1#include "ggml.h"
2#include "gguf.h"
3
4#include "common.h"
5#include "log.h"
6#include "llama.h"
7#include "sampling.h"
8#include "unicode.h"
9
10#include <algorithm>
11#include <cinttypes>
12#include <climits>
13#include <cmath>
14#include <chrono>
15#include <cstdarg>
16#include <cstring>
17#include <ctime>
18#include <filesystem>
19#include <fstream>
20#include <iostream>
21#include <iterator>
22#include <regex>
23#include <sstream>
24#include <string>
25#include <thread>
26#include <unordered_set>
27#include <vector>
28
29#if defined(__APPLE__) && defined(__MACH__)
30#include <sys/types.h>
31#include <sys/sysctl.h>
32#endif
33
34#if defined(_WIN32)
35#define WIN32_LEAN_AND_MEAN
36#ifndef NOMINMAX
37# define NOMINMAX
38#endif
39#include <locale>
40#include <windows.h>
41#include <string.h>
42#include <fcntl.h>
43#include <io.h>
44#else
45#include <sys/ioctl.h>
46#include <sys/stat.h>
47#include <unistd.h>
48#endif
49
50#if defined(__linux__)
51#include <sys/types.h>
52#include <pwd.h>
53#endif
54
55#if defined(_MSC_VER)
56#pragma warning(disable: 4244 4267) // possible loss of data
57#endif
58
59common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
60
61common_time_meas::~common_time_meas() {
62 if (t_start_us >= 0) {
63 t_acc += ggml_time_us() - t_start_us;
64 }
65}
66
67//
68// CPU utils
69//
70
71int32_t cpu_get_num_physical_cores() {
72#ifdef __linux__
73 // enumerate the set of thread siblings, num entries is num cores
74 std::unordered_set<std::string> siblings;
75 for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
76 std::ifstream thread_siblings("/sys/devices/system/cpu/cpu"
77 + std::to_string(cpu) + "/topology/thread_siblings");
78 if (!thread_siblings.is_open()) {
79 break; // no more cpus
80 }
81 std::string line;
82 if (std::getline(thread_siblings, line)) {
83 siblings.insert(line);
84 }
85 }
86 if (!siblings.empty()) {
87 return static_cast<int32_t>(siblings.size());
88 }
89#elif defined(__APPLE__) && defined(__MACH__)
90 int32_t num_physical_cores;
91 size_t len = sizeof(num_physical_cores);
92 int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
93 if (result == 0) {
94 return num_physical_cores;
95 }
96 result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
97 if (result == 0) {
98 return num_physical_cores;
99 }
100#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
101 // TODO: windows + arm64 + mingw64
102 unsigned int n_threads_win = std::thread::hardware_concurrency();
103 unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4;
104
105 DWORD buffer_size = 0;
106 if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
107 if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
108 return default_threads;
109 }
110 }
111
112 std::vector<char> buffer(buffer_size);
113 if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
114 return default_threads;
115 }
116
117 int32_t num_physical_cores = 0;
118 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
119 while (buffer_size > 0) {
120 if (info->Relationship == RelationProcessorCore) {
121 num_physical_cores += info->Processor.GroupCount;
122 }
123 buffer_size -= info->Size;
124 info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
125 }
126
127 return num_physical_cores > 0 ? num_physical_cores : default_threads;
128#endif
129 unsigned int n_threads = std::thread::hardware_concurrency();
130 return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
131}
132
133#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
134#include <pthread.h>
135
136static void cpuid(unsigned leaf, unsigned subleaf,
137 unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx) {
138 __asm__("movq\t%%rbx,%%rsi\n\t"
139 "cpuid\n\t"
140 "xchgq\t%%rbx,%%rsi"
141 : "=a"(*eax), "=S"(*ebx), "=c"(*ecx), "=d"(*edx)
142 : "0"(leaf), "2"(subleaf));
143}
144
145static int pin_cpu(int cpu) {
146 cpu_set_t mask;
147 CPU_ZERO(&mask);
148 CPU_SET(cpu, &mask);
149 return pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
150}
151
152static bool is_hybrid_cpu(void) {
153 unsigned eax, ebx, ecx, edx;
154 cpuid(7, 0, &eax, &ebx, &ecx, &edx);
155 return !!(edx & (1u << 15));
156}
157
158static bool is_running_on_efficiency_core(void) {
159 unsigned eax, ebx, ecx, edx;
160 cpuid(0x1a, 0, &eax, &ebx, &ecx, &edx);
161 int intel_atom = 0x20;
162 int core_type = (eax & 0xff000000u) >> 24;
163 return core_type == intel_atom;
164}
165
166static int cpu_count_math_cpus(int n_cpu) {
167 int result = 0;
168 for (int cpu = 0; cpu < n_cpu; ++cpu) {
169 if (pin_cpu(cpu)) {
170 return -1;
171 }
172 if (is_running_on_efficiency_core()) {
173 continue; // efficiency cores harm lockstep threading
174 }
175 ++cpu; // hyperthreading isn't useful for linear algebra
176 ++result;
177 }
178 return result;
179}
180
181#endif // __x86_64__ && __linux__
182
183/**
184 * Returns number of CPUs on system that are useful for math.
185 */
186int32_t cpu_get_num_math() {
187#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
188 int n_cpu = sysconf(_SC_NPROCESSORS_ONLN);
189 if (n_cpu < 1) {
190 return cpu_get_num_physical_cores();
191 }
192 if (is_hybrid_cpu()) {
193 cpu_set_t affinity;
194 if (!pthread_getaffinity_np(pthread_self(), sizeof(affinity), &affinity)) {
195 int result = cpu_count_math_cpus(n_cpu);
196 pthread_setaffinity_np(pthread_self(), sizeof(affinity), &affinity);
197 if (result > 0) {
198 return result;
199 }
200 }
201 }
202#endif
203 return cpu_get_num_physical_cores();
204}
205
206// Helper for setting process priority
207
208#if defined(_WIN32)
209
210bool set_process_priority(enum ggml_sched_priority prio) {
211 if (prio == GGML_SCHED_PRIO_NORMAL) {
212 return true;
213 }
214
215 DWORD p = NORMAL_PRIORITY_CLASS;
216 switch (prio) {
217 case GGML_SCHED_PRIO_LOW: p = BELOW_NORMAL_PRIORITY_CLASS; break;
218 case GGML_SCHED_PRIO_NORMAL: p = NORMAL_PRIORITY_CLASS; break;
219 case GGML_SCHED_PRIO_MEDIUM: p = ABOVE_NORMAL_PRIORITY_CLASS; break;
220 case GGML_SCHED_PRIO_HIGH: p = HIGH_PRIORITY_CLASS; break;
221 case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS; break;
222 }
223
224 if (!SetPriorityClass(GetCurrentProcess(), p)) {
225 LOG_WRN("failed to set process priority class %d : (%d)\n", prio, (int) GetLastError());
226 return false;
227 }
228
229 return true;
230}
231
232#else // MacOS and POSIX
233#include <sys/types.h>
234#include <sys/resource.h>
235
236bool set_process_priority(enum ggml_sched_priority prio) {
237 if (prio == GGML_SCHED_PRIO_NORMAL) {
238 return true;
239 }
240
241 int p = 0;
242 switch (prio) {
243 case GGML_SCHED_PRIO_LOW: p = 5; break;
244 case GGML_SCHED_PRIO_NORMAL: p = 0; break;
245 case GGML_SCHED_PRIO_MEDIUM: p = -5; break;
246 case GGML_SCHED_PRIO_HIGH: p = -10; break;
247 case GGML_SCHED_PRIO_REALTIME: p = -20; break;
248 }
249
250 if (setpriority(PRIO_PROCESS, 0, p) != 0) {
251 LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
252 return false;
253 }
254 return true;
255}
256
257#endif
258
259//
260// CLI argument parsing
261//
262
263
264void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
265 int32_t n_set = 0;
266
267 if (cpuparams.n_threads < 0) {
268 // Assuming everything about cpuparams is invalid
269 if (role_model != nullptr) {
270 cpuparams = *role_model;
271 } else {
272 cpuparams.n_threads = cpu_get_num_math();
273 }
274 }
275
276 for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
277 if (cpuparams.cpumask[i]) {
278 n_set++;
279 }
280 }
281
282 if (n_set && n_set < cpuparams.n_threads) {
283 // Not enough set bits, may experience performance issues.
284 LOG_WRN("Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads);
285 }
286}
287
288bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) {
289 size_t dash_loc = range.find('-');
290 if (dash_loc == std::string::npos) {
291 LOG_ERR("Format of CPU range is invalid! Expected [<start>]-[<end>].\n");
292 return false;
293 }
294
295 size_t start_i;
296 size_t end_i;
297
298 if (dash_loc == 0) {
299 start_i = 0;
300 } else {
301 start_i = std::stoull(range.substr(0, dash_loc));
302 if (start_i >= GGML_MAX_N_THREADS) {
303 LOG_ERR("Start index out of bounds!\n");
304 return false;
305 }
306 }
307
308 if (dash_loc == range.length() - 1) {
309 end_i = GGML_MAX_N_THREADS - 1;
310 } else {
311 end_i = std::stoull(range.substr(dash_loc + 1));
312 if (end_i >= GGML_MAX_N_THREADS) {
313 LOG_ERR("End index out of bounds!\n");
314 return false;
315 }
316 }
317
318 for (size_t i = start_i; i <= end_i; i++) {
319 boolmask[i] = true;
320 }
321
322 return true;
323}
324
325bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) {
326 // Discard potential 0x prefix
327 size_t start_i = 0;
328 if (mask.length() >= 2 && mask.substr(0, 2) == "0x") {
329 start_i = 2;
330 }
331
332 size_t num_digits = mask.length() - start_i;
333 if (num_digits > 128) num_digits = 128;
334
335 size_t end_i = num_digits + start_i;
336
337 for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) {
338 char c = mask.at(i);
339 int8_t id = c;
340
341 if ((c >= '0' && c <= '9')) {
342 id -= '0';
343 } else if (c >= 'a' && c <= 'f') {
344 id -= 'a' - 10;
345 } else if (c >= 'A' && c <= 'F') {
346 id -= 'A' - 10;
347 } else {
348 LOG_ERR("Invalid hex character '%c' at position %d\n", c, int32_t(i));
349 return false;
350 }
351
352 boolmask[ n ] = boolmask[ n ] || ((id & 8) != 0);
353 boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0);
354 boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0);
355 boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0);
356 }
357
358 return true;
359}
360
361void common_init() {
362 llama_log_set(common_log_default_callback, NULL);
363
364#ifdef NDEBUG
365 const char * build_type = "";
366#else
367 const char * build_type = " (debug)";
368#endif
369
370 LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
371}
372
373std::string common_params_get_system_info(const common_params & params) {
374 std::ostringstream os;
375
376 os << "system_info: n_threads = " << params.cpuparams.n_threads;
377 if (params.cpuparams_batch.n_threads != -1) {
378 os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")";
379 }
380#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
381 // TODO: windows + arm64 + mingw64
382 DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
383 os << " / " << logicalProcessorCount << " | " << llama_print_system_info();
384#else
385 os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
386#endif
387
388 return os.str();
389}
390
391//
392// String utils
393//
394
395std::string string_format(const char * fmt, ...) {
396 va_list ap;
397 va_list ap2;
398 va_start(ap, fmt);
399 va_copy(ap2, ap);
400 int size = vsnprintf(NULL, 0, fmt, ap);
401 GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
402 std::vector<char> buf(size + 1);
403 int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
404 GGML_ASSERT(size2 == size);
405 va_end(ap2);
406 va_end(ap);
407 return std::string(buf.data(), size);
408}
409
410std::string string_strip(const std::string & str) {
411 size_t start = 0;
412 size_t end = str.size();
413 while (start < end && std::isspace(str[start])) {
414 start++;
415 }
416 while (end > start && std::isspace(str[end - 1])) {
417 end--;
418 }
419 return str.substr(start, end - start);
420}
421
422std::string string_get_sortable_timestamp() {
423 using clock = std::chrono::system_clock;
424
425 const clock::time_point current_time = clock::now();
426 const time_t as_time_t = clock::to_time_t(current_time);
427 char timestamp_no_ns[100];
428 std::strftime(timestamp_no_ns, 100, "%Y_%m_%d-%H_%M_%S", std::localtime(&as_time_t));
429
430 const int64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
431 current_time.time_since_epoch() % 1000000000).count();
432 char timestamp_ns[11];
433 snprintf(timestamp_ns, 11, "%09" PRId64, ns);
434
435 return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
436}
437
438void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
439 if (search.empty()) {
440 return;
441 }
442 std::string builder;
443 builder.reserve(s.length());
444 size_t pos = 0;
445 size_t last_pos = 0;
446 while ((pos = s.find(search, last_pos)) != std::string::npos) {
447 builder.append(s, last_pos, pos - last_pos);
448 builder.append(replace);
449 last_pos = pos + search.length();
450 }
451 builder.append(s, last_pos, std::string::npos);
452 s = std::move(builder);
453}
454
455bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
456 return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
457}
458
459bool string_remove_suffix(std::string & str, const std::string_view & suffix) {
460 bool has_suffix = string_ends_with(str, suffix);
461 if (has_suffix) {
462 str = str.substr(0, str.size() - suffix.size());
463 }
464 return has_suffix;
465}
466
467size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
468 if (!str.empty() && !stop.empty()) {
469 const char text_last_char = str.back();
470 for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
471 if (stop[char_index] == text_last_char) {
472 const auto current_partial = stop.substr(0, char_index + 1);
473 if (string_ends_with(str, current_partial)) {
474 return str.size() - char_index - 1;
475 }
476 }
477 }
478 }
479
480 return std::string::npos;
481}
482
483std::string regex_escape(const std::string & s) {
484 static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
485 return std::regex_replace(s, special_chars, "\\$&");
486}
487
488std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
489 std::ostringstream result;
490 for (size_t i = 0; i < values.size(); ++i) {
491 if (i > 0) {
492 result << separator;
493 }
494 result << values[i];
495 }
496 return result.str();
497}
498
499std::vector<std::string> string_split(const std::string & str, const std::string & delimiter) {
500 std::vector<std::string> parts;
501 size_t start = 0;
502 size_t end = str.find(delimiter);
503
504 while (end != std::string::npos) {
505 parts.push_back(str.substr(start, end - start));
506 start = end + delimiter.length();
507 end = str.find(delimiter, start);
508 }
509
510 parts.push_back(str.substr(start));
511
512 return parts;
513}
514
515std::string string_repeat(const std::string & str, size_t n) {
516 if (n == 0) {
517 return "";
518 }
519
520 std::string result;
521 result.reserve(str.length() * n);
522
523 for (size_t i = 0; i < n; ++i) {
524 result += str;
525 }
526
527 return result;
528}
529
530std::string string_from(bool value) {
531 return value ? "true" : "false";
532}
533
534std::string string_from(const std::vector<int> & values) {
535 std::stringstream buf;
536
537 buf << "[ ";
538 bool first = true;
539 for (auto e : values) {
540 if (first) {
541 first = false;
542 } else {
543 buf << ", ";
544 }
545 buf << std::to_string(e);
546 }
547 buf << " ]";
548
549 return buf.str();
550}
551
552std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens) {
553 std::stringstream buf;
554
555 buf << "[ ";
556
557 bool first = true;
558 for (const auto & token : tokens) {
559 if (!first) {
560 buf << ", ";
561 } else {
562 first = false;
563 }
564
565 auto detokenized = common_token_to_piece(ctx, token);
566
567 buf << "'" << detokenized << "'"
568 << ":" << std::to_string(token);
569 }
570
571 buf << " ]";
572
573 return buf.str();
574}
575
576std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch) {
577 std::stringstream buf;
578
579 buf << "[ ";
580
581 bool first = true;
582 for (int i = 0; i < batch.n_tokens; ++i) {
583 if (!first) {
584 buf << ", ";
585 } else {
586 first = false;
587 }
588
589 auto detokenized = common_token_to_piece(ctx, batch.token[i]);
590
591 buf << "\n" << std::to_string(i)
592 << ", token '" << detokenized << "'"
593 << ", pos " << std::to_string(batch.pos[i])
594 << ", n_seq_id " << std::to_string(batch.n_seq_id[i])
595 << ", seq_id " << std::to_string(batch.seq_id[i][0])
596 << ", logits " << std::to_string(batch.logits[i]);
597 }
598
599 buf << " ]";
600
601 return buf.str();
602}
603
604void string_process_escapes(std::string & input) {
605 std::size_t input_len = input.length();
606 std::size_t output_idx = 0;
607
608 for (std::size_t input_idx = 0; input_idx < input_len; ++input_idx) {
609 if (input[input_idx] == '\\' && input_idx + 1 < input_len) {
610 switch (input[++input_idx]) {
611 case 'n': input[output_idx++] = '\n'; break;
612 case 'r': input[output_idx++] = '\r'; break;
613 case 't': input[output_idx++] = '\t'; break;
614 case '\'': input[output_idx++] = '\''; break;
615 case '\"': input[output_idx++] = '\"'; break;
616 case '\\': input[output_idx++] = '\\'; break;
617 case 'x':
618 // Handle \x12, etc
619 if (input_idx + 2 < input_len) {
620 const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 };
621 char *err_p = nullptr;
622 const long val = std::strtol(x, &err_p, 16);
623 if (err_p == x + 2) {
624 input_idx += 2;
625 input[output_idx++] = char(val);
626 break;
627 }
628 }
629 // fall through
630 default: input[output_idx++] = '\\';
631 input[output_idx++] = input[input_idx]; break;
632 }
633 } else {
634 input[output_idx++] = input[input_idx];
635 }
636 }
637
638 input.resize(output_idx);
639}
640
641bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides) {
642 const char * sep = strchr(data, '=');
643 if (sep == nullptr || sep - data >= 128) {
644 LOG_ERR("%s: malformed KV override '%s'\n", __func__, data);
645 return false;
646 }
647 llama_model_kv_override kvo;
648 std::strncpy(kvo.key, data, sep - data);
649 kvo.key[sep - data] = 0;
650 sep++;
651 if (strncmp(sep, "int:", 4) == 0) {
652 sep += 4;
653 kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
654 kvo.val_i64 = std::atol(sep);
655 } else if (strncmp(sep, "float:", 6) == 0) {
656 sep += 6;
657 kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
658 kvo.val_f64 = std::atof(sep);
659 } else if (strncmp(sep, "bool:", 5) == 0) {
660 sep += 5;
661 kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
662 if (std::strcmp(sep, "true") == 0) {
663 kvo.val_bool = true;
664 } else if (std::strcmp(sep, "false") == 0) {
665 kvo.val_bool = false;
666 } else {
667 LOG_ERR("%s: invalid boolean value for KV override '%s'\n", __func__, data);
668 return false;
669 }
670 } else if (strncmp(sep, "str:", 4) == 0) {
671 sep += 4;
672 kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
673 if (strlen(sep) > 127) {
674 LOG_ERR("%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data);
675 return false;
676 }
677 strncpy(kvo.val_str, sep, 127);
678 kvo.val_str[127] = '\0';
679 } else {
680 LOG_ERR("%s: invalid type for KV override '%s'\n", __func__, data);
681 return false;
682 }
683 overrides.emplace_back(std::move(kvo));
684 return true;
685}
686
687//
688// Filesystem utils
689//
690
691// Validate if a filename is safe to use
692// To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
693bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
694 if (!filename.length()) {
695 // Empty filename invalid
696 return false;
697 }
698 if (filename.length() > 255) {
699 // Limit at common largest possible filename on Linux filesystems
700 // to avoid unnecessary further validation
701 // (On systems with smaller limits it will be caught by the OS)
702 return false;
703 }
704
705 size_t offset = 0;
706 while (offset < filename.size()) {
707 utf8_parse_result result = parse_utf8_codepoint(filename, offset);
708
709 if (result.status != utf8_parse_result::SUCCESS) {
710 return false;
711 }
712 uint32_t c = result.codepoint;
713
714 if ((result.bytes_consumed == 2 && c < 0x80) ||
715 (result.bytes_consumed == 3 && c < 0x800) ||
716 (result.bytes_consumed == 4 && c < 0x10000)) {
717 return false;
718 }
719
720 // Check for forbidden codepoints:
721 // - Control characters
722 // - Unicode equivalents of illegal characters
723 // - UTF-16 surrogate pairs
724 // - UTF-8 replacement character
725 // - Byte order mark (BOM)
726 // - Illegal characters: / \ : * ? " < > |
727 if (c <= 0x1F // Control characters (C0)
728 || c == 0x7F // Control characters (DEL)
729 || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
730 || c == 0xFF0E // Fullwidth Full Stop (period equivalent)
731 || c == 0x2215 // Division Slash (forward slash equivalent)
732 || c == 0x2216 // Set Minus (backslash equivalent)
733 || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
734 || c > 0x10FFFF // Max Unicode limit
735 || c == 0xFFFD // Replacement Character (UTF-8)
736 || c == 0xFEFF // Byte Order Mark (BOM)
737 || c == ':' || c == '*' // Illegal characters
738 || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
739 return false;
740 }
741 if (!allow_subdirs && (c == '/' || c == '\\')) {
742 // Subdirectories not allowed, reject path separators
743 return false;
744 }
745 offset += result.bytes_consumed;
746 }
747
748 // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
749 // Unicode and other whitespace is not affected, only 0x20 space
750 if (filename.front() == ' ' || filename.back() == ' ' || filename.back() == '.') {
751 return false;
752 }
753
754 // Reject any ".." (currently stricter than necessary, it should be fine to just check for == ".." instead)
755 if (filename.find("..") != std::string::npos) {
756 return false;
757 }
758
759 // Reject "."
760 if (filename == ".") {
761 return false;
762 }
763
764 return true;
765}
766
767#include <iostream>
768
769
770#ifdef _WIN32
771static std::wstring utf8_to_wstring(const std::string & str) {
772 if (str.empty()) {
773 return std::wstring();
774 }
775
776 int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
777
778 if (size <= 0) {
779 return std::wstring();
780 }
781
782 std::wstring wstr(size, 0);
783 MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
784
785 return wstr;
786}
787#endif
788
789// returns true if successful, false otherwise
790bool fs_create_directory_with_parents(const std::string & path) {
791#ifdef _WIN32
792 std::wstring wpath = utf8_to_wstring(path);
793
794 // if the path already exists, check whether it's a directory
795 const DWORD attributes = GetFileAttributesW(wpath.c_str());
796 if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
797 return true;
798 }
799
800 size_t pos_slash = 0;
801
802 // process path from front to back, procedurally creating directories
803 while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
804 const std::wstring subpath = wpath.substr(0, pos_slash);
805
806 pos_slash += 1;
807
808 // skip the drive letter, in some systems it can return an access denied error
809 if (subpath.length() == 2 && subpath[1] == ':') {
810 continue;
811 }
812
813 const bool success = CreateDirectoryW(subpath.c_str(), NULL);
814
815 if (!success) {
816 const DWORD error = GetLastError();
817
818 // if the path already exists, ensure that it's a directory
819 if (error == ERROR_ALREADY_EXISTS) {
820 const DWORD attributes = GetFileAttributesW(subpath.c_str());
821 if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
822 return false;
823 }
824 } else {
825 return false;
826 }
827 }
828 }
829
830 return true;
831#else
832 // if the path already exists, check whether it's a directory
833 struct stat info;
834 if (stat(path.c_str(), &info) == 0) {
835 return S_ISDIR(info.st_mode);
836 }
837
838 size_t pos_slash = 1; // skip leading slashes for directory creation
839
840 // process path from front to back, procedurally creating directories
841 while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
842 const std::string subpath = path.substr(0, pos_slash);
843 struct stat info;
844
845 // if the path already exists, ensure that it's a directory
846 if (stat(subpath.c_str(), &info) == 0) {
847 if (!S_ISDIR(info.st_mode)) {
848 return false;
849 }
850 } else {
851 // create parent directories
852 const int ret = mkdir(subpath.c_str(), 0755);
853 if (ret != 0) {
854 return false;
855 }
856 }
857
858 pos_slash += 1;
859 }
860
861 return true;
862#endif // _WIN32
863}
864
865bool fs_is_directory(const std::string & path) {
866 std::filesystem::path dir(path);
867 return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
868}
869
870std::string fs_get_cache_directory() {
871 std::string cache_directory = "";
872 auto ensure_trailing_slash = [](std::string p) {
873 // Make sure to add trailing slash
874 if (p.back() != DIRECTORY_SEPARATOR) {
875 p += DIRECTORY_SEPARATOR;
876 }
877 return p;
878 };
879 if (getenv("LLAMA_CACHE")) {
880 cache_directory = std::getenv("LLAMA_CACHE");
881 } else {
882#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
883 if (std::getenv("XDG_CACHE_HOME")) {
884 cache_directory = std::getenv("XDG_CACHE_HOME");
885 } else if (std::getenv("HOME")) {
886 cache_directory = std::getenv("HOME") + std::string("/.cache/");
887 } else {
888#if defined(__linux__)
889 /* no $HOME is defined, fallback to getpwuid */
890 struct passwd *pw = getpwuid(getuid());
891 if ((!pw) || (!pw->pw_dir)) {
892 throw std::runtime_error("Failed to find $HOME directory");
893 }
894
895 cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
896#else /* defined(__linux__) */
897 throw std::runtime_error("Failed to find $HOME directory");
898#endif /* defined(__linux__) */
899 }
900#elif defined(__APPLE__)
901 cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
902#elif defined(_WIN32)
903 cache_directory = std::getenv("LOCALAPPDATA");
904#elif defined(__EMSCRIPTEN__)
905 GGML_ABORT("not implemented on this platform");
906#else
907# error Unknown architecture
908#endif
909 cache_directory = ensure_trailing_slash(cache_directory);
910 cache_directory += "llama.cpp";
911 }
912 return ensure_trailing_slash(cache_directory);
913}
914
915std::string fs_get_cache_file(const std::string & filename) {
916 GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
917 std::string cache_directory = fs_get_cache_directory();
918 const bool success = fs_create_directory_with_parents(cache_directory);
919 if (!success) {
920 throw std::runtime_error("failed to create cache directory: " + cache_directory);
921 }
922 return cache_directory + filename;
923}
924
925std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
926 std::vector<common_file_info> files;
927 if (path.empty()) return files;
928
929 std::filesystem::path dir(path);
930 if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
931 return files;
932 }
933
934 for (const auto & entry : std::filesystem::directory_iterator(dir)) {
935 try {
936 // Only include regular files (skip directories)
937 const auto & p = entry.path();
938 if (std::filesystem::is_regular_file(p)) {
939 common_file_info info;
940 info.path = p.string();
941 info.name = p.filename().string();
942 info.is_dir = false;
943 try {
944 info.size = static_cast<size_t>(std::filesystem::file_size(p));
945 } catch (const std::filesystem::filesystem_error &) {
946 info.size = 0;
947 }
948 files.push_back(std::move(info));
949 } else if (include_directories && std::filesystem::is_directory(p)) {
950 common_file_info info;
951 info.path = p.string();
952 info.name = p.filename().string();
953 info.size = 0; // Directories have no size
954 info.is_dir = true;
955 files.push_back(std::move(info));
956 }
957 } catch (const std::filesystem::filesystem_error &) {
958 // skip entries we cannot inspect
959 continue;
960 }
961 }
962
963 return files;
964}
965
966//
967// TTY utils
968//
969
970bool tty_can_use_colors() {
971 // Check NO_COLOR environment variable (https://no-color.org/)
972 if (const char * no_color = std::getenv("NO_COLOR")) {
973 if (no_color[0] != '\0') {
974 return false;
975 }
976 }
977
978 // Check TERM environment variable
979 if (const char * term = std::getenv("TERM")) {
980 if (std::strcmp(term, "dumb") == 0) {
981 return false;
982 }
983 }
984
985 // Check if stdout and stderr are connected to a terminal
986 // We check both because log messages can go to either
987 bool stdout_is_tty = isatty(fileno(stdout));
988 bool stderr_is_tty = isatty(fileno(stderr));
989
990 return stdout_is_tty || stderr_is_tty;
991}
992
993//
994// Model utils
995//
996
997// TODO: move to common/sampling
998static void common_init_sampler_from_model(
999 const llama_model * model,
1000 common_params_sampling & sparams) {
1001
1002 const uint64_t config = sparams.user_sampling_config;
1003
1004 auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
1005 if (config & user_config) {
1006 return;
1007 }
1008
1009 char buf[64] = {0};
1010 if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
1011 char * end = nullptr;
1012 int32_t v = strtol(buf, &end, 10);
1013 if (end && end != buf) {
1014 dst = v;
1015 }
1016 }
1017 };
1018
1019 auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
1020 if (config & user_config) {
1021 return;
1022 }
1023
1024 char buf[128] = {0};
1025 if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
1026 char * end = nullptr;
1027 float v = strtof(buf, &end);
1028 if (end && end != buf) {
1029 dst = v;
1030 }
1031 }
1032 };
1033
1034 // Sampling sequence
1035 if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
1036 char buf[512] = {0};
1037 if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
1038 const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
1039 if (!sampler_names.empty()) {
1040 sparams.samplers = common_sampler_types_from_names(sampler_names, true);
1041 }
1042 }
1043 }
1044
1045 get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K), sparams.top_k, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
1046 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P), sparams.top_p, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
1047 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P), sparams.min_p, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
1048 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
1049 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD), sparams.xtc_threshold, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
1050 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP), sparams.temp, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
1051 get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N), sparams.penalty_last_n, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
1052 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT), sparams.penalty_repeat, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
1053 get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT), sparams.mirostat, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
1054 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU), sparams.mirostat_tau, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
1055 get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA), sparams.mirostat_eta, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
1056}
1057
1058struct common_init_result::impl {
1059 impl() = default;
1060 ~impl() = default;
1061
1062 // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
1063
1064 llama_model_ptr model;
1065 llama_context_ptr context;
1066
1067 std::vector<llama_adapter_lora_ptr> lora;
1068
1069 std::vector<common_sampler_ptr> samplers;
1070 std::vector<llama_sampler_seq_config> samplers_seq_config;
1071};
1072
1073common_init_result::common_init_result(common_params & params) :
1074 pimpl(new impl{}) {
1075 auto mparams = common_model_params_to_llama(params);
1076 auto cparams = common_context_params_to_llama(params);
1077
1078 if (params.fit_params) {
1079 LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
1080 llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
1081 params.tensor_split,
1082 params.tensor_buft_overrides.data(),
1083 params.fit_params_target.data(),
1084 params.fit_params_min_ctx,
1085 params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
1086 }
1087
1088 llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
1089 if (model == NULL) {
1090 return;
1091 }
1092
1093 pimpl->model.reset(model);
1094
1095 const llama_vocab * vocab = llama_model_get_vocab(model);
1096
1097 // load and optionally apply lora adapters (must be loaded before context creation)
1098 for (auto & la : params.lora_adapters) {
1099 llama_adapter_lora_ptr lora;
1100 lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
1101 if (lora == nullptr) {
1102 LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
1103 pimpl->model.reset(model);
1104 return;
1105 }
1106
1107 char buf[1024];
1108 la.ptr = lora.get();
1109 llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
1110 la.task_name = buf;
1111 llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
1112 la.prompt_prefix = buf;
1113 pimpl->lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
1114 }
1115
1116 // updates params.sampling
1117 // TODO: fix naming
1118 common_init_sampler_from_model(model, params.sampling);
1119
1120 if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
1121 LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
1122 params.sampling.ignore_eos = false;
1123 }
1124
1125 // initialize once
1126 for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
1127 if (llama_vocab_is_eog(vocab, i)) {
1128 LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
1129 params.sampling.logit_bias_eog.push_back({i, -INFINITY});
1130 }
1131 }
1132
1133 if (params.sampling.ignore_eos) {
1134 // add EOG biases to the active set of logit biases
1135 params.sampling.logit_bias.insert(
1136 params.sampling.logit_bias.end(),
1137 params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
1138 }
1139
1140 //if (params.sampling.penalty_last_n == -1) {
1141 // LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
1142 // params.sampling.penalty_last_n = llama_n_ctx(lctx);
1143 //}
1144
1145 //if (params.sampling.dry_penalty_last_n == -1) {
1146 // LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
1147 // params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
1148 //}
1149
1150 // init the backend samplers as part of the context creation
1151 pimpl->samplers.resize(cparams.n_seq_max);
1152 pimpl->samplers_seq_config.resize(cparams.n_seq_max);
1153
1154 for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
1155 pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
1156 pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
1157 }
1158
1159 if (params.sampling.backend_sampling) {
1160 cparams.samplers = pimpl->samplers_seq_config.data();
1161 cparams.n_samplers = pimpl->samplers_seq_config.size();
1162 }
1163
1164 llama_context * lctx = llama_init_from_model(model, cparams);
1165 if (lctx == NULL) {
1166 LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
1167 return;
1168 }
1169
1170 pimpl->context.reset(lctx);
1171}
1172
1173llama_model * common_init_result::model() {
1174 return pimpl->model.get();
1175}
1176
1177llama_context * common_init_result::context() {
1178 return pimpl->context.get();
1179}
1180
1181common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
1182 return pimpl->samplers[seq_id].get();
1183}
1184
1185void common_init_result::reset_samplers() {
1186 for (int i = 0; i < (int) pimpl->samplers.size(); ++i) {
1187 llama_sampler_reset(common_sampler_get(pimpl->samplers[i].get()));
1188 }
1189}
1190
1191std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
1192 return pimpl->lora;
1193}
1194
1195common_init_result_ptr common_init_from_params(common_params & params) {
1196 common_init_result_ptr res(new common_init_result(params));
1197
1198 llama_model * model = res->model();
1199 if (model == NULL) {
1200 LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
1201 return res;
1202 }
1203
1204 llama_context * lctx = res->context();
1205 if (lctx == NULL) {
1206 LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
1207 return res;
1208 }
1209
1210 const llama_vocab * vocab = llama_model_get_vocab(model);
1211
1212 if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
1213 LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
1214 params.ctx_shift = false;
1215 }
1216
1217 if (!params.control_vectors.empty()) {
1218 if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
1219 if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_model_n_layer(model);
1220
1221 const auto cvec = common_control_vector_load(params.control_vectors);
1222 if (cvec.n_embd == -1) {
1223 return res;
1224 }
1225
1226 int err = llama_apply_adapter_cvec(
1227 lctx,
1228 cvec.data.data(),
1229 cvec.data.size(),
1230 cvec.n_embd,
1231 params.control_vector_layer_start,
1232 params.control_vector_layer_end);
1233 if (err) {
1234 return res;
1235 }
1236 }
1237
1238 if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
1239 bool ok = true;
1240
1241 if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
1242 LOG_WRN("%s: warning: vocab does not have a BOS token, reranking will not work\n", __func__);
1243 ok = false;
1244 }
1245
1246 bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
1247 bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
1248 bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL;
1249
1250 if (!has_eos && !has_sep && !has_rerank_prompt) {
1251 LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__);
1252 ok = false;
1253 } else if (!has_eos) {
1254 LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
1255 }
1256
1257 if (!ok) {
1258 return res;
1259 }
1260 }
1261
1262 if (!params.lora_init_without_apply) {
1263 common_set_adapter_lora(lctx, params.lora_adapters);
1264 }
1265
1266 if (params.warmup) {
1267 LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
1268
1269 llama_set_warmup(lctx, true);
1270
1271 std::vector<llama_token> tmp;
1272 llama_token bos = llama_vocab_bos(vocab);
1273 llama_token eos = llama_vocab_eos(vocab);
1274
1275 // some models (e.g. T5) don't have a BOS token
1276 if (bos != LLAMA_TOKEN_NULL) {
1277 tmp.push_back(bos);
1278 }
1279 if (eos != LLAMA_TOKEN_NULL) {
1280 tmp.push_back(eos);
1281 }
1282 if (tmp.empty()) {
1283 tmp.push_back(0);
1284 }
1285
1286 if (llama_model_has_encoder(model)) {
1287 llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
1288 llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
1289 if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
1290 decoder_start_token_id = bos;
1291 }
1292 tmp.clear();
1293 tmp.push_back(decoder_start_token_id);
1294 }
1295 if (llama_model_has_decoder(model)) {
1296 llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
1297 }
1298 llama_memory_clear(llama_get_memory(lctx), true);
1299 llama_synchronize(lctx);
1300 llama_perf_context_reset(lctx);
1301 llama_set_warmup(lctx, false);
1302
1303 // reset samplers to reset RNG state after warmup to the seeded state
1304 res->reset_samplers();
1305 }
1306
1307 return res;
1308}
1309
1310common_init_result::~common_init_result() = default;
1311
1312std::string get_model_endpoint() {
1313 const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
1314 // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
1315 const char * hf_endpoint_env = getenv("HF_ENDPOINT");
1316 const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
1317 std::string model_endpoint = "https://huggingface.co/";
1318 if (endpoint_env) {
1319 model_endpoint = endpoint_env;
1320 if (model_endpoint.back() != '/') {
1321 model_endpoint += '/';
1322 }
1323 }
1324 return model_endpoint;
1325}
1326
1327void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
1328 llama_clear_adapter_lora(ctx);
1329 for (auto & la : lora) {
1330 if (la.scale != 0.0f) {
1331 llama_set_adapter_lora(ctx, la.ptr, la.scale);
1332 }
1333 }
1334}
1335
1336struct llama_model_params common_model_params_to_llama(common_params & params) {
1337 auto mparams = llama_model_default_params();
1338
1339 if (!params.devices.empty()) {
1340 mparams.devices = params.devices.data();
1341 }
1342
1343 mparams.n_gpu_layers = params.n_gpu_layers;
1344 mparams.main_gpu = params.main_gpu;
1345 mparams.split_mode = params.split_mode;
1346 mparams.tensor_split = params.tensor_split;
1347 mparams.use_mmap = params.use_mmap;
1348 mparams.use_direct_io = params.use_direct_io;
1349 mparams.use_mlock = params.use_mlock;
1350 mparams.check_tensors = params.check_tensors;
1351 mparams.use_extra_bufts = !params.no_extra_bufts;
1352 mparams.no_host = params.no_host;
1353
1354 if (params.kv_overrides.empty()) {
1355 mparams.kv_overrides = NULL;
1356 } else {
1357 GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
1358 mparams.kv_overrides = params.kv_overrides.data();
1359 }
1360
1361 if (params.tensor_buft_overrides.empty()) {
1362 mparams.tensor_buft_overrides = NULL;
1363 } else {
1364 GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
1365 mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
1366 }
1367
1368 mparams.progress_callback = params.load_progress_callback;
1369 mparams.progress_callback_user_data = params.load_progress_callback_user_data;
1370
1371 return mparams;
1372}
1373
1374struct llama_context_params common_context_params_to_llama(const common_params & params) {
1375 auto cparams = llama_context_default_params();
1376
1377 cparams.n_ctx = params.n_ctx;
1378 cparams.n_seq_max = params.n_parallel;
1379 cparams.n_batch = params.n_batch;
1380 cparams.n_ubatch = params.n_ubatch;
1381 cparams.n_threads = params.cpuparams.n_threads;
1382 cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ?
1383 params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
1384 cparams.embeddings = params.embedding;
1385 cparams.rope_scaling_type = params.rope_scaling_type;
1386 cparams.rope_freq_base = params.rope_freq_base;
1387 cparams.rope_freq_scale = params.rope_freq_scale;
1388 cparams.yarn_ext_factor = params.yarn_ext_factor;
1389 cparams.yarn_attn_factor = params.yarn_attn_factor;
1390 cparams.yarn_beta_fast = params.yarn_beta_fast;
1391 cparams.yarn_beta_slow = params.yarn_beta_slow;
1392 cparams.yarn_orig_ctx = params.yarn_orig_ctx;
1393 cparams.pooling_type = params.pooling_type;
1394 cparams.attention_type = params.attention_type;
1395 cparams.flash_attn_type = params.flash_attn_type;
1396 cparams.cb_eval = params.cb_eval;
1397 cparams.cb_eval_user_data = params.cb_eval_user_data;
1398 cparams.offload_kqv = !params.no_kv_offload;
1399 cparams.no_perf = params.no_perf;
1400 cparams.op_offload = !params.no_op_offload;
1401 cparams.swa_full = params.swa_full;
1402 cparams.kv_unified = params.kv_unified;
1403
1404 cparams.type_k = params.cache_type_k;
1405 cparams.type_v = params.cache_type_v;
1406
1407 return cparams;
1408}
1409
1410struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) {
1411 struct ggml_threadpool_params tpp;
1412
1413 ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults
1414
1415 if (params.mask_valid) {
1416 std::memcpy(&tpp.cpumask, ¶ms.cpumask, GGML_MAX_N_THREADS);
1417 }
1418
1419 tpp.prio = params.priority;
1420 tpp.poll = params.poll;
1421 tpp.strict_cpu = params.strict_cpu;
1422
1423 return tpp;
1424}
1425
1426//
1427// Batch utils
1428//
1429
1430void common_batch_clear(struct llama_batch & batch) {
1431 batch.n_tokens = 0;
1432}
1433
1434void common_batch_add(
1435 struct llama_batch & batch,
1436 llama_token id,
1437 llama_pos pos,
1438 const std::vector<llama_seq_id> & seq_ids,
1439 bool logits) {
1440 GGML_ASSERT(batch.seq_id[batch.n_tokens] && "llama_batch size exceeded");
1441
1442 batch.token [batch.n_tokens] = id;
1443 batch.pos [batch.n_tokens] = pos;
1444 batch.n_seq_id[batch.n_tokens] = seq_ids.size();
1445 for (size_t i = 0; i < seq_ids.size(); ++i) {
1446 batch.seq_id[batch.n_tokens][i] = seq_ids[i];
1447 }
1448 batch.logits [batch.n_tokens] = logits;
1449
1450 batch.n_tokens++;
1451}
1452
1453//
1454// Vocab utils
1455//
1456
1457std::vector<llama_token> common_tokenize(
1458 const struct llama_context * ctx,
1459 const std::string & text,
1460 bool add_special,
1461 bool parse_special) {
1462 const llama_model * model = llama_get_model(ctx);
1463 const llama_vocab * vocab = llama_model_get_vocab(model);
1464 return common_tokenize(vocab, text, add_special, parse_special);
1465}
1466
1467std::vector<llama_token> common_tokenize(
1468 const struct llama_vocab * vocab,
1469 const std::string & text,
1470 bool add_special,
1471 bool parse_special) {
1472 // upper limit for the number of tokens
1473 int n_tokens = text.length() + 2 * add_special;
1474 std::vector<llama_token> result(n_tokens);
1475 n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
1476 if (n_tokens == std::numeric_limits<int32_t>::min()) {
1477 throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
1478 }
1479 if (n_tokens < 0) {
1480 result.resize(-n_tokens);
1481 int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
1482 GGML_ASSERT(check == -n_tokens);
1483 } else {
1484 result.resize(n_tokens);
1485 }
1486 return result;
1487}
1488
1489std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
1490 const llama_model * model = llama_get_model(ctx);
1491 const llama_vocab * vocab = llama_model_get_vocab(model);
1492 return common_token_to_piece(vocab, token, special);
1493}
1494
1495std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token token, bool special) {
1496 std::string piece;
1497 piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n'
1498 const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
1499 if (n_chars < 0) {
1500 piece.resize(-n_chars);
1501 int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
1502 GGML_ASSERT(check == -n_chars);
1503 }
1504 else {
1505 piece.resize(n_chars);
1506 }
1507
1508 return piece;
1509}
1510
1511std::string common_detokenize(const struct llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
1512 const llama_model * model = llama_get_model(ctx);
1513 const llama_vocab * vocab = llama_model_get_vocab(model);
1514 return common_detokenize(vocab, tokens, special);
1515}
1516
1517std::string common_detokenize(const struct llama_vocab * vocab, const std::vector<llama_token> & tokens, bool special) {
1518 std::string text;
1519 text.resize(std::max(text.capacity(), tokens.size()));
1520 int32_t n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
1521 if (n_chars < 0) {
1522 text.resize(-n_chars);
1523 n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
1524 GGML_ASSERT(n_chars <= (int32_t)text.size()); // whitespace trimming is performed after per-token detokenization
1525 }
1526
1527 text.resize(n_chars);
1528
1529 // NOTE: the original tokenizer decodes bytes after collecting the pieces.
1530 return text;
1531}
1532
1533//
1534// Embedding utils
1535//
1536
1537void common_embd_normalize(const float * inp, float * out, int n, int embd_norm) {
1538 double sum = 0.0;
1539
1540 switch (embd_norm) {
1541 case -1: // no normalisation
1542 sum = 1.0;
1543 break;
1544 case 0: // max absolute
1545 for (int i = 0; i < n; i++) {
1546 if (sum < std::abs(inp[i])) {
1547 sum = std::abs(inp[i]);
1548 }
1549 }
1550 sum /= 32760.0; // make an int16 range
1551 break;
1552 case 2: // euclidean
1553 for (int i = 0; i < n; i++) {
1554 sum += inp[i] * inp[i];
1555 }
1556 sum = std::sqrt(sum);
1557 break;
1558 default: // p-norm (euclidean is p-norm p=2)
1559 for (int i = 0; i < n; i++) {
1560 sum += std::pow(std::abs(inp[i]), embd_norm);
1561 }
1562 sum = std::pow(sum, 1.0 / embd_norm);
1563 break;
1564 }
1565
1566 const float norm = sum > 0.0 ? 1.0 / sum : 0.0f;
1567
1568 for (int i = 0; i < n; i++) {
1569 out[i] = inp[i] * norm;
1570 }
1571}
1572
1573float common_embd_similarity_cos(const float * embd1, const float * embd2, int n){
1574 double sum = 0.0;
1575 double sum1 = 0.0;
1576 double sum2 = 0.0;
1577
1578 for (int i = 0; i < n; i++) {
1579 sum += embd1[i] * embd2[i];
1580 sum1 += embd1[i] * embd1[i];
1581 sum2 += embd2[i] * embd2[i];
1582 }
1583
1584 // Handle the case where one or both vectors are zero vectors
1585 if (sum1 == 0.0 || sum2 == 0.0) {
1586 if (sum1 == 0.0 && sum2 == 0.0) {
1587 return 1.0f; // two zero vectors are similar
1588 }
1589 return 0.0f;
1590 }
1591
1592 return sum / (sqrt(sum1) * sqrt(sum2));
1593}
1594
1595//
1596// Control vector utils
1597//
1598
1599static common_control_vector_data common_control_vector_load_one(const common_control_vector_load_info & load_info) {
1600 common_control_vector_data result = { -1, {} };
1601
1602 ggml_context * ctx = nullptr;
1603 struct gguf_init_params meta_gguf_params = {
1604 /* .no_alloc = */ false,
1605 /* .ctx = */ &ctx,
1606 };
1607 struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params);
1608 if (!ctx_gguf) {
1609 LOG_ERR("%s: failed to load control vector file from %s\n", __func__, load_info.fname.c_str());
1610 return result;
1611 }
1612
1613 int32_t n_tensors = gguf_get_n_tensors(ctx_gguf);
1614 if (n_tensors == 0) {
1615 LOG_WRN("%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str());
1616 }
1617
1618 for (int i = 0; i < n_tensors; i++) {
1619 std::string name = gguf_get_tensor_name(ctx_gguf, i);
1620
1621 int layer_idx = -1;
1622
1623 // split on '.'
1624 size_t dotpos = name.find('.');
1625 if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") {
1626 try {
1627 layer_idx = std::stoi(name.substr(dotpos + 1));
1628 } catch (...) {
1629 layer_idx = -1;
1630 }
1631 }
1632 if (layer_idx < 0) {
1633 LOG_ERR("%s: invalid/unparsable direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
1634 result.n_embd = -1;
1635 break;
1636 } else if (layer_idx == 0) {
1637 LOG_ERR("%s: invalid (zero) direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
1638 result.n_embd = -1;
1639 break;
1640 }
1641
1642 struct ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str());
1643 if (tensor->type != GGML_TYPE_F32) {
1644 LOG_ERR("%s: invalid (non-F32) direction tensor type in %s\n", __func__, load_info.fname.c_str());
1645 result.n_embd = -1;
1646 break;
1647 }
1648 if (ggml_n_dims(tensor) != 1) {
1649 LOG_ERR("%s: invalid (non-1D) direction tensor shape in %s\n", __func__, load_info.fname.c_str());
1650 result.n_embd = -1;
1651 break;
1652 }
1653
1654 if (result.n_embd == -1) {
1655 result.n_embd = ggml_nelements(tensor);
1656 } else if (ggml_nelements(tensor) != result.n_embd) {
1657 LOG_ERR("%s: direction tensor in %s does not match previous dimensions\n", __func__, load_info.fname.c_str());
1658 result.n_embd = -1;
1659 break;
1660 }
1661
1662 // extend if necessary - do not store data for layer 0 (it's not used)
1663 result.data.resize(std::max(result.data.size(), static_cast<size_t>(result.n_embd * layer_idx)), 0.0f);
1664
1665 const float * src = (const float *) tensor->data;
1666 float * dst = result.data.data() + result.n_embd * (layer_idx - 1); // layer 1 at [0]
1667 for (int j = 0; j < result.n_embd; j++) {
1668 dst[j] += src[j] * load_info.strength; // allows multiple directions for same layer in same file
1669 }
1670
1671 }
1672
1673 if (result.n_embd == -1) {
1674 LOG_WRN("%s: skipping %s due to invalid direction tensors\n", __func__, load_info.fname.c_str());
1675 result.data.clear();
1676 }
1677
1678 gguf_free(ctx_gguf);
1679 ggml_free(ctx);
1680
1681 return result;
1682}
1683
1684common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos) {
1685 common_control_vector_data result = { -1, {} };
1686
1687 for (const auto & info : load_infos) {
1688 auto cur = common_control_vector_load_one(info);
1689
1690 if (cur.n_embd == -1) {
1691 result.n_embd = -1;
1692 break;
1693 }
1694 if (result.n_embd != -1 && result.n_embd != cur.n_embd) {
1695 LOG_ERR("%s: control vectors in %s does not match previous dimensions\n", __func__, info.fname.c_str());
1696 result.n_embd = -1;
1697 break;
1698 }
1699
1700 if (result.n_embd == -1) {
1701 result = std::move(cur);
1702 } else {
1703 result.data.resize(std::max(result.data.size(), cur.data.size()), 0.0f); // extend if necessary
1704 for (size_t i = 0; i < cur.data.size(); i++) {
1705 result.data[i] += cur.data[i];
1706 }
1707 }
1708 }
1709
1710 if (result.n_embd == -1) {
1711 LOG_ERR("%s: no valid control vector files passed\n", __func__);
1712 result.data.clear();
1713 }
1714
1715 return result;
1716}
1717
1718ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
1719 const int64_t ne_datapoint = llama_n_ctx(ctx);
1720 const int64_t ndata = (tokens.size() - ne_datapoint - 1) / stride;
1721 ggml_opt_dataset_t result = ggml_opt_dataset_init(
1722 GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
1723
1724 llama_token * data = (llama_token *) ggml_opt_dataset_data(result)->data;
1725 llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
1726
1727 for (int64_t idata = 0; idata < ndata; ++idata) {
1728 memcpy(data + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
1729 memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
1730 }
1731
1732 return result;
1733}
1734
1735ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
1736 ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
1737 const lr_opt & d = *(lr_opt *) userdata;
1738 result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
1739 result.sgd.wd = result.adamw.wd = d.wd;
1740 return result;
1741}
1742
1743// TODO make all command line args case-insensitive
1744static inline bool eq_case_insensitive(char const* a, char const* b) {
1745 return !
1746#if defined(_MSC_VER)
1747 _stricmp
1748#else
1749 strcasecmp
1750#endif // defined(_MSC_VER)
1751 (a, b);
1752}
1753
1754enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
1755 if (eq_case_insensitive("adamw", n)) {
1756 return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
1757 }
1758 if (eq_case_insensitive("sgd", n)) {
1759 return GGML_OPT_OPTIMIZER_TYPE_SGD;
1760 }
1761 return GGML_OPT_OPTIMIZER_TYPE_COUNT;
1762}
1763
1764// TODO simplify to use just log and exp
1765static float const k_log_2 = std::log(2.f);
1766
1767void lr_opt::init() {
1768 if (lr_min > 0 && lr_min < lr0) {
1769 float nhalf = std::log(lr0 / lr_min) / k_log_2;
1770 float e = epochs;
1771 if (decay_epochs > 0 && decay_epochs < e) {
1772 e = decay_epochs;
1773 } else {
1774 decay_epochs = e;
1775 }
1776 scale_epoch = nhalf / e;
1777 }
1778}
1779
1780float lr_opt::get_lr(float epoch) const {
1781 float r = lr_min <= 0 ? lr0 :
1782 epoch >= decay_epochs ? lr_min :
1783 lr0 * std::pow(0.5f, epoch * scale_epoch);
1784 LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
1785 return r;
1786}