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/log.cpp CHANGED
@@ -1,401 +1,401 @@
1
- #include "log.h"
2
-
3
- #include <condition_variable>
4
- #include <cstdarg>
5
- #include <cstdio>
6
- #include <mutex>
7
- #include <sstream>
8
- #include <thread>
9
- #include <vector>
10
-
11
- int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
12
-
13
- void common_log_set_verbosity_thold(int verbosity) {
14
- common_log_verbosity_thold = verbosity;
15
- }
16
-
17
- #define LOG_COL_DEFAULT "\033[0m"
18
- #define LOG_COL_BOLD "\033[1m"
19
- #define LOG_COL_RED "\033[31m"
20
- #define LOG_COL_GREEN "\033[32m"
21
- #define LOG_COL_YELLOW "\033[33m"
22
- #define LOG_COL_BLUE "\033[34m"
23
- #define LOG_COL_MAGENTA "\033[35m"
24
- #define LOG_COL_CYAN "\033[36m"
25
- #define LOG_COL_WHITE "\033[37m"
26
-
27
- static int64_t t_us() {
28
- return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
29
- }
30
-
31
- // colors
32
- enum common_log_col : int {
33
- COMMON_LOG_COL_DEFAULT = 0,
34
- COMMON_LOG_COL_BOLD,
35
- COMMON_LOG_COL_RED,
36
- COMMON_LOG_COL_GREEN,
37
- COMMON_LOG_COL_YELLOW,
38
- COMMON_LOG_COL_BLUE,
39
- COMMON_LOG_COL_MAGENTA,
40
- COMMON_LOG_COL_CYAN,
41
- COMMON_LOG_COL_WHITE,
42
- };
43
-
44
- // disable colors by default
45
- static std::vector<const char *> g_col = {
46
- "",
47
- "",
48
- "",
49
- "",
50
- "",
51
- "",
52
- "",
53
- "",
54
- "",
55
- };
56
-
57
- struct common_log_entry {
58
- enum lm_ggml_log_level level;
59
-
60
- bool prefix;
61
-
62
- int64_t timestamp;
63
-
64
- std::vector<char> msg;
65
-
66
- // signals the worker thread to stop
67
- bool is_end;
68
-
69
- void print(FILE * file = nullptr) const {
70
- FILE * fcur = file;
71
- if (!fcur) {
72
- // stderr displays DBG messages only when their verbosity level is not higher than the threshold
73
- // these messages will still be logged to a file
74
- if (level == LM_GGML_LOG_LEVEL_DEBUG && common_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
75
- return;
76
- }
77
-
78
- fcur = stdout;
79
-
80
- if (level != LM_GGML_LOG_LEVEL_NONE) {
81
- fcur = stderr;
82
- }
83
- }
84
-
85
- if (level != LM_GGML_LOG_LEVEL_NONE && level != LM_GGML_LOG_LEVEL_CONT && prefix) {
86
- if (timestamp) {
87
- // [M.s.ms.us]
88
- fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
89
- g_col[COMMON_LOG_COL_BLUE],
90
- (int) (timestamp / 1000000 / 60),
91
- (int) (timestamp / 1000000 % 60),
92
- (int) (timestamp / 1000 % 1000),
93
- (int) (timestamp % 1000),
94
- g_col[COMMON_LOG_COL_DEFAULT]);
95
- }
96
-
97
- switch (level) {
98
- case LM_GGML_LOG_LEVEL_INFO: fprintf(fcur, "%sI %s", g_col[COMMON_LOG_COL_GREEN], g_col[COMMON_LOG_COL_DEFAULT]); break;
99
- case LM_GGML_LOG_LEVEL_WARN: fprintf(fcur, "%sW %s", g_col[COMMON_LOG_COL_MAGENTA], "" ); break;
100
- case LM_GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[COMMON_LOG_COL_RED], "" ); break;
101
- case LM_GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[COMMON_LOG_COL_YELLOW], "" ); break;
102
- default:
103
- break;
104
- }
105
- }
106
-
107
- fprintf(fcur, "%s", msg.data());
108
-
109
- if (level == LM_GGML_LOG_LEVEL_WARN || level == LM_GGML_LOG_LEVEL_ERROR || level == LM_GGML_LOG_LEVEL_DEBUG) {
110
- fprintf(fcur, "%s", g_col[COMMON_LOG_COL_DEFAULT]);
111
- }
112
-
113
- fflush(fcur);
114
- }
115
- };
116
-
117
- struct common_log {
118
- // default capacity - will be expanded if needed
119
- common_log() : common_log(256) {}
120
-
121
- common_log(size_t capacity) {
122
- file = nullptr;
123
- prefix = false;
124
- timestamps = false;
125
- running = false;
126
- t_start = t_us();
127
-
128
- // initial message size - will be expanded if longer messages arrive
129
- entries.resize(capacity);
130
- for (auto & entry : entries) {
131
- entry.msg.resize(256);
132
- }
133
-
134
- head = 0;
135
- tail = 0;
136
-
137
- resume();
138
- }
139
-
140
- ~common_log() {
141
- pause();
142
- if (file) {
143
- fclose(file);
144
- }
145
- }
146
-
147
- private:
148
- std::mutex mtx;
149
- std::thread thrd;
150
- std::condition_variable cv;
151
-
152
- FILE * file;
153
-
154
- bool prefix;
155
- bool timestamps;
156
- bool running;
157
-
158
- int64_t t_start;
159
-
160
- // ring buffer of entries
161
- std::vector<common_log_entry> entries;
162
- size_t head;
163
- size_t tail;
164
-
165
- // worker thread copies into this
166
- common_log_entry cur;
167
-
168
- public:
169
- void add(enum lm_ggml_log_level level, const char * fmt, va_list args) {
170
- std::lock_guard<std::mutex> lock(mtx);
171
-
172
- if (!running) {
173
- // discard messages while the worker thread is paused
174
- return;
175
- }
176
-
177
- auto & entry = entries[tail];
178
-
179
- {
180
- // cannot use args twice, so make a copy in case we need to expand the buffer
181
- va_list args_copy;
182
- va_copy(args_copy, args);
183
-
184
- #if 1
185
- const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
186
- if (n >= entry.msg.size()) {
187
- entry.msg.resize(n + 1);
188
- vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args_copy);
189
- }
190
- #else
191
- // hack for bolding arguments
192
-
193
- std::stringstream ss;
194
- for (int i = 0; fmt[i] != 0; i++) {
195
- if (fmt[i] == '%') {
196
- ss << LOG_COL_BOLD;
197
- while (fmt[i] != ' ' && fmt[i] != ')' && fmt[i] != ']' && fmt[i] != 0) ss << fmt[i++];
198
- ss << LOG_COL_DEFAULT;
199
- if (fmt[i] == 0) break;
200
- }
201
- ss << fmt[i];
202
- }
203
- const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
204
- if (n >= entry.msg.size()) {
205
- entry.msg.resize(n + 1);
206
- vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
207
- }
208
- #endif
209
- }
210
-
211
- entry.level = level;
212
- entry.prefix = prefix;
213
- entry.timestamp = 0;
214
- if (timestamps) {
215
- entry.timestamp = t_us() - t_start;
216
- }
217
- entry.is_end = false;
218
-
219
- tail = (tail + 1) % entries.size();
220
- if (tail == head) {
221
- // expand the buffer
222
- std::vector<common_log_entry> new_entries(2*entries.size());
223
-
224
- size_t new_tail = 0;
225
-
226
- do {
227
- new_entries[new_tail] = std::move(entries[head]);
228
-
229
- head = (head + 1) % entries.size();
230
- new_tail = (new_tail + 1);
231
- } while (head != tail);
232
-
233
- head = 0;
234
- tail = new_tail;
235
-
236
- for (size_t i = tail; i < new_entries.size(); i++) {
237
- new_entries[i].msg.resize(256);
238
- }
239
-
240
- entries = std::move(new_entries);
241
- }
242
-
243
- cv.notify_one();
244
- }
245
-
246
- void resume() {
247
- std::lock_guard<std::mutex> lock(mtx);
248
-
249
- if (running) {
250
- return;
251
- }
252
-
253
- running = true;
254
-
255
- thrd = std::thread([this]() {
256
- while (true) {
257
- {
258
- std::unique_lock<std::mutex> lock(mtx);
259
- cv.wait(lock, [this]() { return head != tail; });
260
-
261
- cur = entries[head];
262
-
263
- head = (head + 1) % entries.size();
264
- }
265
-
266
- if (cur.is_end) {
267
- break;
268
- }
269
-
270
- cur.print(); // stdout and stderr
271
-
272
- if (file) {
273
- cur.print(file);
274
- }
275
- }
276
- });
277
- }
278
-
279
- void pause() {
280
- {
281
- std::lock_guard<std::mutex> lock(mtx);
282
-
283
- if (!running) {
284
- return;
285
- }
286
-
287
- running = false;
288
-
289
- // push an entry to signal the worker thread to stop
290
- {
291
- auto & entry = entries[tail];
292
- entry.is_end = true;
293
-
294
- tail = (tail + 1) % entries.size();
295
- }
296
-
297
- cv.notify_one();
298
- }
299
-
300
- thrd.join();
301
- }
302
-
303
- void set_file(const char * path) {
304
- pause();
305
-
306
- if (file) {
307
- fclose(file);
308
- }
309
-
310
- if (path) {
311
- file = fopen(path, "w");
312
- } else {
313
- file = nullptr;
314
- }
315
-
316
- resume();
317
- }
318
-
319
- void set_colors(bool colors) {
320
- pause();
321
-
322
- if (colors) {
323
- g_col[COMMON_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
324
- g_col[COMMON_LOG_COL_BOLD] = LOG_COL_BOLD;
325
- g_col[COMMON_LOG_COL_RED] = LOG_COL_RED;
326
- g_col[COMMON_LOG_COL_GREEN] = LOG_COL_GREEN;
327
- g_col[COMMON_LOG_COL_YELLOW] = LOG_COL_YELLOW;
328
- g_col[COMMON_LOG_COL_BLUE] = LOG_COL_BLUE;
329
- g_col[COMMON_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
330
- g_col[COMMON_LOG_COL_CYAN] = LOG_COL_CYAN;
331
- g_col[COMMON_LOG_COL_WHITE] = LOG_COL_WHITE;
332
- } else {
333
- for (size_t i = 0; i < g_col.size(); i++) {
334
- g_col[i] = "";
335
- }
336
- }
337
-
338
- resume();
339
- }
340
-
341
- void set_prefix(bool prefix) {
342
- std::lock_guard<std::mutex> lock(mtx);
343
-
344
- this->prefix = prefix;
345
- }
346
-
347
- void set_timestamps(bool timestamps) {
348
- std::lock_guard<std::mutex> lock(mtx);
349
-
350
- this->timestamps = timestamps;
351
- }
352
- };
353
-
354
- //
355
- // public API
356
- //
357
-
358
- struct common_log * common_log_init() {
359
- return new common_log;
360
- }
361
-
362
- struct common_log * common_log_main() {
363
- static struct common_log log;
364
-
365
- return &log;
366
- }
367
-
368
- void common_log_pause(struct common_log * log) {
369
- log->pause();
370
- }
371
-
372
- void common_log_resume(struct common_log * log) {
373
- log->resume();
374
- }
375
-
376
- void common_log_free(struct common_log * log) {
377
- delete log;
378
- }
379
-
380
- void common_log_add(struct common_log * log, enum lm_ggml_log_level level, const char * fmt, ...) {
381
- va_list args;
382
- va_start(args, fmt);
383
- log->add(level, fmt, args);
384
- va_end(args);
385
- }
386
-
387
- void common_log_set_file(struct common_log * log, const char * file) {
388
- log->set_file(file);
389
- }
390
-
391
- void common_log_set_colors(struct common_log * log, bool colors) {
392
- log->set_colors(colors);
393
- }
394
-
395
- void common_log_set_prefix(struct common_log * log, bool prefix) {
396
- log->set_prefix(prefix);
397
- }
398
-
399
- void common_log_set_timestamps(struct common_log * log, bool timestamps) {
400
- log->set_timestamps(timestamps);
401
- }
1
+ #include "log.h"
2
+
3
+ #include <condition_variable>
4
+ #include <cstdarg>
5
+ #include <cstdio>
6
+ #include <mutex>
7
+ #include <sstream>
8
+ #include <thread>
9
+ #include <vector>
10
+
11
+ int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
12
+
13
+ void common_log_set_verbosity_thold(int verbosity) {
14
+ common_log_verbosity_thold = verbosity;
15
+ }
16
+
17
+ #define LOG_COL_DEFAULT "\033[0m"
18
+ #define LOG_COL_BOLD "\033[1m"
19
+ #define LOG_COL_RED "\033[31m"
20
+ #define LOG_COL_GREEN "\033[32m"
21
+ #define LOG_COL_YELLOW "\033[33m"
22
+ #define LOG_COL_BLUE "\033[34m"
23
+ #define LOG_COL_MAGENTA "\033[35m"
24
+ #define LOG_COL_CYAN "\033[36m"
25
+ #define LOG_COL_WHITE "\033[37m"
26
+
27
+ static int64_t t_us() {
28
+ return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
29
+ }
30
+
31
+ // colors
32
+ enum common_log_col : int {
33
+ COMMON_LOG_COL_DEFAULT = 0,
34
+ COMMON_LOG_COL_BOLD,
35
+ COMMON_LOG_COL_RED,
36
+ COMMON_LOG_COL_GREEN,
37
+ COMMON_LOG_COL_YELLOW,
38
+ COMMON_LOG_COL_BLUE,
39
+ COMMON_LOG_COL_MAGENTA,
40
+ COMMON_LOG_COL_CYAN,
41
+ COMMON_LOG_COL_WHITE,
42
+ };
43
+
44
+ // disable colors by default
45
+ static std::vector<const char *> g_col = {
46
+ "",
47
+ "",
48
+ "",
49
+ "",
50
+ "",
51
+ "",
52
+ "",
53
+ "",
54
+ "",
55
+ };
56
+
57
+ struct common_log_entry {
58
+ enum lm_ggml_log_level level;
59
+
60
+ bool prefix;
61
+
62
+ int64_t timestamp;
63
+
64
+ std::vector<char> msg;
65
+
66
+ // signals the worker thread to stop
67
+ bool is_end;
68
+
69
+ void print(FILE * file = nullptr) const {
70
+ FILE * fcur = file;
71
+ if (!fcur) {
72
+ // stderr displays DBG messages only when their verbosity level is not higher than the threshold
73
+ // these messages will still be logged to a file
74
+ if (level == LM_GGML_LOG_LEVEL_DEBUG && common_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
75
+ return;
76
+ }
77
+
78
+ fcur = stdout;
79
+
80
+ if (level != LM_GGML_LOG_LEVEL_NONE) {
81
+ fcur = stderr;
82
+ }
83
+ }
84
+
85
+ if (level != LM_GGML_LOG_LEVEL_NONE && level != LM_GGML_LOG_LEVEL_CONT && prefix) {
86
+ if (timestamp) {
87
+ // [M.s.ms.us]
88
+ fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
89
+ g_col[COMMON_LOG_COL_BLUE],
90
+ (int) (timestamp / 1000000 / 60),
91
+ (int) (timestamp / 1000000 % 60),
92
+ (int) (timestamp / 1000 % 1000),
93
+ (int) (timestamp % 1000),
94
+ g_col[COMMON_LOG_COL_DEFAULT]);
95
+ }
96
+
97
+ switch (level) {
98
+ case LM_GGML_LOG_LEVEL_INFO: fprintf(fcur, "%sI %s", g_col[COMMON_LOG_COL_GREEN], g_col[COMMON_LOG_COL_DEFAULT]); break;
99
+ case LM_GGML_LOG_LEVEL_WARN: fprintf(fcur, "%sW %s", g_col[COMMON_LOG_COL_MAGENTA], "" ); break;
100
+ case LM_GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[COMMON_LOG_COL_RED], "" ); break;
101
+ case LM_GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[COMMON_LOG_COL_YELLOW], "" ); break;
102
+ default:
103
+ break;
104
+ }
105
+ }
106
+
107
+ fprintf(fcur, "%s", msg.data());
108
+
109
+ if (level == LM_GGML_LOG_LEVEL_WARN || level == LM_GGML_LOG_LEVEL_ERROR || level == LM_GGML_LOG_LEVEL_DEBUG) {
110
+ fprintf(fcur, "%s", g_col[COMMON_LOG_COL_DEFAULT]);
111
+ }
112
+
113
+ fflush(fcur);
114
+ }
115
+ };
116
+
117
+ struct common_log {
118
+ // default capacity - will be expanded if needed
119
+ common_log() : common_log(256) {}
120
+
121
+ common_log(size_t capacity) {
122
+ file = nullptr;
123
+ prefix = false;
124
+ timestamps = false;
125
+ running = false;
126
+ t_start = t_us();
127
+
128
+ // initial message size - will be expanded if longer messages arrive
129
+ entries.resize(capacity);
130
+ for (auto & entry : entries) {
131
+ entry.msg.resize(256);
132
+ }
133
+
134
+ head = 0;
135
+ tail = 0;
136
+
137
+ resume();
138
+ }
139
+
140
+ ~common_log() {
141
+ pause();
142
+ if (file) {
143
+ fclose(file);
144
+ }
145
+ }
146
+
147
+ private:
148
+ std::mutex mtx;
149
+ std::thread thrd;
150
+ std::condition_variable cv;
151
+
152
+ FILE * file;
153
+
154
+ bool prefix;
155
+ bool timestamps;
156
+ bool running;
157
+
158
+ int64_t t_start;
159
+
160
+ // ring buffer of entries
161
+ std::vector<common_log_entry> entries;
162
+ size_t head;
163
+ size_t tail;
164
+
165
+ // worker thread copies into this
166
+ common_log_entry cur;
167
+
168
+ public:
169
+ void add(enum lm_ggml_log_level level, const char * fmt, va_list args) {
170
+ std::lock_guard<std::mutex> lock(mtx);
171
+
172
+ if (!running) {
173
+ // discard messages while the worker thread is paused
174
+ return;
175
+ }
176
+
177
+ auto & entry = entries[tail];
178
+
179
+ {
180
+ // cannot use args twice, so make a copy in case we need to expand the buffer
181
+ va_list args_copy;
182
+ va_copy(args_copy, args);
183
+
184
+ #if 1
185
+ const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
186
+ if (n >= entry.msg.size()) {
187
+ entry.msg.resize(n + 1);
188
+ vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args_copy);
189
+ }
190
+ #else
191
+ // hack for bolding arguments
192
+
193
+ std::stringstream ss;
194
+ for (int i = 0; fmt[i] != 0; i++) {
195
+ if (fmt[i] == '%') {
196
+ ss << LOG_COL_BOLD;
197
+ while (fmt[i] != ' ' && fmt[i] != ')' && fmt[i] != ']' && fmt[i] != 0) ss << fmt[i++];
198
+ ss << LOG_COL_DEFAULT;
199
+ if (fmt[i] == 0) break;
200
+ }
201
+ ss << fmt[i];
202
+ }
203
+ const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
204
+ if (n >= entry.msg.size()) {
205
+ entry.msg.resize(n + 1);
206
+ vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
207
+ }
208
+ #endif
209
+ }
210
+
211
+ entry.level = level;
212
+ entry.prefix = prefix;
213
+ entry.timestamp = 0;
214
+ if (timestamps) {
215
+ entry.timestamp = t_us() - t_start;
216
+ }
217
+ entry.is_end = false;
218
+
219
+ tail = (tail + 1) % entries.size();
220
+ if (tail == head) {
221
+ // expand the buffer
222
+ std::vector<common_log_entry> new_entries(2*entries.size());
223
+
224
+ size_t new_tail = 0;
225
+
226
+ do {
227
+ new_entries[new_tail] = std::move(entries[head]);
228
+
229
+ head = (head + 1) % entries.size();
230
+ new_tail = (new_tail + 1);
231
+ } while (head != tail);
232
+
233
+ head = 0;
234
+ tail = new_tail;
235
+
236
+ for (size_t i = tail; i < new_entries.size(); i++) {
237
+ new_entries[i].msg.resize(256);
238
+ }
239
+
240
+ entries = std::move(new_entries);
241
+ }
242
+
243
+ cv.notify_one();
244
+ }
245
+
246
+ void resume() {
247
+ std::lock_guard<std::mutex> lock(mtx);
248
+
249
+ if (running) {
250
+ return;
251
+ }
252
+
253
+ running = true;
254
+
255
+ thrd = std::thread([this]() {
256
+ while (true) {
257
+ {
258
+ std::unique_lock<std::mutex> lock(mtx);
259
+ cv.wait(lock, [this]() { return head != tail; });
260
+
261
+ cur = entries[head];
262
+
263
+ head = (head + 1) % entries.size();
264
+ }
265
+
266
+ if (cur.is_end) {
267
+ break;
268
+ }
269
+
270
+ cur.print(); // stdout and stderr
271
+
272
+ if (file) {
273
+ cur.print(file);
274
+ }
275
+ }
276
+ });
277
+ }
278
+
279
+ void pause() {
280
+ {
281
+ std::lock_guard<std::mutex> lock(mtx);
282
+
283
+ if (!running) {
284
+ return;
285
+ }
286
+
287
+ running = false;
288
+
289
+ // push an entry to signal the worker thread to stop
290
+ {
291
+ auto & entry = entries[tail];
292
+ entry.is_end = true;
293
+
294
+ tail = (tail + 1) % entries.size();
295
+ }
296
+
297
+ cv.notify_one();
298
+ }
299
+
300
+ thrd.join();
301
+ }
302
+
303
+ void set_file(const char * path) {
304
+ pause();
305
+
306
+ if (file) {
307
+ fclose(file);
308
+ }
309
+
310
+ if (path) {
311
+ file = fopen(path, "w");
312
+ } else {
313
+ file = nullptr;
314
+ }
315
+
316
+ resume();
317
+ }
318
+
319
+ void set_colors(bool colors) {
320
+ pause();
321
+
322
+ if (colors) {
323
+ g_col[COMMON_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
324
+ g_col[COMMON_LOG_COL_BOLD] = LOG_COL_BOLD;
325
+ g_col[COMMON_LOG_COL_RED] = LOG_COL_RED;
326
+ g_col[COMMON_LOG_COL_GREEN] = LOG_COL_GREEN;
327
+ g_col[COMMON_LOG_COL_YELLOW] = LOG_COL_YELLOW;
328
+ g_col[COMMON_LOG_COL_BLUE] = LOG_COL_BLUE;
329
+ g_col[COMMON_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
330
+ g_col[COMMON_LOG_COL_CYAN] = LOG_COL_CYAN;
331
+ g_col[COMMON_LOG_COL_WHITE] = LOG_COL_WHITE;
332
+ } else {
333
+ for (size_t i = 0; i < g_col.size(); i++) {
334
+ g_col[i] = "";
335
+ }
336
+ }
337
+
338
+ resume();
339
+ }
340
+
341
+ void set_prefix(bool prefix) {
342
+ std::lock_guard<std::mutex> lock(mtx);
343
+
344
+ this->prefix = prefix;
345
+ }
346
+
347
+ void set_timestamps(bool timestamps) {
348
+ std::lock_guard<std::mutex> lock(mtx);
349
+
350
+ this->timestamps = timestamps;
351
+ }
352
+ };
353
+
354
+ //
355
+ // public API
356
+ //
357
+
358
+ struct common_log * common_log_init() {
359
+ return new common_log;
360
+ }
361
+
362
+ struct common_log * common_log_main() {
363
+ static struct common_log log;
364
+
365
+ return &log;
366
+ }
367
+
368
+ void common_log_pause(struct common_log * log) {
369
+ log->pause();
370
+ }
371
+
372
+ void common_log_resume(struct common_log * log) {
373
+ log->resume();
374
+ }
375
+
376
+ void common_log_free(struct common_log * log) {
377
+ delete log;
378
+ }
379
+
380
+ void common_log_add(struct common_log * log, enum lm_ggml_log_level level, const char * fmt, ...) {
381
+ va_list args;
382
+ va_start(args, fmt);
383
+ log->add(level, fmt, args);
384
+ va_end(args);
385
+ }
386
+
387
+ void common_log_set_file(struct common_log * log, const char * file) {
388
+ log->set_file(file);
389
+ }
390
+
391
+ void common_log_set_colors(struct common_log * log, bool colors) {
392
+ log->set_colors(colors);
393
+ }
394
+
395
+ void common_log_set_prefix(struct common_log * log, bool prefix) {
396
+ log->set_prefix(prefix);
397
+ }
398
+
399
+ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
400
+ log->set_timestamps(timestamps);
401
+ }