,
361
-
362
- // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
363
- // be an inclusive range ([a-z])
364
- LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
365
-
366
- // modifies a preceding LLAMA_GRETYPE_CHAR or
367
- // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
368
- LLAMA_GRETYPE_CHAR_ALT = 6,
369
-
370
- // any character (.)
371
- LLAMA_GRETYPE_CHAR_ANY = 7,
372
- };
373
-
374
- typedef struct llama_grammar_element {
375
- enum llama_gretype type;
376
- uint32_t value; // Unicode code point or rule ID
377
- } llama_grammar_element;
378
-
379
- // performance timing information
380
- struct llama_timings {
381
- double t_start_ms;
382
- double t_end_ms;
383
- double t_load_ms;
384
- double t_sample_ms;
385
- double t_p_eval_ms;
386
- double t_eval_ms;
387
-
388
- int32_t n_sample;
389
- int32_t n_p_eval;
390
- int32_t n_eval;
391
- };
392
-
393
- // used in chat template
394
- typedef struct llama_chat_message {
395
- const char * role;
396
- const char * content;
397
- } llama_chat_message;
398
-
399
- // Helpers for getting default parameters
400
- LLAMA_API struct llama_model_params llama_model_default_params(void);
401
- LLAMA_API struct llama_context_params llama_context_default_params(void);
402
- LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
403
-
404
- // Initialize the llama + ggml backend
405
- // If numa is true, use NUMA optimizations
406
- // Call once at the start of the program
407
- LLAMA_API void llama_backend_init(void);
408
-
409
- //optional:
410
- LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
411
-
412
- // Call once at the end of the program - currently only used for MPI
413
- LLAMA_API void llama_backend_free(void);
414
-
415
- LLAMA_API struct llama_model * llama_load_model_from_file(
416
- const char * path_model,
417
- struct llama_model_params params);
418
-
419
- LLAMA_API void llama_free_model(struct llama_model * model);
420
-
421
- LLAMA_API struct llama_context * llama_new_context_with_model(
422
- struct llama_model * model,
423
- struct llama_context_params params);
424
-
425
- // Frees all allocated memory
426
- LLAMA_API void llama_free(struct llama_context * ctx);
427
-
428
- LLAMA_API int64_t llama_time_us(void);
429
-
430
- LLAMA_API size_t llama_max_devices(void);
431
-
432
- LLAMA_API bool llama_supports_mmap (void);
433
- LLAMA_API bool llama_supports_mlock (void);
434
- LLAMA_API bool llama_supports_gpu_offload(void);
435
-
436
- LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
437
-
438
- LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx);
439
- LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx);
440
- LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx);
441
- LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx);
442
-
443
- LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
444
-
445
- LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model);
446
- LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model);
447
-
448
- LLAMA_API int32_t llama_n_vocab (const struct llama_model * model);
449
- LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
450
- LLAMA_API int32_t llama_n_embd (const struct llama_model * model);
451
- LLAMA_API int32_t llama_n_layer (const struct llama_model * model);
452
-
453
- // Get the model's RoPE frequency scaling factor
454
- LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
455
-
456
- // Functions to access the model's GGUF metadata scalar values
457
- // - The functions return the length of the string on success, or -1 on failure
458
- // - The output string is always null-terminated and cleared on failure
459
- // - GGUF array values are not supported by these functions
460
-
461
- // Get metadata value as a string by key name
462
- LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size);
463
-
464
- // Get the number of metadata key/value pairs
465
- LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
466
-
467
- // Get metadata key name by index
468
- LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
469
-
470
- // Get metadata value as a string by index
471
- LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
472
-
473
- // Get a string describing the model type
474
- LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
475
-
476
- // Returns the total size of all the tensors in the model in bytes
477
- LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
478
-
479
- // Returns the total number of parameters in the model
480
- LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
481
-
482
- // Get a llama model tensor
483
- LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
484
-
485
- // Returns 0 on success
486
- LLAMA_API uint32_t llama_model_quantize(
487
- const char * fname_inp,
488
- const char * fname_out,
489
- const llama_model_quantize_params * params);
490
-
491
- // Apply a LoRA adapter to a loaded model
492
- // path_base_model is the path to a higher quality model to use as a base for
493
- // the layers modified by the adapter. Can be NULL to use the current loaded model.
494
- // The model needs to be reloaded before applying a new adapter, otherwise the adapter
495
- // will be applied on top of the previous one
496
- // Returns 0 on success
497
- LLAMA_API int32_t llama_model_apply_lora_from_file(
498
- const struct llama_model * model,
499
- const char * path_lora,
500
- float scale,
501
- const char * path_base_model,
502
- int32_t n_threads);
503
-
504
- // Apply a loaded control vector to a llama_context, or if data is NULL, clear
505
- // the currently loaded vector.
506
- // n_embd should be the size of a single layer's control, and data should point
507
- // to an n_embd x n_layers buffer starting from layer 1.
508
- // il_start and il_end are the layer range the vector should apply to (both inclusive)
509
- // See llama_control_vector_load in common to load a control vector.
510
- LLAMA_API int32_t llama_control_vector_apply(
511
- struct llama_context * lctx,
512
- const float * data,
513
- size_t len,
514
- int32_t n_embd,
515
- int32_t il_start,
516
- int32_t il_end);
517
-
518
- //
519
- // KV cache
520
- //
521
-
522
- // Information associated with an individual cell in the KV cache view.
523
- struct llama_kv_cache_view_cell {
524
- // The position for this cell. Takes KV cache shifts into account.
525
- // May be negative if the cell is not populated.
526
- llama_pos pos;
527
- };
528
-
529
- // An updateable view of the KV cache.
530
- struct llama_kv_cache_view {
531
- // Number of KV cache cells. This will be the same as the context size.
532
- int32_t n_cells;
533
-
534
- // Maximum number of sequences that can exist in a cell. It's not an error
535
- // if there are more sequences in a cell than this value, however they will
536
- // not be visible in the view cells_sequences.
537
- int32_t n_seq_max;
538
-
539
- // Number of tokens in the cache. For example, if there are two populated
540
- // cells, the first with 1 sequence id in it and the second with 2 sequence
541
- // ids then you'll have 3 tokens.
542
- int32_t token_count;
543
-
544
- // Number of populated cache cells.
545
- int32_t used_cells;
546
-
547
- // Maximum contiguous empty slots in the cache.
548
- int32_t max_contiguous;
549
-
550
- // Index to the start of the max_contiguous slot range. Can be negative
551
- // when cache is full.
552
- int32_t max_contiguous_idx;
553
-
554
- // Information for an individual cell.
555
- struct llama_kv_cache_view_cell * cells;
556
-
557
- // The sequences for each cell. There will be n_seq_max items per cell.
558
- llama_seq_id * cells_sequences;
559
- };
560
-
561
- // Create an empty KV cache view. (use only for debugging purposes)
562
- LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max);
563
-
564
- // Free a KV cache view. (use only for debugging purposes)
565
- LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
566
-
567
- // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
568
- LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
569
-
570
- // Returns the number of tokens in the KV cache (slow, use only for debug)
571
- // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
572
- LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
573
-
574
- // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
575
- LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
576
-
577
- // Clear the KV cache - both cell info is erased and KV data is zeroed
578
- LLAMA_API void llama_kv_cache_clear(
579
- struct llama_context * ctx);
580
-
581
- // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
582
- // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
583
- // seq_id < 0 : match any sequence
584
- // p0 < 0 : [0, p1]
585
- // p1 < 0 : [p0, inf)
586
- LLAMA_API bool llama_kv_cache_seq_rm(
587
- struct llama_context * ctx,
588
- llama_seq_id seq_id,
589
- llama_pos p0,
590
- llama_pos p1);
591
-
592
- // Copy all tokens that belong to the specified sequence to another sequence
593
- // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
594
- // p0 < 0 : [0, p1]
595
- // p1 < 0 : [p0, inf)
596
- LLAMA_API void llama_kv_cache_seq_cp(
597
- struct llama_context * ctx,
598
- llama_seq_id seq_id_src,
599
- llama_seq_id seq_id_dst,
600
- llama_pos p0,
601
- llama_pos p1);
602
-
603
- // Removes all tokens that do not belong to the specified sequence
604
- LLAMA_API void llama_kv_cache_seq_keep(
605
- struct llama_context * ctx,
606
- llama_seq_id seq_id);
607
-
608
- // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
609
- // If the KV cache is RoPEd, the KV data is updated accordingly:
610
- // - lazily on next llama_decode()
611
- // - explicitly with llama_kv_cache_update()
612
- // p0 < 0 : [0, p1]
613
- // p1 < 0 : [p0, inf)
614
- LLAMA_API void llama_kv_cache_seq_add(
615
- struct llama_context * ctx,
616
- llama_seq_id seq_id,
617
- llama_pos p0,
618
- llama_pos p1,
619
- llama_pos delta);
620
-
621
- // Integer division of the positions by factor of `d > 1`
622
- // If the KV cache is RoPEd, the KV data is updated accordingly:
623
- // - lazily on next llama_decode()
624
- // - explicitly with llama_kv_cache_update()
625
- // p0 < 0 : [0, p1]
626
- // p1 < 0 : [p0, inf)
627
- LLAMA_API void llama_kv_cache_seq_div(
628
- struct llama_context * ctx,
629
- llama_seq_id seq_id,
630
- llama_pos p0,
631
- llama_pos p1,
632
- int d);
633
-
634
- // Returns the largest position present in the KV cache for the specified sequence
635
- LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
636
- struct llama_context * ctx,
637
- llama_seq_id seq_id);
638
-
639
- // Defragment the KV cache
640
- // This will be applied:
641
- // - lazily on next llama_decode()
642
- // - explicitly with llama_kv_cache_update()
643
- LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
644
-
645
- // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
646
- LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
647
-
648
- //
649
- // State / sessions
650
- //
651
-
652
- // Returns the maximum size in bytes of the state (rng, logits, embedding
653
- // and kv_cache) - will often be smaller after compacting tokens
654
- LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
655
- LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
656
- "use llama_state_get_size instead");
657
-
658
- // Copies the state to the specified destination address.
659
- // Destination needs to have allocated enough memory.
660
- // Returns the number of bytes copied
661
- LLAMA_API size_t llama_state_get_data(
662
- struct llama_context * ctx,
663
- uint8_t * dst);
664
- LLAMA_API DEPRECATED(size_t llama_copy_state_data(
665
- struct llama_context * ctx,
666
- uint8_t * dst),
667
- "use llama_state_get_data instead");
668
-
669
- // Set the state reading from the specified address
670
- // Returns the number of bytes read
671
- LLAMA_API size_t llama_state_set_data(
672
- struct llama_context * ctx,
673
- const uint8_t * src);
674
- LLAMA_API DEPRECATED(size_t llama_set_state_data(
675
- struct llama_context * ctx,
676
- const uint8_t * src),
677
- "use llama_state_set_data instead");
678
-
679
- // Save/load session file
680
- LLAMA_API bool llama_state_load_file(
681
- struct llama_context * ctx,
682
- const char * path_session,
683
- llama_token * tokens_out,
684
- size_t n_token_capacity,
685
- size_t * n_token_count_out);
686
- LLAMA_API DEPRECATED(bool llama_load_session_file(
687
- struct llama_context * ctx,
688
- const char * path_session,
689
- llama_token * tokens_out,
690
- size_t n_token_capacity,
691
- size_t * n_token_count_out),
692
- "use llama_state_load_file instead");
693
-
694
- LLAMA_API bool llama_state_save_file(
695
- struct llama_context * ctx,
696
- const char * path_session,
697
- const llama_token * tokens,
698
- size_t n_token_count);
699
- LLAMA_API DEPRECATED(bool llama_save_session_file(
700
- struct llama_context * ctx,
701
- const char * path_session,
702
- const llama_token * tokens,
703
- size_t n_token_count),
704
- "use llama_state_save_file instead");
705
-
706
- // Get the exact size needed to copy the KV cache of a single sequence
707
- LLAMA_API size_t llama_state_seq_get_size(
708
- struct llama_context * ctx,
709
- llama_seq_id seq_id);
710
-
711
- // Copy the KV cache of a single sequence into the specified buffer
712
- LLAMA_API size_t llama_state_seq_get_data(
713
- struct llama_context * ctx,
714
- uint8_t * dst,
715
- llama_seq_id seq_id);
716
-
717
- // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
718
- // Returns:
719
- // - Positive: Ok
720
- // - Zero: Failed to load
721
- LLAMA_API size_t llama_state_seq_set_data(
722
- struct llama_context * ctx,
723
- const uint8_t * src,
724
- llama_seq_id dest_seq_id);
725
-
726
- LLAMA_API size_t llama_state_seq_save_file(
727
- struct llama_context * ctx,
728
- const char * filepath,
729
- llama_seq_id seq_id,
730
- const llama_token * tokens,
731
- size_t n_token_count);
732
-
733
- LLAMA_API size_t llama_state_seq_load_file(
734
- struct llama_context * ctx,
735
- const char * filepath,
736
- llama_seq_id dest_seq_id,
737
- llama_token * tokens_out,
738
- size_t n_token_capacity,
739
- size_t * n_token_count_out);
740
-
741
- //
742
- // Decoding
743
- //
744
-
745
- // Return batch for single sequence of tokens starting at pos_0
746
- //
747
- // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
748
- //
749
- LLAMA_API struct llama_batch llama_batch_get_one(
750
- llama_token * tokens,
751
- int32_t n_tokens,
752
- llama_pos pos_0,
753
- llama_seq_id seq_id);
754
-
755
- // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
756
- // Each token can be assigned up to n_seq_max sequence ids
757
- // The batch has to be freed with llama_batch_free()
758
- // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
759
- // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
760
- // The rest of the llama_batch members are allocated with size n_tokens
761
- // All members are left uninitialized
762
- LLAMA_API struct llama_batch llama_batch_init(
763
- int32_t n_tokens,
764
- int32_t embd,
765
- int32_t n_seq_max);
766
-
767
- // Frees a batch of tokens allocated with llama_batch_init()
768
- LLAMA_API void llama_batch_free(struct llama_batch batch);
769
-
770
- // Positive return values does not mean a fatal error, but rather a warning.
771
- // 0 - success
772
- // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
773
- // < 0 - error
774
- LLAMA_API int32_t llama_decode(
775
- struct llama_context * ctx,
776
- struct llama_batch batch);
777
-
778
- // Set the number of threads used for decoding
779
- // n_threads is the number of threads used for generation (single token)
780
- // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
781
- LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
782
-
783
- // Get the number of threads used for generation of a single token.
784
- LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
785
-
786
- // Get the number of threads used for prompt and batch processing (multiple token).
787
- LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
788
-
789
- // Set whether the model is in embeddings model or not
790
- // If true, embeddings will be returned but logits will not
791
- LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
792
-
793
- // Set whether to use causal attention or not
794
- // If set to true, the model will only attend to the past tokens
795
- LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
796
-
797
- // Set abort callback
798
- LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
799
-
800
- // Wait until all computations are finished
801
- // This is automatically done when using one of the functions below to obtain the computation results
802
- // and is not necessary to call it explicitly in most cases
803
- LLAMA_API void llama_synchronize(struct llama_context * ctx);
804
-
805
- // Token logits obtained from the last call to llama_decode()
806
- // The logits for which llama_batch.logits[i] != 0 are stored contiguously
807
- // in the order they have appeared in the batch.
808
- // Rows: number of tokens for which llama_batch.logits[i] != 0
809
- // Cols: n_vocab
810
- LLAMA_API float * llama_get_logits(struct llama_context * ctx);
811
-
812
- // Logits for the ith token. For positive indices, Equivalent to:
813
- // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
814
- // Negative indicies can be used to access logits in reverse order, -1 is the last logit.
815
- // returns NULL for invalid ids.
816
- LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
817
-
818
- // Get all output token embeddings.
819
- // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model,
820
- // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously
821
- // in the order they have appeared in the batch.
822
- // shape: [n_outputs*n_embd]
823
- // Otherwise, returns NULL.
824
- LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
825
-
826
- // Get the embeddings for the ith token. For positive indices, Equivalent to:
827
- // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
828
- // Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding.
829
- // shape: [n_embd] (1-dimensional)
830
- // returns NULL for invalid ids.
831
- LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
832
-
833
- // Get the embeddings for a sequence id
834
- // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
835
- // shape: [n_embd] (1-dimensional)
836
- LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
837
-
838
- //
839
- // Vocab
840
- //
841
-
842
- LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
843
-
844
- LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
845
-
846
- LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token);
847
-
848
- // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
849
- LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
850
-
851
- // Identify if Token Id is a control token or a render-able token
852
- LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
853
-
854
- // Special tokens
855
- LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
856
- LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
857
- LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
858
- LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
859
- LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
860
-
861
- // Returns -1 if unknown, 1 for true or 0 for false.
862
- LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model);
863
-
864
- // Returns -1 if unknown, 1 for true or 0 for false.
865
- LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model);
866
-
867
- // Codellama infill tokens
868
- LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
869
- LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
870
- LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
871
- LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle
872
-
873
- //
874
- // Tokenization
875
- //
876
-
877
- /// @details Convert the provided text into tokens.
878
- /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
879
- /// @return Returns the number of tokens on success, no more than n_tokens_max
880
- /// @return Returns a negative number on failure - the number of tokens that would have been returned
881
- /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
882
- /// as plaintext. Does not insert a leading space.
883
- LLAMA_API int32_t llama_tokenize(
884
- const struct llama_model * model,
885
- const char * text,
886
- int32_t text_len,
887
- llama_token * tokens,
888
- int32_t n_tokens_max,
889
- bool add_special,
890
- bool parse_special);
891
-
892
- // Token Id -> Piece.
893
- // Uses the vocabulary in the provided context.
894
- // Does not write null terminator to the buffer.
895
- // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
896
- // @param special If true, special tokens are rendered in the output.
897
- LLAMA_API int32_t llama_token_to_piece(
898
- const struct llama_model * model,
899
- llama_token token,
900
- char * buf,
901
- int32_t length,
902
- bool special);
903
-
904
- /// Apply chat template. Inspired by hf apply_chat_template() on python.
905
- /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
906
- /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
907
- /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
908
- /// @param chat Pointer to a list of multiple llama_chat_message
909
- /// @param n_msg Number of llama_chat_message in this chat
910
- /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
911
- /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
912
- /// @param length The size of the allocated buffer
913
- /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
914
- LLAMA_API int32_t llama_chat_apply_template(
915
- const struct llama_model * model,
916
- const char * tmpl,
917
- const struct llama_chat_message * chat,
918
- size_t n_msg,
919
- bool add_ass,
920
- char * buf,
921
- int32_t length);
922
-
923
- //
924
- // Grammar
925
- //
926
-
927
- LLAMA_API struct llama_grammar * llama_grammar_init(
928
- const llama_grammar_element ** rules,
929
- size_t n_rules,
930
- size_t start_rule_index);
931
-
932
- LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
933
-
934
- LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
935
-
936
- //
937
- // Sampling functions
938
- //
939
-
940
- // Sets the current rng seed.
941
- LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
942
-
943
- /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
944
- /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
945
- LLAMA_API void llama_sample_repetition_penalties(
946
- struct llama_context * ctx,
947
- llama_token_data_array * candidates,
948
- const llama_token * last_tokens,
949
- size_t penalty_last_n,
950
- float penalty_repeat,
951
- float penalty_freq,
952
- float penalty_present);
953
-
954
- /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
955
- /// @param logits Logits extracted from the original generation context.
956
- /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
957
- /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
958
- LLAMA_API void llama_sample_apply_guidance(
959
- struct llama_context * ctx,
960
- float * logits,
961
- float * logits_guidance,
962
- float scale);
963
-
964
- /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
965
- LLAMA_API void llama_sample_softmax(
966
- struct llama_context * ctx,
967
- llama_token_data_array * candidates);
968
-
969
- /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
970
- LLAMA_API void llama_sample_top_k(
971
- struct llama_context * ctx,
972
- llama_token_data_array * candidates,
973
- int32_t k,
974
- size_t min_keep);
975
-
976
- /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
977
- LLAMA_API void llama_sample_top_p(
978
- struct llama_context * ctx,
979
- llama_token_data_array * candidates,
980
- float p,
981
- size_t min_keep);
982
-
983
- /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
984
- LLAMA_API void llama_sample_min_p(
985
- struct llama_context * ctx,
986
- llama_token_data_array * candidates,
987
- float p,
988
- size_t min_keep);
989
-
990
- /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
991
- LLAMA_API void llama_sample_tail_free(
992
- struct llama_context * ctx,
993
- llama_token_data_array * candidates,
994
- float z,
995
- size_t min_keep);
996
-
997
- /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
998
- LLAMA_API void llama_sample_typical(
999
- struct llama_context * ctx,
1000
- llama_token_data_array * candidates,
1001
- float p,
1002
- size_t min_keep);
1003
-
1004
- /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
1005
- LLAMA_API void llama_sample_entropy(
1006
- struct llama_context * ctx,
1007
- llama_token_data_array * candidates_p,
1008
- float min_temp,
1009
- float max_temp,
1010
- float exponent_val);
1011
-
1012
- LLAMA_API void llama_sample_temp(
1013
- struct llama_context * ctx,
1014
- llama_token_data_array * candidates,
1015
- float temp);
1016
-
1017
- /// @details Apply constraints from grammar
1018
- LLAMA_API void llama_sample_grammar(
1019
- struct llama_context * ctx,
1020
- llama_token_data_array * candidates,
1021
- const struct llama_grammar * grammar);
1022
-
1023
- /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
1024
- /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
1025
- /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
1026
- /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
1027
- /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
1028
- /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
1029
- LLAMA_API llama_token llama_sample_token_mirostat(
1030
- struct llama_context * ctx,
1031
- llama_token_data_array * candidates,
1032
- float tau,
1033
- float eta,
1034
- int32_t m,
1035
- float * mu);
1036
-
1037
- /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
1038
- /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
1039
- /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
1040
- /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
1041
- /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
1042
- LLAMA_API llama_token llama_sample_token_mirostat_v2(
1043
- struct llama_context * ctx,
1044
- llama_token_data_array * candidates,
1045
- float tau,
1046
- float eta,
1047
- float * mu);
1048
-
1049
- /// @details Selects the token with the highest probability.
1050
- /// Does not compute the token probabilities. Use llama_sample_softmax() instead.
1051
- LLAMA_API llama_token llama_sample_token_greedy(
1052
- struct llama_context * ctx,
1053
- llama_token_data_array * candidates);
1054
-
1055
- /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
1056
- LLAMA_API llama_token llama_sample_token(
1057
- struct llama_context * ctx,
1058
- llama_token_data_array * candidates);
1059
-
1060
- /// @details Accepts the sampled token into the grammar
1061
- LLAMA_API void llama_grammar_accept_token(
1062
- struct llama_context * ctx,
1063
- struct llama_grammar * grammar,
1064
- llama_token token);
1065
-
1066
- //
1067
- // Model split
1068
- //
1069
-
1070
- /// @details Build a split GGUF final path for this chunk.
1071
- /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
1072
- // Returns the split_path length.
1073
- LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
1074
-
1075
- /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
1076
- /// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
1077
- // Returns the split_prefix length.
1078
- LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
1079
-
1080
- // Performance information
1081
- LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
1082
-
1083
- LLAMA_API void llama_print_timings(struct llama_context * ctx);
1084
- LLAMA_API void llama_reset_timings(struct llama_context * ctx);
1085
-
1086
- // Print system information
1087
- LLAMA_API const char * llama_print_system_info(void);
1088
-
1089
- // Set callback for all future logging events.
1090
- // If this is not called, or NULL is supplied, everything is output on stderr.
1091
- LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
1092
-
1093
- LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
1094
-
1095
- #ifdef __cplusplus
1096
- }
1097
- #endif
1098
-
1099
- // Internal API to be implemented by llama.cpp and used by tests/benchmarks only
1100
- #ifdef LLAMA_API_INTERNAL
1101
-
1102
- #include <random>
1103
- #include <string>
1104
- #include <vector>
1105
-
1106
- struct ggml_tensor;
1107
-
1108
- struct llama_partial_utf8 {
1109
- uint32_t value; // bit value so far (unshifted)
1110
- int n_remain; // num bytes remaining; -1 indicates invalid sequence
1111
- };
1112
-
1113
- struct llama_grammar {
1114
- const std::vector<std::vector<llama_grammar_element>> rules;
1115
- std::vector<std::vector<const llama_grammar_element *>> stacks;
1116
-
1117
- // buffer for partially generated UTF-8 sequence from accepted tokens
1118
- llama_partial_utf8 partial_utf8;
1119
- };
1120
-
1121
- struct llama_grammar_candidate {
1122
- size_t index;
1123
- const uint32_t * code_points;
1124
- llama_partial_utf8 partial_utf8;
1125
- };
1126
-
1127
- const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
1128
- struct llama_context * ctx
1129
- );
1130
-
1131
- void llama_grammar_accept(
1132
- const std::vector<std::vector<llama_grammar_element>> & rules,
1133
- const std::vector<std::vector<const llama_grammar_element *>> & stacks,
1134
- const uint32_t chr,
1135
- std::vector<std::vector<const llama_grammar_element *>> & new_stacks);
1136
-
1137
- std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
1138
- const std::string & src,
1139
- llama_partial_utf8 partial_start);
1140
-
1141
- // Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
1142
- // This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
1143
- llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
1144
-
1145
- #endif // LLAMA_API_INTERNAL
1146
-
1147
- #endif // LLAMA_H