cui-llama.rn 1.3.6 → 1.4.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (101) hide show
  1. package/README.md +22 -1
  2. package/android/src/main/CMakeLists.txt +25 -26
  3. package/android/src/main/java/com/rnllama/LlamaContext.java +31 -9
  4. package/android/src/main/java/com/rnllama/RNLlama.java +98 -0
  5. package/android/src/main/jni-utils.h +94 -0
  6. package/android/src/main/jni.cpp +133 -63
  7. package/android/src/newarch/java/com/rnllama/RNLlamaModule.java +15 -0
  8. package/android/src/oldarch/java/com/rnllama/RNLlamaModule.java +15 -0
  9. package/cpp/common.cpp +2085 -1982
  10. package/cpp/common.h +696 -664
  11. package/cpp/ggml-alloc.c +1042 -1037
  12. package/cpp/ggml-backend-impl.h +255 -256
  13. package/cpp/ggml-backend-reg.cpp +582 -582
  14. package/cpp/ggml-backend.cpp +2002 -2002
  15. package/cpp/ggml-backend.h +354 -352
  16. package/cpp/ggml-common.h +1853 -1853
  17. package/cpp/ggml-cpp.h +39 -39
  18. package/cpp/ggml-cpu-aarch64.cpp +4247 -4247
  19. package/cpp/ggml-cpu-aarch64.h +8 -8
  20. package/cpp/ggml-cpu-impl.h +386 -386
  21. package/cpp/ggml-cpu-quants.c +10920 -10839
  22. package/cpp/ggml-cpu-traits.cpp +36 -36
  23. package/cpp/ggml-cpu-traits.h +38 -38
  24. package/cpp/ggml-cpu.c +14391 -14122
  25. package/cpp/ggml-cpu.cpp +635 -627
  26. package/cpp/ggml-cpu.h +135 -135
  27. package/cpp/ggml-impl.h +567 -567
  28. package/cpp/ggml-metal-impl.h +288 -0
  29. package/cpp/ggml-metal.m +4884 -4884
  30. package/cpp/ggml-opt.cpp +854 -0
  31. package/cpp/ggml-opt.h +216 -0
  32. package/cpp/ggml-quants.c +5238 -5238
  33. package/cpp/ggml-threading.h +14 -14
  34. package/cpp/ggml.c +6514 -6448
  35. package/cpp/ggml.h +2194 -2163
  36. package/cpp/gguf.cpp +1329 -1325
  37. package/cpp/gguf.h +202 -202
  38. package/cpp/json-schema-to-grammar.cpp +1045 -1045
  39. package/cpp/json-schema-to-grammar.h +8 -8
  40. package/cpp/json.hpp +24766 -24766
  41. package/cpp/llama-adapter.cpp +347 -346
  42. package/cpp/llama-adapter.h +74 -73
  43. package/cpp/llama-arch.cpp +1487 -1434
  44. package/cpp/llama-arch.h +400 -395
  45. package/cpp/llama-batch.cpp +368 -368
  46. package/cpp/llama-batch.h +88 -88
  47. package/cpp/llama-chat.cpp +578 -567
  48. package/cpp/llama-chat.h +52 -51
  49. package/cpp/llama-context.cpp +1775 -1771
  50. package/cpp/llama-context.h +128 -128
  51. package/cpp/llama-cparams.cpp +1 -1
  52. package/cpp/llama-cparams.h +37 -37
  53. package/cpp/llama-cpp.h +30 -30
  54. package/cpp/llama-grammar.cpp +1139 -1139
  55. package/cpp/llama-grammar.h +143 -143
  56. package/cpp/llama-hparams.cpp +71 -71
  57. package/cpp/llama-hparams.h +139 -140
  58. package/cpp/llama-impl.cpp +167 -167
  59. package/cpp/llama-impl.h +61 -61
  60. package/cpp/llama-kv-cache.cpp +718 -718
  61. package/cpp/llama-kv-cache.h +218 -218
  62. package/cpp/llama-mmap.cpp +590 -589
  63. package/cpp/llama-mmap.h +67 -67
  64. package/cpp/llama-model-loader.cpp +1124 -1011
  65. package/cpp/llama-model-loader.h +167 -158
  66. package/cpp/llama-model.cpp +3997 -2202
  67. package/cpp/llama-model.h +370 -391
  68. package/cpp/llama-sampling.cpp +2408 -2406
  69. package/cpp/llama-sampling.h +32 -48
  70. package/cpp/llama-vocab.cpp +3247 -1982
  71. package/cpp/llama-vocab.h +125 -182
  72. package/cpp/llama.cpp +10077 -12544
  73. package/cpp/llama.h +1323 -1285
  74. package/cpp/log.cpp +401 -401
  75. package/cpp/log.h +121 -121
  76. package/cpp/rn-llama.hpp +123 -116
  77. package/cpp/sampling.cpp +505 -500
  78. package/cpp/sgemm.cpp +2597 -2597
  79. package/cpp/sgemm.h +14 -14
  80. package/cpp/speculative.cpp +277 -274
  81. package/cpp/speculative.h +28 -28
  82. package/cpp/unicode.cpp +2 -3
  83. package/ios/RNLlama.mm +47 -0
  84. package/ios/RNLlamaContext.h +3 -1
  85. package/ios/RNLlamaContext.mm +71 -14
  86. package/jest/mock.js +15 -3
  87. package/lib/commonjs/NativeRNLlama.js.map +1 -1
  88. package/lib/commonjs/index.js +33 -37
  89. package/lib/commonjs/index.js.map +1 -1
  90. package/lib/module/NativeRNLlama.js.map +1 -1
  91. package/lib/module/index.js +31 -35
  92. package/lib/module/index.js.map +1 -1
  93. package/lib/typescript/NativeRNLlama.d.ts +26 -6
  94. package/lib/typescript/NativeRNLlama.d.ts.map +1 -1
  95. package/lib/typescript/index.d.ts +21 -36
  96. package/lib/typescript/index.d.ts.map +1 -1
  97. package/llama-rn.podspec +4 -18
  98. package/package.json +2 -3
  99. package/src/NativeRNLlama.ts +32 -13
  100. package/src/index.ts +52 -47
  101. package/cpp/llama.cpp.rej +0 -23
package/cpp/ggml-metal-impl.h ADDED
@@ -0,0 +1,288 @@
1
+ #ifndef GGML_METAL_IMPL
2
+ #define GGML_METAL_IMPL
3
+
4
+ // kernel argument structs
5
+ //
6
+ // - element counters (e.g. ne00) typically use int32_t to reduce register usage
7
+ // however, be careful from int overflows when using those in the kernel implementation
8
+ //
9
+ // - strides (e.g. nb00) use uint64_t
10
+
11
+ typedef struct {
12
+ int32_t ne00;
13
+ int32_t ne01;
14
+ int32_t ne02;
15
+ int32_t ne03;
16
+ uint64_t nb00;
17
+ uint64_t nb01;
18
+ uint64_t nb02;
19
+ uint64_t nb03;
20
+ int32_t ne10;
21
+ int32_t ne11;
22
+ int32_t ne12;
23
+ int32_t ne13;
24
+ uint64_t nb10;
25
+ uint64_t nb11;
26
+ uint64_t nb12;
27
+ uint64_t nb13;
28
+ int32_t ne0;
29
+ int32_t ne1;
30
+ int32_t ne2;
31
+ int32_t ne3;
32
+ uint64_t nb0;
33
+ uint64_t nb1;
34
+ uint64_t nb2;
35
+ uint64_t nb3;
36
+ int32_t dim;
37
+ } ggml_metal_kargs_concat;
38
+
39
+ typedef struct {
40
+ int32_t ne00;
41
+ int32_t ne01;
42
+ int32_t ne02;
43
+ int32_t ne03;
44
+ uint64_t nb00;
45
+ uint64_t nb01;
46
+ uint64_t nb02;
47
+ uint64_t nb03;
48
+ int32_t ne10;
49
+ int32_t ne11;
50
+ int32_t ne12;
51
+ int32_t ne13;
52
+ uint64_t nb10;
53
+ uint64_t nb11;
54
+ uint64_t nb12;
55
+ uint64_t nb13;
56
+ int32_t ne0;
57
+ int32_t ne1;
58
+ int32_t ne2;
59
+ int32_t ne3;
60
+ uint64_t nb0;
61
+ uint64_t nb1;
62
+ uint64_t nb2;
63
+ uint64_t nb3;
64
+ uint64_t offs;
65
+ } ggml_metal_kargs_bin;
66
+
67
+ typedef struct {
68
+ int32_t ne00;
69
+ int32_t ne01;
70
+ int32_t ne02;
71
+ int32_t ne03;
72
+ uint64_t nb00;
73
+ uint64_t nb01;
74
+ uint64_t nb02;
75
+ uint64_t nb03;
76
+ int32_t ne0;
77
+ int32_t ne1;
78
+ int32_t ne2;
79
+ int32_t ne3;
80
+ uint64_t nb0;
81
+ uint64_t nb1;
82
+ uint64_t nb2;
83
+ uint64_t nb3;
84
+ } ggml_metal_kargs_repeat;
85
+
86
+ typedef struct {
87
+ int64_t ne00;
88
+ int64_t ne01;
89
+ int64_t ne02;
90
+ int64_t ne03;
91
+ uint64_t nb00;
92
+ uint64_t nb01;
93
+ uint64_t nb02;
94
+ uint64_t nb03;
95
+ int64_t ne0;
96
+ int64_t ne1;
97
+ int64_t ne2;
98
+ int64_t ne3;
99
+ uint64_t nb0;
100
+ uint64_t nb1;
101
+ uint64_t nb2;
102
+ uint64_t nb3;
103
+ } ggml_metal_kargs_cpy;
104
+
105
+ typedef struct {
106
+ int64_t ne10;
107
+ int64_t ne11;
108
+ int64_t ne12;
109
+ uint64_t nb10;
110
+ uint64_t nb11;
111
+ uint64_t nb12;
112
+ uint64_t nb13;
113
+ uint64_t nb1;
114
+ uint64_t nb2;
115
+ uint64_t nb3;
116
+ uint64_t offs;
117
+ bool inplace;
118
+ } ggml_metal_kargs_set;
119
+
120
+ typedef struct {
121
+ int32_t ne00;
122
+ int32_t ne01;
123
+ int32_t ne02;
124
+ int32_t ne03;
125
+ uint64_t nb00;
126
+ uint64_t nb01;
127
+ uint64_t nb02;
128
+ uint64_t nb03;
129
+ int32_t ne0;
130
+ int32_t ne1;
131
+ int32_t ne2;
132
+ int32_t ne3;
133
+ uint64_t nb0;
134
+ uint64_t nb1;
135
+ uint64_t nb2;
136
+ uint64_t nb3;
137
+ int32_t n_past;
138
+ int32_t n_dims;
139
+ int32_t n_ctx_orig;
140
+ float freq_base;
141
+ float freq_scale;
142
+ float ext_factor;
143
+ float attn_factor;
144
+ float beta_fast;
145
+ float beta_slow;
146
+ } ggml_metal_kargs_rope;
147
+
148
+ typedef struct {
149
+ int32_t ne01;
150
+ int32_t ne02;
151
+ int32_t ne03;
152
+ uint64_t nb01;
153
+ uint64_t nb02;
154
+ uint64_t nb03;
155
+ int32_t ne11;
156
+ int32_t ne_12_2; // assume K and V are same shape
157
+ int32_t ne_12_3;
158
+ uint64_t nb_12_1;
159
+ uint64_t nb_12_2;
160
+ uint64_t nb_12_3;
161
+ uint64_t nb31;
162
+ int32_t ne1;
163
+ int32_t ne2;
164
+ float scale;
165
+ float max_bias;
166
+ float m0;
167
+ float m1;
168
+ uint16_t n_head_log2;
169
+ float logit_softcap;
170
+ } ggml_metal_kargs_flash_attn_ext;
171
+
172
+ typedef struct {
173
+ int32_t ne00;
174
+ int32_t ne02;
175
+ uint64_t nb01;
176
+ uint64_t nb02;
177
+ uint64_t nb03;
178
+ int32_t ne12;
179
+ uint64_t nb10;
180
+ uint64_t nb11;
181
+ uint64_t nb12;
182
+ uint64_t nb13;
183
+ int32_t ne0;
184
+ int32_t ne1;
185
+ int16_t r2;
186
+ int16_t r3;
187
+ } ggml_metal_kargs_mul_mm;
188
+
189
+ typedef struct {
190
+ int32_t ne00;
191
+ int32_t ne01;
192
+ int32_t ne02;
193
+ uint64_t nb00;
194
+ uint64_t nb01;
195
+ uint64_t nb02;
196
+ uint64_t nb03;
197
+ int32_t ne10;
198
+ int32_t ne11;
199
+ int32_t ne12;
200
+ uint64_t nb10;
201
+ uint64_t nb11;
202
+ uint64_t nb12;
203
+ uint64_t nb13;
204
+ int32_t ne0;
205
+ int32_t ne1;
206
+ int16_t r2;
207
+ int16_t r3;
208
+ } ggml_metal_kargs_mul_mv;
209
+
210
+ typedef struct {
211
+ int32_t ne00;
212
+ int32_t ne01;
213
+ int32_t ne02;
214
+ uint64_t nb00;
215
+ uint64_t nb01;
216
+ uint64_t nb02;
217
+ uint64_t nb03;
218
+ int32_t ne10;
219
+ int32_t ne11;
220
+ int32_t ne12;
221
+ uint64_t nb10;
222
+ uint64_t nb11;
223
+ uint64_t nb12;
224
+ uint64_t nb13;
225
+ int32_t ne0;
226
+ int32_t ne1;
227
+ int16_t r2;
228
+ int16_t r3;
229
+ int16_t nsg;
230
+ int16_t nxpsg;
231
+ int16_t r1ptg;
232
+ } ggml_metal_kargs_mul_mv_ext;
233
+
234
+ typedef struct {
235
+ int32_t nei0;
236
+ int32_t nei1;
237
+ uint64_t nbi1;
238
+ int32_t ne00;
239
+ int32_t ne02;
240
+ uint64_t nb01;
241
+ uint64_t nb02;
242
+ int32_t ne11;
243
+ int32_t ne12;
244
+ int32_t ne13;
245
+ uint64_t nb10;
246
+ uint64_t nb11;
247
+ uint64_t nb12;
248
+ int32_t ne0;
249
+ int32_t ne1;
250
+ } ggml_metal_kargs_mul_mm_id;
251
+
252
+ typedef struct {
253
+ int32_t nei0;
254
+ int32_t nei1;
255
+ uint64_t nbi1;
256
+ int32_t ne00;
257
+ int32_t ne01;
258
+ int32_t ne02;
259
+ uint64_t nb00;
260
+ uint64_t nb01;
261
+ uint64_t nb02;
262
+ int32_t ne10;
263
+ int32_t ne11;
264
+ int32_t ne12;
265
+ int32_t ne13;
266
+ uint64_t nb10;
267
+ uint64_t nb11;
268
+ uint64_t nb12;
269
+ int32_t ne0;
270
+ int32_t ne1;
271
+ uint64_t nb1;
272
+ } ggml_metal_kargs_mul_mv_id;
273
+
274
+ typedef struct {
275
+ int32_t ne00;
276
+ int32_t ne00_4;
277
+ uint64_t nb01;
278
+ float eps;
279
+ } ggml_metal_kargs_norm;
280
+
281
+ typedef struct {
282
+ int32_t ne00;
283
+ int32_t ne00_4;
284
+ uint64_t nb01;
285
+ float eps;
286
+ } ggml_metal_kargs_rms_norm;
287
+
288
+ #endif // GGML_METAL_IMPL